Source code for tensorlayer.layers

#! /usr/bin/python
# -*- coding: utf8 -*-



import tensorflow as tf
import time
from . import visualize
from . import utils
from . import files
from . import cost
from . import iterate
from . import ops
import numpy as np
from six.moves import xrange
import random, warnings
import copy
import inspect
# __all__ = [
#     "Layer",
#     "DenseLayer",
# ]


# set_keep = locals()
set_keep = globals()
set_keep['_layers_name_list'] =[]
set_keep['name_reuse'] = False

try:  # For TF12 and later
    TF_GRAPHKEYS_VARIABLES = tf.GraphKeys.GLOBAL_VARIABLES
except:  # For TF11 and before
    TF_GRAPHKEYS_VARIABLES = tf.GraphKeys.VARIABLES

## Variable Operation
[docs]def flatten_reshape(variable, name=''): """Reshapes high-dimension input to a vector. [batch_size, mask_row, mask_col, n_mask] ---> [batch_size, mask_row * mask_col * n_mask] Parameters ---------- variable : a tensorflow variable name : a string or None An optional name to attach to this layer. Examples -------- >>> W_conv2 = weight_variable([5, 5, 100, 32]) # 64 features for each 5x5 patch >>> b_conv2 = bias_variable([32]) >>> W_fc1 = weight_variable([7 * 7 * 32, 256]) >>> h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) >>> h_pool2 = max_pool_2x2(h_conv2) >>> h_pool2.get_shape()[:].as_list() = [batch_size, 7, 7, 32] ... [batch_size, mask_row, mask_col, n_mask] >>> h_pool2_flat = tl.layers.flatten_reshape(h_pool2) ... [batch_size, mask_row * mask_col * n_mask] >>> h_pool2_flat_drop = tf.nn.dropout(h_pool2_flat, keep_prob) ... """ dim = 1 for d in variable.get_shape()[1:].as_list(): dim *= d return tf.reshape(variable, shape=[-1, dim], name=name)
[docs]def clear_layers_name(): """Clear all layer names in set_keep['_layers_name_list'], enable layer name reuse. Examples --------- >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.DenseLayer(network, n_units=800, name='relu1') ... >>> tl.layers.clear_layers_name() >>> network2 = tl.layers.InputLayer(x, name='input_layer') >>> network2 = tl.layers.DenseLayer(network2, n_units=800, name='relu1') ... """ set_keep['_layers_name_list'] =[]
[docs]def set_name_reuse(enable=True): """Enable or disable reuse layer name. By default, each layer must has unique name. When you want two or more input placeholder (inference) share the same model parameters, you need to enable layer name reuse, then allow the parameters have same name scope. Parameters ------------ enable : boolean, enable name reuse. (None means False). Examples ------------ >>> def embed_seq(input_seqs, is_train, reuse): >>> with tf.variable_scope("model", reuse=reuse): >>> tl.layers.set_name_reuse(reuse) >>> network = tl.layers.EmbeddingInputlayer( ... inputs = input_seqs, ... vocabulary_size = vocab_size, ... embedding_size = embedding_size, ... name = 'e_embedding') >>> network = tl.layers.DynamicRNNLayer(network, ... cell_fn = tf.nn.rnn_cell.BasicLSTMCell, ... n_hidden = embedding_size, ... dropout = (0.7 if is_train else None), ... initializer = w_init, ... sequence_length = tl.layers.retrieve_seq_length_op2(input_seqs), ... return_last = True, ... name = 'e_dynamicrnn',) >>> return network >>> >>> net_train = embed_seq(t_caption, is_train=True, reuse=False) >>> net_test = embed_seq(t_caption, is_train=False, reuse=True) - see ``tutorial_ptb_lstm.py`` for example. """ set_keep['name_reuse'] = enable
[docs]def initialize_rnn_state(state): """Return the initialized RNN state. The input is LSTMStateTuple or State of RNNCells. Parameters ----------- state : a RNN state. """ try: # TF1.0 LSTMStateTuple = tf.contrib.rnn.LSTMStateTuple except: LSTMStateTuple = tf.nn.rnn_cell.LSTMStateTuple if isinstance(state, LSTMStateTuple): c = state.c.eval() h = state.h.eval() return (c, h) else: new_state = state.eval() return new_state
[docs]def get_variables_with_name(name, train_only=True, printable=False): """Get variable list by a given name scope. Examples --------- >>> dense_vars = tl.layers.get_variable_with_name('dense', True, True) """ print(" [*] geting variables with %s" % name) # tvar = tf.trainable_variables() if train_only else tf.all_variables() if train_only: t_vars = tf.trainable_variables() else: try: # TF1.0 t_vars = tf.global_variables() except: # TF0.12 t_vars = tf.all_variables() d_vars = [var for var in t_vars if name in var.name] if printable: for idx, v in enumerate(d_vars): print(" got {:3}: {:15} {}".format(idx, v.name, str(v.get_shape()))) return d_vars
[docs]def get_layers_with_name(network=None, name="", printable=False): """Get layer list in a network by a given name scope. Examples --------- >>> layers = tl.layers.get_layers_with_name(network, "CNN", True) """ assert network is not None print(" [*] geting layers with %s" % name) layers = [] i = 0 for layer in network.all_layers: # print(type(layer.name)) if name in layer.name: layers.append(layer) if printable: # print(layer.name) print(" got {:3}: {:15} {}".format(i, layer.name, str(layer.get_shape()))) i = i + 1 return layers
[docs]def list_remove_repeat(l=None): """Remove the repeated items in a list, and return the processed list. You may need it to create merged layer like Concat, Elementwise and etc. Parameters ---------- l : a list Examples --------- >>> l = [2, 3, 4, 2, 3] >>> l = list_remove_repeat(l) ... [2, 3, 4] """ l2 = [] [l2.append(i) for i in l if not i in l2] return l2
[docs]def initialize_global_variables(sess=None): """Excute ``sess.run(tf.global_variables_initializer())`` for TF12+ or sess.run(tf.initialize_all_variables()) for TF11. Parameters ---------- sess : a Session """ assert sess is not None try: # TF12 sess.run(tf.global_variables_initializer()) except: # TF11 sess.run(tf.initialize_all_variables())
## Basic layer
[docs]class Layer(object): """ The :class:`Layer` class represents a single layer of a neural network. It should be subclassed when implementing new types of layers. Because each layer can keep track of the layer(s) feeding into it, a network's output :class:`Layer` instance can double as a handle to the full network. Parameters ---------- inputs : a :class:`Layer` instance The `Layer` class feeding into this layer. name : a string or None An optional name to attach to this layer. """ def __init__( self, inputs = None, name ='layer' ): self.inputs = inputs scope_name=tf.get_variable_scope().name if scope_name: name = scope_name + '/' + name if (name in set_keep['_layers_name_list']) and name_reuse == False: raise Exception("Layer '%s' already exists, please choice other 'name' or reuse this layer\ \nHint : Use different name for different 'Layer' (The name is used to control parameter sharing)" % name) else: self.name = name if name not in ['', None, False]: set_keep['_layers_name_list'].append(name) def print_params(self, details=True): ''' Print all info of parameters in the network''' for i, p in enumerate(self.all_params): if details: try: # print(" param {:3}: {:15} (mean: {:<18}, median: {:<18}, std: {:<18}) {}".format(i, str(p.eval().shape), p.eval().mean(), np.median(p.eval()), p.eval().std(), p.name)) val = p.eval() print(" param {:3}: {:20} {:15} {} (mean: {:<18}, median: {:<18}, std: {:<18}) ".format(i, p.name, str(val.shape), p.dtype.name, val.mean(), np.median(val), val.std())) except Exception as e: print(str(e)) raise Exception("Hint: print params details after tl.layers.initialize_global_variables(sess) or use network.print_params(False).") else: print(" param {:3}: {:20} {:15} {}".format(i, p.name, str(p.get_shape()), p.dtype.name)) print(" num of params: %d" % self.count_params()) def print_layers(self): ''' Print all info of layers in the network ''' for i, layer in enumerate(self.all_layers): # print(" layer %d: %s" % (i, str(layer))) print(" layer {:3}: {:20} {:15} {}".format(i, layer.name, str(layer.get_shape()), layer.dtype.name)) def count_params(self): ''' Return the number of parameters in the network ''' n_params = 0 for i, p in enumerate(self.all_params): n = 1 # for s in p.eval().shape: for s in p.get_shape(): try: s = int(s) except: s = 1 if s: n = n * s n_params = n_params + n return n_params def __str__(self): # print("\nIt is a Layer class") # self.print_params(False) # self.print_layers() return " Last layer is: %s" % self.__class__.__name__
## Input layer
[docs]class InputLayer(Layer): """ The :class:`InputLayer` class is the starting layer of a neural network. Parameters ---------- inputs : a placeholder or tensor The input tensor data. name : a string or None An optional name to attach to this layer. """ def __init__( self, inputs = None, name ='input_layer' ): Layer.__init__(self, inputs=inputs, name=name) print(" [TL] InputLayer %s: %s" % (self.name, inputs.get_shape())) self.outputs = inputs self.all_layers = [] self.all_params = [] self.all_drop = {}
## OneHot layer
[docs]class OneHotInputLayer(Layer): """ The :class:`OneHotInputLayer` class is the starting layer of a neural network, see ``tf.one_hot``. Parameters ---------- inputs : a placeholder or tensor The input tensor data. name : a string or None An optional name to attach to this layer. depth : If the input indices is rank N, the output will have rank N+1. The new axis is created at dimension axis (default: the new axis is appended at the end). on_value : If on_value is not provided, it will default to the value 1 with type dtype. default, None off_value : If off_value is not provided, it will default to the value 0 with type dtype. default, None axis : default, None dtype : default, None """ def __init__( self, inputs = None, depth = None, on_value = None, off_value = None, axis = None, dtype=None, name ='input_layer' ): Layer.__init__(self, inputs=inputs, name=name) assert depth != None, "depth is not given" print(" [TL]:Instantiate OneHotInputLayer %s: %s" % (self.name, inputs.get_shape())) self.outputs = tf.one_hot(inputs, depth, on_value=on_value, off_value=off_value, axis=axis, dtype=dtype) self.all_layers = [] self.all_params = [] self.all_drop = {}
## Word Embedding Input layer
[docs]class Word2vecEmbeddingInputlayer(Layer): """ The :class:`Word2vecEmbeddingInputlayer` class is a fully connected layer, for Word Embedding. Words are input as integer index. The output is the embedded word vector. Parameters ---------- inputs : placeholder For word inputs. integer index format. train_labels : placeholder For word labels. integer index format. vocabulary_size : int The size of vocabulary, number of words. embedding_size : int The number of embedding dimensions. num_sampled : int The Number of negative examples for NCE loss. nce_loss_args : a dictionary The arguments for tf.nn.nce_loss() E_init : embedding initializer The initializer for initializing the embedding matrix. E_init_args : a dictionary The arguments for embedding initializer nce_W_init : NCE decoder biases initializer The initializer for initializing the nce decoder weight matrix. nce_W_init_args : a dictionary The arguments for initializing the nce decoder weight matrix. nce_b_init : NCE decoder biases initializer The initializer for tf.get_variable() of the nce decoder bias vector. nce_b_init_args : a dictionary The arguments for tf.get_variable() of the nce decoder bias vector. name : a string or None An optional name to attach to this layer. Variables -------------- nce_cost : a tensor The NCE loss. outputs : a tensor The outputs of embedding layer. normalized_embeddings : tensor Normalized embedding matrix Examples -------- - Without TensorLayer : see tensorflow/examples/tutorials/word2vec/word2vec_basic.py >>> train_inputs = tf.placeholder(tf.int32, shape=[batch_size]) >>> train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1]) >>> embeddings = tf.Variable( ... tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0)) >>> embed = tf.nn.embedding_lookup(embeddings, train_inputs) >>> nce_weights = tf.Variable( ... tf.truncated_normal([vocabulary_size, embedding_size], ... stddev=1.0 / math.sqrt(embedding_size))) >>> nce_biases = tf.Variable(tf.zeros([vocabulary_size])) >>> cost = tf.reduce_mean( ... tf.nn.nce_loss(weights=nce_weights, biases=nce_biases, ... inputs=embed, labels=train_labels, ... num_sampled=num_sampled, num_classes=vocabulary_size, ... num_true=1)) - With TensorLayer : see tutorial_word2vec_basic.py >>> train_inputs = tf.placeholder(tf.int32, shape=[batch_size]) >>> train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1]) >>> emb_net = tl.layers.Word2vecEmbeddingInputlayer( ... inputs = train_inputs, ... train_labels = train_labels, ... vocabulary_size = vocabulary_size, ... embedding_size = embedding_size, ... num_sampled = num_sampled, ... name ='word2vec_layer', ... ) >>> cost = emb_net.nce_cost >>> train_params = emb_net.all_params >>> train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize( ... cost, var_list=train_params) >>> normalized_embeddings = emb_net.normalized_embeddings References ---------- - `tensorflow/examples/tutorials/word2vec/word2vec_basic.py <https://github.com/tensorflow/tensorflow/blob/r0.7/tensorflow/examples/tutorials/word2vec/word2vec_basic.py>`_ """ def __init__( self, inputs = None, train_labels = None, vocabulary_size = 80000, embedding_size = 200, num_sampled = 64, nce_loss_args = {}, E_init = tf.random_uniform_initializer(minval=-1.0, maxval=1.0), E_init_args = {}, nce_W_init = tf.truncated_normal_initializer(stddev=0.03), nce_W_init_args = {}, nce_b_init = tf.constant_initializer(value=0.0), nce_b_init_args = {}, name ='word2vec_layer', ): Layer.__init__(self, name=name) self.inputs = inputs print(" [TL] Word2vecEmbeddingInputlayer %s: (%d, %d)" % (self.name, vocabulary_size, embedding_size)) # Look up embeddings for inputs. # Note: a row of 'embeddings' is the vector representation of a word. # for the sake of speed, it is better to slice the embedding matrix # instead of transfering a word id to one-hot-format vector and then # multiply by the embedding matrix. # embed is the outputs of the hidden layer (embedding layer), it is a # row vector with 'embedding_size' values. with tf.variable_scope(name) as vs: embeddings = tf.get_variable(name='embeddings', shape=(vocabulary_size, embedding_size), initializer=E_init, **E_init_args) embed = tf.nn.embedding_lookup(embeddings, self.inputs) # Construct the variables for the NCE loss (i.e. negative sampling) nce_weights = tf.get_variable(name='nce_weights', shape=(vocabulary_size, embedding_size), initializer=nce_W_init, **nce_W_init_args) nce_biases = tf.get_variable(name='nce_biases', shape=(vocabulary_size), initializer=nce_b_init, **nce_b_init_args) # Compute the average NCE loss for the batch. # tf.nce_loss automatically draws a new sample of the negative labels # each time we evaluate the loss. self.nce_cost = tf.reduce_mean( tf.nn.nce_loss(weights=nce_weights, biases=nce_biases, inputs=embed, labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size, **nce_loss_args)) self.outputs = embed self.normalized_embeddings = tf.nn.l2_normalize(embeddings, 1) self.all_layers = [self.outputs] self.all_params = [embeddings, nce_weights, nce_biases] self.all_drop = {}
[docs]class EmbeddingInputlayer(Layer): """ The :class:`EmbeddingInputlayer` class is a fully connected layer, for Word Embedding. Words are input as integer index. The output is the embedded word vector. If you have a pre-train matrix, you can assign the matrix into it. To train a word embedding matrix, you can used class:`Word2vecEmbeddingInputlayer`. Note that, do not update this embedding matrix. Parameters ---------- inputs : placeholder For word inputs. integer index format. a 2D tensor : [batch_size, num_steps(num_words)] vocabulary_size : int The size of vocabulary, number of words. embedding_size : int The number of embedding dimensions. E_init : embedding initializer The initializer for initializing the embedding matrix. E_init_args : a dictionary The arguments for embedding initializer name : a string or None An optional name to attach to this layer. Variables ------------ outputs : a tensor The outputs of embedding layer. the outputs 3D tensor : [batch_size, num_steps(num_words), embedding_size] Examples -------- >>> vocabulary_size = 50000 >>> embedding_size = 200 >>> model_file_name = "model_word2vec_50k_200" >>> batch_size = None ... >>> all_var = tl.files.load_npy_to_any(name=model_file_name+'.npy') >>> data = all_var['data']; count = all_var['count'] >>> dictionary = all_var['dictionary'] >>> reverse_dictionary = all_var['reverse_dictionary'] >>> tl.files.save_vocab(count, name='vocab_'+model_file_name+'.txt') >>> del all_var, data, count ... >>> load_params = tl.files.load_npz(name=model_file_name+'.npz') >>> x = tf.placeholder(tf.int32, shape=[batch_size]) >>> y_ = tf.placeholder(tf.int32, shape=[batch_size, 1]) >>> emb_net = tl.layers.EmbeddingInputlayer( ... inputs = x, ... vocabulary_size = vocabulary_size, ... embedding_size = embedding_size, ... name ='embedding_layer') >>> tl.layers.initialize_global_variables(sess) >>> tl.files.assign_params(sess, [load_params[0]], emb_net) >>> word = b'hello' >>> word_id = dictionary[word] >>> print('word_id:', word_id) ... 6428 ... >>> words = [b'i', b'am', b'hao', b'dong'] >>> word_ids = tl.files.words_to_word_ids(words, dictionary) >>> context = tl.files.word_ids_to_words(word_ids, reverse_dictionary) >>> print('word_ids:', word_ids) ... [72, 1226, 46744, 20048] >>> print('context:', context) ... [b'i', b'am', b'hao', b'dong'] ... >>> vector = sess.run(emb_net.outputs, feed_dict={x : [word_id]}) >>> print('vector:', vector.shape) ... (1, 200) >>> vectors = sess.run(emb_net.outputs, feed_dict={x : word_ids}) >>> print('vectors:', vectors.shape) ... (4, 200) """ def __init__( self, inputs = None, vocabulary_size = 80000, embedding_size = 200, E_init = tf.random_uniform_initializer(-0.1, 0.1), E_init_args = {}, name ='embedding_layer', ): Layer.__init__(self, name=name) self.inputs = inputs print(" [TL] EmbeddingInputlayer %s: (%d, %d)" % (self.name, vocabulary_size, embedding_size)) with tf.variable_scope(name) as vs: embeddings = tf.get_variable(name='embeddings', shape=(vocabulary_size, embedding_size), initializer=E_init, **E_init_args) embed = tf.nn.embedding_lookup(embeddings, self.inputs) self.outputs = embed self.all_layers = [self.outputs] self.all_params = [embeddings] self.all_drop = {}
## Dense layer
[docs]class DenseLayer(Layer): """ The :class:`DenseLayer` class is a fully connected layer. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. n_units : int The number of units of the layer. act : activation function The function that is applied to the layer activations. W_init : weights initializer The initializer for initializing the weight matrix. b_init : biases initializer or None The initializer for initializing the bias vector. If None, skip biases. W_init_args : dictionary The arguments for the weights tf.get_variable. b_init_args : dictionary The arguments for the biases tf.get_variable. name : a string or None An optional name to attach to this layer. Examples -------- >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.DenseLayer( ... network, ... n_units=800, ... act = tf.nn.relu, ... W_init=tf.truncated_normal_initializer(stddev=0.1), ... name ='relu_layer' ... ) >>> Without TensorLayer, you can do as follow. >>> W = tf.Variable( ... tf.random_uniform([n_in, n_units], -1.0, 1.0), name='W') >>> b = tf.Variable(tf.zeros(shape=[n_units]), name='b') >>> y = tf.nn.relu(tf.matmul(inputs, W) + b) Notes ----- If the input to this layer has more than two axes, it need to flatten the input by using :class:`FlattenLayer` in this case. """ def __init__( self, layer = None, n_units = 100, act = tf.identity, W_init = tf.truncated_normal_initializer(stddev=0.1), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='dense_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs if self.inputs.get_shape().ndims != 2: raise Exception("The input dimension must be rank 2, please reshape or flatten it") n_in = int(self.inputs.get_shape()[-1]) self.n_units = n_units print(" [TL] DenseLayer %s: %d %s" % (self.name, self.n_units, act.__name__)) with tf.variable_scope(name) as vs: W = tf.get_variable(name='W', shape=(n_in, n_units), initializer=W_init, **W_init_args ) if b_init is not None: try: b = tf.get_variable(name='b', shape=(n_units), initializer=b_init, **b_init_args ) except: # If initializer is a constant, do not specify shape. b = tf.get_variable(name='b', initializer=b_init, **b_init_args ) self.outputs = act(tf.matmul(self.inputs, W) + b) else: self.outputs = act(tf.matmul(self.inputs, W)) # Hint : list(), dict() is pass by value (shallow), without them, it is # pass by reference. self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) if b_init is not None: self.all_params.extend( [W, b] ) else: self.all_params.extend( [W] )
[docs]class ReconLayer(DenseLayer): """ The :class:`ReconLayer` class is a reconstruction layer `DenseLayer` which use to pre-train a `DenseLayer`. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. x_recon : tensorflow variable The variables used for reconstruction. name : a string or None An optional name to attach to this layer. n_units : int The number of units of the layer, should be equal to x_recon act : activation function The activation function that is applied to the reconstruction layer. Normally, for sigmoid layer, the reconstruction activation is sigmoid; for rectifying layer, the reconstruction activation is softplus. Examples -------- >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.DenseLayer(network, n_units=196, ... act=tf.nn.sigmoid, name='sigmoid1') >>> recon_layer1 = tl.layers.ReconLayer(network, x_recon=x, n_units=784, ... act=tf.nn.sigmoid, name='recon_layer1') >>> recon_layer1.pretrain(sess, x=x, X_train=X_train, X_val=X_val, ... denoise_name=None, n_epoch=1200, batch_size=128, ... print_freq=10, save=True, save_name='w1pre_') Methods ------- pretrain(self, sess, x, X_train, X_val, denoise_name=None, n_epoch=100, batch_size=128, print_freq=10, save=True, save_name='w1pre_') Start to pre-train the parameters of previous DenseLayer. Notes ----- The input layer should be `DenseLayer` or a layer has only one axes. You may need to modify this part to define your own cost function. By default, the cost is implemented as follow: - For sigmoid layer, the implementation can be `UFLDL <http://deeplearning.stanford.edu/wiki/index.php/UFLDL_Tutorial>`_ - For rectifying layer, the implementation can be `Glorot (2011). Deep Sparse Rectifier Neural Networks <http://doi.org/10.1.1.208.6449>`_ """ def __init__( self, layer = None, x_recon = None, name = 'recon_layer', n_units = 784, act = tf.nn.softplus, ): DenseLayer.__init__(self, layer=layer, n_units=n_units, act=act, name=name) print(" [TL] %s is a ReconLayer" % self.name) # y : reconstruction outputs; train_params : parameters to train # Note that: train_params = [W_encoder, b_encoder, W_decoder, b_encoder] y = self.outputs self.train_params = self.all_params[-4:] # ===================================================================== # # You need to modify the below cost function and optimizer so as to # implement your own pre-train method. # # ===================================================================== lambda_l2_w = 0.004 learning_rate = 0.0001 print(" lambda_l2_w: %f" % lambda_l2_w) print(" learning_rate: %f" % learning_rate) # Mean-squre-error i.e. quadratic-cost mse = tf.reduce_sum(tf.squared_difference(y, x_recon), 1) mse = tf.reduce_mean(mse) # in theano: mse = ((y - x) ** 2 ).sum(axis=1).mean() # mse = tf.reduce_mean(tf.reduce_sum(tf.square(tf.sub(y, x_recon)), 1)) # mse = tf.reduce_mean(tf.squared_difference(y, x_recon)) # <haodong>: Error # mse = tf.sqrt(tf.reduce_mean(tf.square(y - x_recon))) # <haodong>: Error # Cross-entropy # ce = cost.cross_entropy(y, x_recon) # <haodong>: list , list , Error (only be used for softmax output) # ce = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y, x_recon)) # <haodong>: list , list , Error (only be used for softmax output) # ce = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(y, x_recon)) # <haodong>: list , index , Error (only be used for softmax output) L2_w = tf.contrib.layers.l2_regularizer(lambda_l2_w)(self.train_params[0]) \ + tf.contrib.layers.l2_regularizer(lambda_l2_w)(self.train_params[2]) # faster than the code below # L2_w = lambda_l2_w * tf.reduce_mean(tf.square(self.train_params[0])) + lambda_l2_w * tf.reduce_mean( tf.square(self.train_params[2])) # DropNeuro P_o = cost.lo_regularizer(0.03)(self.train_params[0]) # + cost.lo_regularizer(0.5)(self.train_params[2]) # <haodong>: if add lo on decoder, no neuron will be broken P_i = cost.li_regularizer(0.03)(self.train_params[0]) # + cost.li_regularizer(0.001)(self.train_params[2]) # L1 of activation outputs activation_out = self.all_layers[-2] L1_a = 0.001 * tf.reduce_mean(activation_out) # <haodong>: theano: T.mean( self.a[i] ) # some neuron are broken, white and black # L1_a = 0.001 * tf.reduce_mean( tf.reduce_sum(activation_out, 0) ) # <haodong>: some neuron are broken, white and black # L1_a = 0.001 * 100 * tf.reduce_mean( tf.reduce_sum(activation_out, 1) ) # <haodong>: some neuron are broken, white and black # KL Divergence beta = 4 rho = 0.15 p_hat = tf.reduce_mean(activation_out, 0) # theano: p_hat = T.mean( self.a[i], axis=0 ) try: ## TF1.0 KLD = beta * tf.reduce_sum( rho * tf.log(tf.divide(rho, p_hat)) + (1- rho) * tf.log((1- rho)/ (tf.subtract(float(1), p_hat))) ) except: ## TF0.12 KLD = beta * tf.reduce_sum( rho * tf.log(tf.div(rho, p_hat)) + (1- rho) * tf.log((1- rho)/ (tf.sub(float(1), p_hat))) ) # KLD = beta * tf.reduce_sum( rho * tf.log(rho/ p_hat) + (1- rho) * tf.log((1- rho)/(1- p_hat)) ) # theano: L1_a = l1_a[i] * T.sum( rho[i] * T.log(rho[i]/ p_hat) + (1- rho[i]) * T.log((1- rho[i])/(1- p_hat)) ) # Total cost if act == tf.nn.softplus: print(' use: mse, L2_w, L1_a') self.cost = mse + L1_a + L2_w elif act == tf.nn.sigmoid: # ---------------------------------------------------- # Cross-entropy was used in Denoising AE # print(' use: ce, L2_w, KLD') # self.cost = ce + L2_w + KLD # ---------------------------------------------------- # Mean-squared-error was used in Vanilla AE print(' use: mse, L2_w, KLD') self.cost = mse + L2_w + KLD # ---------------------------------------------------- # Add DropNeuro penalty (P_o) can remove neurons of AE # print(' use: mse, L2_w, KLD, P_o') # self.cost = mse + L2_w + KLD + P_o # ---------------------------------------------------- # Add DropNeuro penalty (P_i) can remove neurons of previous layer # If previous layer is InputLayer, it means remove useless features # print(' use: mse, L2_w, KLD, P_i') # self.cost = mse + L2_w + KLD + P_i else: raise Exception("Don't support the given reconstruct activation function") self.train_op = tf.train.AdamOptimizer(learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-08, use_locking=False).minimize(self.cost, var_list=self.train_params) # self.train_op = tf.train.GradientDescentOptimizer(1.0).minimize(self.cost, var_list=self.train_params) def pretrain(self, sess, x, X_train, X_val, denoise_name=None, n_epoch=100, batch_size=128, print_freq=10, save=True, save_name='w1pre_'): # ==================================================== # # You need to modify the cost function in __init__() so as to # get your own pre-train method. # # ==================================================== print(" [*] %s start pretrain" % self.name) print(" batch_size: %d" % batch_size) if denoise_name: print(" denoising layer keep: %f" % self.all_drop[set_keep[denoise_name]]) dp_denoise = self.all_drop[set_keep[denoise_name]] else: print(" no denoising layer") for epoch in range(n_epoch): start_time = time.time() for X_train_a, _ in iterate.minibatches(X_train, X_train, batch_size, shuffle=True): dp_dict = utils.dict_to_one( self.all_drop ) if denoise_name: dp_dict[set_keep[denoise_name]] = dp_denoise feed_dict = {x: X_train_a} feed_dict.update(dp_dict) sess.run(self.train_op, feed_dict=feed_dict) if epoch + 1 == 1 or (epoch + 1) % print_freq == 0: print("Epoch %d of %d took %fs" % (epoch + 1, n_epoch, time.time() - start_time)) train_loss, n_batch = 0, 0 for X_train_a, _ in iterate.minibatches(X_train, X_train, batch_size, shuffle=True): dp_dict = utils.dict_to_one( self.all_drop ) feed_dict = {x: X_train_a} feed_dict.update(dp_dict) err = sess.run(self.cost, feed_dict=feed_dict) train_loss += err n_batch += 1 print(" train loss: %f" % (train_loss/ n_batch)) val_loss, n_batch = 0, 0 for X_val_a, _ in iterate.minibatches(X_val, X_val, batch_size, shuffle=True): dp_dict = utils.dict_to_one( self.all_drop ) feed_dict = {x: X_val_a} feed_dict.update(dp_dict) err = sess.run(self.cost, feed_dict=feed_dict) val_loss += err n_batch += 1 print(" val loss: %f" % (val_loss/ n_batch)) if save: try: visualize.W(self.train_params[0].eval(), second=10, saveable=True, shape=[28,28], name=save_name+str(epoch+1), fig_idx=2012) files.save_npz([self.all_params[0]] , name=save_name+str(epoch+1)+'.npz') except: raise Exception("You should change the visualize.W() in ReconLayer.pretrain(), if you want to save the feature images for different dataset")
## Noise layer
[docs]class DropoutLayer(Layer): """ The :class:`DropoutLayer` class is a noise layer which randomly set some values to zero by a given keeping probability. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. keep : float The keeping probability, the lower more values will be set to zero. is_fix : boolean Default False, if True, the keeping probability is fixed and cannot be changed via feed_dict. is_train : boolean If False, skip this layer, default is True. seed : int or None An integer or None to create random seed. name : a string or None An optional name to attach to this layer. Examples -------- - Define network >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.DropoutLayer(network, keep=0.8, name='drop1') >>> network = tl.layers.DenseLayer(network, n_units=800, act = tf.nn.relu, name='relu1') >>> ... - For training, enable dropout as follow. >>> feed_dict = {x: X_train_a, y_: y_train_a} >>> feed_dict.update( network.all_drop ) # enable noise layers >>> sess.run(train_op, feed_dict=feed_dict) >>> ... - For testing, disable dropout as follow. >>> dp_dict = tl.utils.dict_to_one( network.all_drop ) # disable noise layers >>> feed_dict = {x: X_val_a, y_: y_val_a} >>> feed_dict.update(dp_dict) >>> err, ac = sess.run([cost, acc], feed_dict=feed_dict) >>> ... Notes ------- - A frequent question regarding :class:`DropoutLayer` is that why it donot have `is_train` like :class:`BatchNormLayer`. In many simple cases, user may find it is better to use one inference instead of two inferences for training and testing seperately, :class:`DropoutLayer` allows you to control the dropout rate via `feed_dict`. However, you can fix the keeping probability by setting `is_fix` to True. """ def __init__( self, layer = None, keep = 0.5, is_fix = False, is_train = True, seed = None, name = 'dropout_layer', ): Layer.__init__(self, name=name) if is_train is False: print(" [TL] skip DropoutLayer") self.outputs = layer.outputs self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) else: self.inputs = layer.outputs print(" [TL] DropoutLayer %s: keep:%f is_fix:%s" % (self.name, keep, is_fix)) # The name of placeholder for keep_prob is the same with the name # of the Layer. if is_fix: self.outputs = tf.nn.dropout(self.inputs, keep, seed=seed, name=name) else: set_keep[name] = tf.placeholder(tf.float32) self.outputs = tf.nn.dropout(self.inputs, set_keep[name], seed=seed, name=name) # 1.2 self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) if is_fix is False: self.all_drop.update( {set_keep[name]: keep} ) self.all_layers.extend( [self.outputs] )
# print(set_keep[name]) # Tensor("Placeholder_2:0", dtype=float32) # print(denoising1) # Tensor("Placeholder_2:0", dtype=float32) # print(self.all_drop[denoising1]) # 0.8 # # https://www.tensorflow.org/versions/r0.8/tutorials/mnist/tf/index.html # The optional feed_dict argument allows the caller to override the # value of tensors in the graph. Each key in feed_dict can be one of # the following types: # If the key is a Tensor, the value may be a Python scalar, string, # list, or numpy ndarray that can be converted to the same dtype as that # tensor. Additionally, if the key is a placeholder, the shape of the # value will be checked for compatibility with the placeholder. # If the key is a SparseTensor, the value should be a SparseTensorValue.
[docs]class GaussianNoiseLayer(Layer): """ The :class:`GaussianNoiseLayer` class is noise layer that adding noise with normal distribution to the activation. Parameters ------------ layer : a :class:`Layer` instance The `Layer` class feeding into this layer. mean : float stddev : float is_train : boolean If False, skip this layer, default is True. seed : int or None An integer or None to create random seed. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, mean = 0.0, stddev = 1.0, is_train = True, seed = None, name = 'gaussian_noise_layer', ): Layer.__init__(self, name=name) if is_train is False: print(" [TL] skip GaussianNoiseLayer") self.outputs = layer.outputs self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) else: self.inputs = layer.outputs print(" [TL] GaussianNoiseLayer %s: mean:%f stddev:%f" % (self.name, mean, stddev)) with tf.variable_scope(name) as vs: # noise = np.random.normal(0.0 , sigma , tf.to_int64(self.inputs).get_shape()) noise = tf.random_normal(shape = self.inputs.get_shape(), mean=mean, stddev=stddev, seed=seed) self.outputs = self.inputs + noise self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop)
[docs]class DropconnectDenseLayer(Layer): """ The :class:`DropconnectDenseLayer` class is ``DenseLayer`` with DropConnect behaviour which randomly remove connection between this layer to previous layer by a given keeping probability. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. keep : float The keeping probability, the lower more values will be set to zero. n_units : int The number of units of the layer. act : activation function The function that is applied to the layer activations. W_init : weights initializer The initializer for initializing the weight matrix. b_init : biases initializer The initializer for initializing the bias vector. W_init_args : dictionary The arguments for the weights tf.get_variable(). b_init_args : dictionary The arguments for the biases tf.get_variable(). name : a string or None An optional name to attach to this layer. Examples -------- >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.DropconnectDenseLayer(network, keep = 0.8, ... n_units=800, act = tf.nn.relu, name='dropconnect_relu1') >>> network = tl.layers.DropconnectDenseLayer(network, keep = 0.5, ... n_units=800, act = tf.nn.relu, name='dropconnect_relu2') >>> network = tl.layers.DropconnectDenseLayer(network, keep = 0.5, ... n_units=10, act = tl.activation.identity, name='output_layer') References ---------- - `Wan, L. (2013). Regularization of neural networks using dropconnect <http://machinelearning.wustl.edu/mlpapers/papers/icml2013_wan13>`_ """ def __init__( self, layer = None, keep = 0.5, n_units = 100, act = tf.identity, W_init = tf.truncated_normal_initializer(stddev=0.1), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='dropconnect_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs if self.inputs.get_shape().ndims != 2: raise Exception("The input dimension must be rank 2") n_in = int(self.inputs.get_shape()[-1]) self.n_units = n_units print(" [TL] DropconnectDenseLayer %s: %d %s" % (self.name, self.n_units, act.__name__)) with tf.variable_scope(name) as vs: W = tf.get_variable(name='W', shape=(n_in, n_units), initializer=W_init, **W_init_args ) b = tf.get_variable(name='b', shape=(n_units), initializer=b_init, **b_init_args ) self.outputs = act(tf.matmul(self.inputs, W) + b)#, name=name) # 1.2 set_keep[name] = tf.placeholder(tf.float32) W_dropcon = tf.nn.dropout(W, set_keep[name]) self.outputs = act(tf.matmul(self.inputs, W_dropcon) + b) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_drop.update( {set_keep[name]: keep} ) self.all_layers.extend( [self.outputs] ) self.all_params.extend( [W, b] )
## Convolutional layer (Pro)
[docs]class Conv1dLayer(Layer): """ The :class:`Conv1dLayer` class is a 1D CNN layer, see `tf.nn.convolution <https://www.tensorflow.org/api_docs/python/tf/nn/convolution>`_. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer, [batch, in_width, in_channels]. act : activation function, None for identity. shape : list of shape shape of the filters, [filter_length, in_channels, out_channels]. stride : an int. The number of entries by which the filter is moved right at each step. dilation_rate : an int. Specifies the filter upsampling/input downsampling rate. padding : a string from: "SAME", "VALID". The type of padding algorithm to use. use_cudnn_on_gpu : An optional bool. Defaults to True. data_format : As it is 1D conv, default is 'NWC'. W_init : weights initializer The initializer for initializing the weight matrix. b_init : biases initializer or None The initializer for initializing the bias vector. If None, skip biases. W_init_args : dictionary The arguments for the weights tf.get_variable(). b_init_args : dictionary The arguments for the biases tf.get_variable(). name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, act = tf.identity, shape = [5, 1, 5], stride = 1, dilation_rate = 1, padding='SAME', use_cudnn_on_gpu=None, data_format='NWC', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='cnn_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] Conv1dLayer %s: shape:%s stride:%s pad:%s act:%s" % (self.name, str(shape), str(stride), padding, act.__name__)) if act is None: act = tf.identity with tf.variable_scope(name) as vs: W = tf.get_variable(name='W_conv1d', shape=shape, initializer=W_init, **W_init_args ) self.outputs = tf.nn.convolution( self.inputs, W, strides=(stride,), padding=padding, dilation_rate=(dilation_rate,), data_format=data_format ) #1.2 if b_init: b = tf.get_variable(name='b_conv1d', shape=(shape[-1]), initializer=b_init, **b_init_args ) self.outputs = self.outputs + b self.outputs = act(self.outputs) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) if b_init: self.all_params.extend( [W, b] ) else: self.all_params.extend( [W] )
[docs]class Conv2dLayer(Layer): """ The :class:`Conv2dLayer` class is a 2D CNN layer, see `tf.nn.conv2d <https://www.tensorflow.org/versions/master/api_docs/python/nn.html#conv2d>`_. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. act : activation function The function that is applied to the layer activations. shape : list of shape shape of the filters, [filter_height, filter_width, in_channels, out_channels]. strides : a list of ints. The stride of the sliding window for each dimension of input.\n It Must be in the same order as the dimension specified with format. padding : a string from: "SAME", "VALID". The type of padding algorithm to use. W_init : weights initializer The initializer for initializing the weight matrix. b_init : biases initializer or None The initializer for initializing the bias vector. If None, skip biases. W_init_args : dictionary The arguments for the weights tf.get_variable(). b_init_args : dictionary The arguments for the biases tf.get_variable(). use_cudnn_on_gpu : an optional string from: "NHWC", "NCHW". Defaults to "NHWC". data_format : an optional bool. Defaults to True. name : a string or None An optional name to attach to this layer. Notes ------ - shape = [h, w, the number of output channel of previous layer, the number of output channels] - the number of output channel of a layer is its last dimension. Examples -------- >>> x = tf.placeholder(tf.float32, shape=[None, 28, 28, 1]) >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.Conv2dLayer(network, ... act = tf.nn.relu, ... shape = [5, 5, 1, 32], # 32 features for each 5x5 patch ... strides=[1, 1, 1, 1], ... padding='SAME', ... W_init=tf.truncated_normal_initializer(stddev=5e-2), ... W_init_args={}, ... b_init = tf.constant_initializer(value=0.0), ... b_init_args = {}, ... name ='cnn_layer1') # output: (?, 28, 28, 32) >>> network = tl.layers.PoolLayer(network, ... ksize=[1, 2, 2, 1], ... strides=[1, 2, 2, 1], ... padding='SAME', ... pool = tf.nn.max_pool, ... name ='pool_layer1',) # output: (?, 14, 14, 32) >>> Without TensorLayer, you can implement 2d convolution as follow. >>> W = tf.Variable(W_init(shape=[5, 5, 1, 32], ), name='W_conv') >>> b = tf.Variable(b_init(shape=[32], ), name='b_conv') >>> outputs = tf.nn.relu( tf.nn.conv2d(inputs, W, ... strides=[1, 1, 1, 1], ... padding='SAME') + b ) """ def __init__( self, layer = None, act = tf.identity, shape = [5, 5, 1, 100], strides=[1, 1, 1, 1], padding='SAME', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, use_cudnn_on_gpu = None, data_format = None, name ='cnn_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] Conv2dLayer %s: shape:%s strides:%s pad:%s act:%s" % (self.name, str(shape), str(strides), padding, act.__name__)) with tf.variable_scope(name) as vs: W = tf.get_variable(name='W_conv2d', shape=shape, initializer=W_init, **W_init_args ) if b_init: b = tf.get_variable(name='b_conv2d', shape=(shape[-1]), initializer=b_init, **b_init_args ) self.outputs = act( tf.nn.conv2d(self.inputs, W, strides=strides, padding=padding, use_cudnn_on_gpu=use_cudnn_on_gpu, data_format=data_format) + b ) else: self.outputs = act( tf.nn.conv2d(self.inputs, W, strides=strides, padding=padding, use_cudnn_on_gpu=use_cudnn_on_gpu, data_format=data_format)) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) if b_init: self.all_params.extend( [W, b] ) else: self.all_params.extend( [W] )
[docs]class DeConv2dLayer(Layer): """ The :class:`DeConv2dLayer` class is deconvolutional 2D layer, see `tf.nn.conv2d_transpose <https://www.tensorflow.org/versions/master/api_docs/python/nn.html#conv2d_transpose>`_. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. act : activation function The function that is applied to the layer activations. shape : list of shape shape of the filters, [height, width, output_channels, in_channels], filter's in_channels dimension must match that of value. output_shape : list of output shape representing the output shape of the deconvolution op. strides : a list of ints. The stride of the sliding window for each dimension of the input tensor. padding : a string from: "SAME", "VALID". The type of padding algorithm to use. W_init : weights initializer The initializer for initializing the weight matrix. b_init : biases initializer The initializer for initializing the bias vector. If None, skip biases. W_init_args : dictionary The arguments for the weights initializer. b_init_args : dictionary The arguments for the biases initializer. name : a string or None An optional name to attach to this layer. Notes ----- - shape = [h, w, the number of output channels of this layer, the number of output channel of previous layer] - output_shape = [batch_size, any, any, the number of output channels of this layer] - the number of output channel of a layer is its last dimension. Examples --------- - A part of the generator in DCGAN example >>> batch_size = 64 >>> inputs = tf.placeholder(tf.float32, [batch_size, 100], name='z_noise') >>> net_in = tl.layers.InputLayer(inputs, name='g/in') >>> net_h0 = tl.layers.DenseLayer(net_in, n_units = 8192, ... W_init = tf.random_normal_initializer(stddev=0.02), ... act = tf.identity, name='g/h0/lin') >>> print(net_h0.outputs._shape) ... (64, 8192) >>> net_h0 = tl.layers.ReshapeLayer(net_h0, shape = [-1, 4, 4, 512], name='g/h0/reshape') >>> net_h0 = tl.layers.BatchNormLayer(net_h0, act=tf.nn.relu, is_train=is_train, name='g/h0/batch_norm') >>> print(net_h0.outputs._shape) ... (64, 4, 4, 512) >>> net_h1 = tl.layers.DeConv2dLayer(net_h0, ... shape = [5, 5, 256, 512], ... output_shape = [batch_size, 8, 8, 256], ... strides=[1, 2, 2, 1], ... act=tf.identity, name='g/h1/decon2d') >>> net_h1 = tl.layers.BatchNormLayer(net_h1, act=tf.nn.relu, is_train=is_train, name='g/h1/batch_norm') >>> print(net_h1.outputs._shape) ... (64, 8, 8, 256) - U-Net >>> .... >>> conv10 = tl.layers.Conv2dLayer(conv9, act=tf.nn.relu, ... shape=[3,3,1024,1024], strides=[1,1,1,1], padding='SAME', ... W_init=w_init, b_init=b_init, name='conv10') >>> print(conv10.outputs) ... (batch_size, 32, 32, 1024) >>> deconv1 = tl.layers.DeConv2dLayer(conv10, act=tf.nn.relu, ... shape=[3,3,512,1024], strides=[1,2,2,1], output_shape=[batch_size,64,64,512], ... padding='SAME', W_init=w_init, b_init=b_init, name='devcon1_1') """ def __init__( self, layer = None, act = tf.identity, shape = [3, 3, 128, 256], output_shape = [1, 256, 256, 128], strides = [1, 2, 2, 1], padding = 'SAME', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='decnn2d_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] DeConv2dLayer %s: shape:%s out_shape:%s strides:%s pad:%s act:%s" % (self.name, str(shape), str(output_shape), str(strides), padding, act.__name__)) # print(" DeConv2dLayer: Untested") with tf.variable_scope(name) as vs: W = tf.get_variable(name='W_deconv2d', shape=shape, initializer=W_init, **W_init_args ) if b_init: b = tf.get_variable(name='b_deconv2d', shape=(shape[-2]), initializer=b_init, **b_init_args ) self.outputs = act( tf.nn.conv2d_transpose(self.inputs, W, output_shape=output_shape, strides=strides, padding=padding) + b ) else: self.outputs = act( tf.nn.conv2d_transpose(self.inputs, W, output_shape=output_shape, strides=strides, padding=padding)) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) if b_init: self.all_params.extend( [W, b] ) else: self.all_params.extend( [W] )
[docs]class Conv3dLayer(Layer): """ The :class:`Conv3dLayer` class is a 3D CNN layer, see `tf.nn.conv3d <https://www.tensorflow.org/versions/master/api_docs/python/nn.html#conv3d>`_. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. act : activation function The function that is applied to the layer activations. shape : list of shape shape of the filters, [filter_depth, filter_height, filter_width, in_channels, out_channels]. strides : a list of ints. 1-D of length 4. The stride of the sliding window for each dimension of input. Must be in the same order as the dimension specified with format. padding : a string from: "SAME", "VALID". The type of padding algorithm to use. W_init : weights initializer The initializer for initializing the weight matrix. b_init : biases initializer The initializer for initializing the bias vector. W_init_args : dictionary The arguments for the weights initializer. b_init_args : dictionary The arguments for the biases initializer. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, act = tf.identity, shape = [2, 2, 2, 64, 128], strides=[1, 2, 2, 2, 1], padding='SAME', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='cnn3d_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] Conv3dLayer %s: shape:%s strides:%s pad:%s act:%s" % (self.name, str(shape), str(strides), padding, act.__name__)) with tf.variable_scope(name) as vs: # W = tf.Variable(W_init(shape=shape, **W_init_args), name='W_conv') # b = tf.Variable(b_init(shape=[shape[-1]], **b_init_args), name='b_conv') W = tf.get_variable(name='W_conv3d', shape=shape, initializer=W_init, **W_init_args ) b = tf.get_variable(name='b_conv3d', shape=(shape[-1]), initializer=b_init, **b_init_args ) self.outputs = act( tf.nn.conv3d(self.inputs, W, strides=strides, padding=padding, name=None) + b ) # self.outputs = act( tf.nn.conv3d(self.inputs, W, strides=strides, padding=padding, name=None) + b ) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( [W, b] )
[docs]class DeConv3dLayer(Layer): """The :class:`DeConv3dLayer` class is deconvolutional 3D layer, see `tf.nn.conv3d_transpose <https://www.tensorflow.org/versions/master/api_docs/python/nn.html#conv3d_transpose>`_. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. act : activation function The function that is applied to the layer activations. shape : list of shape shape of the filters, [depth, height, width, output_channels, in_channels], filter's in_channels dimension must match that of value. output_shape : list of output shape representing the output shape of the deconvolution op. strides : a list of ints. The stride of the sliding window for each dimension of the input tensor. padding : a string from: "SAME", "VALID". The type of padding algorithm to use. W_init : weights initializer The initializer for initializing the weight matrix. b_init : biases initializer The initializer for initializing the bias vector. W_init_args : dictionary The arguments for the weights initializer. b_init_args : dictionary The arguments for the biases initializer. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, act = tf.identity, shape = [2, 2, 2, 128, 256], output_shape = [1, 12, 32, 32, 128], strides = [1, 2, 2, 2, 1], padding = 'SAME', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='decnn3d_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] DeConv3dLayer %s: shape:%s out_shape:%s strides:%s pad:%s act:%s" % (self.name, str(shape), str(output_shape), str(strides), padding, act.__name__)) with tf.variable_scope(name) as vs: W = tf.get_variable(name='W_deconv3d', shape=shape, initializer=W_init, **W_init_args ) b = tf.get_variable(name='b_deconv3d', shape=(shape[-2]), initializer=b_init, **b_init_args ) self.outputs = act( tf.nn.conv3d_transpose(self.inputs, W, output_shape=output_shape, strides=strides, padding=padding) + b ) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( [W, b] )
[docs]class UpSampling2dLayer(Layer): """The :class:`UpSampling2dLayer` class is upSampling 2d layer, see `tf.image.resize_images <https://www.tensorflow.org/versions/master/api_docs/python/image/resizing#resize_images>`_. Parameters ----------- layer : a layer class with 4-D Tensor of shape [batch, height, width, channels] or 3-D Tensor of shape [height, width, channels]. size : a tuple of int or float. (height, width) scale factor or new size of height and width. is_scale : boolean, if True (default), size is scale factor, otherwise, size is number of pixels of height and width. method : 0, 1, 2, 3. ResizeMethod. Defaults to ResizeMethod.BILINEAR. - ResizeMethod.BILINEAR, Bilinear interpolation. - ResizeMethod.NEAREST_NEIGHBOR, Nearest neighbor interpolation. - ResizeMethod.BICUBIC, Bicubic interpolation. - ResizeMethod.AREA, Area interpolation. align_corners : bool. If true, exactly align all 4 corners of the input and output. Defaults to false. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, size = [], is_scale = True, method = 0, align_corners = False, name ='upsample2d_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs if len(self.inputs.get_shape()) == 3: if is_scale: size_h = size[0] * int(self.inputs.get_shape()[0]) size_w = size[1] * int(self.inputs.get_shape()[1]) size = [int(size_h), int(size_w)] elif len(self.inputs.get_shape()) == 4: if is_scale: size_h = size[0] * int(self.inputs.get_shape()[1]) size_w = size[1] * int(self.inputs.get_shape()[2]) size = [int(size_h), int(size_w)] else: raise Exception("Donot support shape %s" % self.inputs.get_shape()) print(" [TL] UpSampling2dLayer %s: is_scale:%s size:%s method:%d align_corners:%s" % (name, is_scale, size, method, align_corners)) with tf.variable_scope(name) as vs: try: self.outputs = tf.image.resize_images(self.inputs, size=size, method=method, align_corners=align_corners) except: # for TF 0.10 self.outputs = tf.image.resize_images(self.inputs, new_height=size[0], new_width=size[1], method=method, align_corners=align_corners) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
[docs]class DownSampling2dLayer(Layer): """The :class:`DownSampling2dLayer` class is downSampling 2d layer, see `tf.image.resize_images <https://www.tensorflow.org/versions/master/api_docs/python/image/resizing#resize_images>`_. Parameters ----------- layer : a layer class with 4-D Tensor of shape [batch, height, width, channels] or 3-D Tensor of shape [height, width, channels]. size : a tupe of int or float. (height, width) scale factor or new size of height and width. is_scale : boolean, if True (default), size is scale factor, otherwise, size is number of pixels of height and width. method : 0, 1, 2, 3. ResizeMethod. Defaults to ResizeMethod.BILINEAR. - ResizeMethod.BILINEAR, Bilinear interpolation. - ResizeMethod.NEAREST_NEIGHBOR, Nearest neighbor interpolation. - ResizeMethod.BICUBIC, Bicubic interpolation. - ResizeMethod.AREA, Area interpolation. align_corners : bool. If true, exactly align all 4 corners of the input and output. Defaults to false. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, size = [], is_scale = True, method = 0, align_corners = False, name ='downsample2d_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs if len(self.inputs.get_shape()) == 3: if is_scale: size_h = size[0] * int(self.inputs.get_shape()[0]) size_w = size[1] * int(self.inputs.get_shape()[1]) size = [int(size_h), int(size_w)] elif len(self.inputs.get_shape()) == 4: if is_scale: size_h = size[0] * int(self.inputs.get_shape()[1]) size_w = size[1] * int(self.inputs.get_shape()[2]) size = [int(size_h), int(size_w)] else: raise Exception("Donot support shape %s" % self.inputs.get_shape()) print(" [TL] DownSampling2dLayer %s: is_scale:%s size:%s method:%d, align_corners:%s" % (name, is_scale, size, method, align_corners)) with tf.variable_scope(name) as vs: try: self.outputs = tf.image.resize_images(self.inputs, size=size, method=method, align_corners=align_corners) except: # for TF 0.10 self.outputs = tf.image.resize_images(self.inputs, new_height=size[0], new_width=size[1], method=method, align_corners=align_corners) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
[docs]def AtrousConv1dLayer(net, n_filter=32, filter_size=2, stride=1, dilation=1, act=None, padding='SAME', use_cudnn_on_gpu=None,data_format='NWC', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {},name ='conv1d',): """Wrapper for :class:`AtrousConv1dLayer`, if you don't understand how to use :class:`Conv1dLayer`, this function may be easier. Parameters ---------- net : TensorLayer layer. n_filter : number of filter. filter_size : an int. stride : an int. dilation : an int, filter dilation size. act : None or activation function. others : see :class:`Conv1dLayer`. """ if act is None: act = tf.identity net = Conv1dLayer(layer = net, act = act, shape = [filter_size, int(net.outputs.get_shape()[-1]), n_filter], stride = stride, padding = padding, dilation_rate = dilation, use_cudnn_on_gpu = use_cudnn_on_gpu, data_format = data_format, W_init = W_init, b_init = b_init, W_init_args = W_init_args, b_init_args = b_init_args, name = name, ) return net
[docs]class AtrousConv2dLayer(Layer): """The :class:`AtrousConv2dLayer` class is Atrous convolution (a.k.a. convolution with holes or dilated convolution) 2D layer, see `tf.nn.atrous_conv2d <https://www.tensorflow.org/versions/master/api_docs/python/nn.html#atrous_conv2d>`_. Parameters ----------- layer : a layer class with 4-D Tensor of shape [batch, height, width, channels]. filters : A 4-D Tensor with the same type as value and shape [filter_height, filter_width, in_channels, out_channels]. filters' in_channels dimension must match that of value. Atrous convolution is equivalent to standard convolution with upsampled filters with effective height filter_height + (filter_height - 1) * (rate - 1) and effective width filter_width + (filter_width - 1) * (rate - 1), produced by inserting rate - 1 zeros along consecutive elements across the filters' spatial dimensions. n_filter : number of filter. filter_size : tuple (height, width) for filter size. rate : A positive int32. The stride with which we sample input values across the height and width dimensions. Equivalently, the rate by which we upsample the filter values by inserting zeros across the height and width dimensions. In the literature, the same parameter is sometimes called input stride or dilation. act : activation function, None for linear. padding : A string, either 'VALID' or 'SAME'. The padding algorithm. W_init : weights initializer. The initializer for initializing the weight matrix. b_init : biases initializer or None. The initializer for initializing the bias vector. If None, skip biases. W_init_args : dictionary. The arguments for the weights tf.get_variable(). b_init_args : dictionary. The arguments for the biases tf.get_variable(). name : a string or None, an optional name to attach to this layer. """ def __init__( self, layer = None, n_filter = 32, filter_size = (3,3), rate = 2, act = None, padding = 'SAME', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name = 'atrou2d' ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] AtrousConv2dLayer %s: n_filter:%d filter_size:%s rate:%d pad:%s act:%s" % (self.name, n_filter, filter_size, rate, padding, act.__name__)) if act is None: act = tf.identity with tf.variable_scope(name) as vs: shape = [filter_size[0], filter_size[1], int(self.inputs.get_shape()[-1]), n_filter] filters = tf.get_variable(name='filter', shape=shape, initializer=W_init, **W_init_args ) if b_init: b = tf.get_variable(name='b', shape=(n_filter), initializer=b_init, **b_init_args ) self.outputs = act(tf.nn.atrous_conv2d(self.inputs, filters, rate, padding) + b) else: self.outputs = act(tf.nn.atrous_conv2d(self.inputs, filters, rate, padding)) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) if b_init: self.all_params.extend( [filters, b] ) else: self.all_params.extend( [filters] )
[docs]class SeparableConv2dLayer(Layer):# Untested """The :class:`SeparableConv2dLayer` class is 2-D convolution with separable filters, see `tf.layers.separable_conv2d <https://www.tensorflow.org/api_docs/python/tf/layers/separable_conv2d>`_. Parameters ----------- layer : a layer class filters : integer, the dimensionality of the output space (i.e. the number output of filters in the convolution). kernel_size : a tuple or list of N positive integers specifying the spatial dimensions of of the filters. Can be a single integer to specify the same value for all spatial dimensions. strides : a tuple or list of N positive integers specifying the strides of the convolution. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any dilation_rate value != 1. padding : one of "valid" or "same" (case-insensitive). data_format : A string, one of channels_last (default) or channels_first. The ordering of the dimensions in the inputs. channels_last corresponds to inputs with shapedata_format = 'NWHC' (batch, width, height, channels) while channels_first corresponds to inputs with shape (batch, channels, width, height). dilation_rate : an integer or tuple/list of 2 integers, specifying the dilation rate to use for dilated convolution. Can be a single integer to specify the same value for all spatial dimensions. Currently, specifying any dilation_rate value != 1 is incompatible with specifying any stride value != 1. depth_multiplier : The number of depthwise convolution output channels for each input channel. The total number of depthwise convolution output channels will be equal to num_filters_in * depth_multiplier. act (activation) : Activation function. Set it to None to maintain a linear activation. use_bias : Boolean, whether the layer uses a bias. depthwise_initializer : An initializer for the depthwise convolution kernel. pointwise_initializer : An initializer for the pointwise convolution kernel. bias_initializer : An initializer for the bias vector. If None, no bias will be applied. depthwise_regularizer : Optional regularizer for the depthwise convolution kernel. pointwise_regularizer : Optional regularizer for the pointwise convolution kernel. bias_regularizer : Optional regularizer for the bias vector. activity_regularizer : Regularizer function for the output. name : a string or None, an optional name to attach to this layer. """ def __init__( self, layer = None, filters = None, kernel_size=5, strides=(1, 1), padding='valid', data_format='channels_last', dilation_rate=(1, 1), depth_multiplier=1, act=None, use_bias=True, depthwise_initializer=None, pointwise_initializer=None, bias_initializer=tf.zeros_initializer, depthwise_regularizer=None, pointwise_regularizer=None, bias_regularizer=None, activity_regularizer=None, name = 'atrou2d' ): Layer.__init__(self, name=name) self.inputs = layer.outputs assert filters is not None assert tf.__version__ > "0.12.1", "This layer only supports for TF 1.0+" if act is None: act = tf.identity bias_initializer = bias_initializer() print(" [TL] SeparableConv2dLayer %s: filters:%s kernel_size:%s strides:%s padding:%s dilation_rate:%s depth_multiplier:%s act:%s" % (self.name, str(filters), str(kernel_size), str(strides), padding, str(dilation_rate), str(depth_multiplier), act.__name__)) with tf.variable_scope(name) as vs: self.outputs = tf.layers.separable_conv2d(self.inputs, filters, kernel_size, strides=strides, padding=padding, data_format=data_format, dilation_rate=dilation_rate, depth_multiplier=depth_multiplier, activation=act, use_bias=use_bias, depthwise_initializer=depthwise_initializer, pointwise_initializer=pointwise_initializer, bias_initializer=bias_initializer, depthwise_regularizer=depthwise_regularizer, pointwise_regularizer=pointwise_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer,) #trainable=True, name=None, reuse=None) variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( variables )
## Initializers for Convuolutional Layers def deconv2d_bilinear_upsampling_initializer(shape): """Returns initializer that can be passed to DeConv2dLayer to initalize the weights to correspond to channel wise bilinear upsampling. Used in some segmantic segmentation approches such as [FCN](https://arxiv.org/abs/1605.06211) Parameters ---------- shape : list of shape shape of the filters, [height, width, output_channels, in_channels], must match that passed to DeConv2dLayer Returns ---------- tf.constant_initializer with weights set to correspond to per channel bilinear upsampling when passed as W_int in DeConv2dLayer Examples -------- >>> rescale_factor = 2 #upsampling by a factor of 2, ie e.g 100->200 >>> filter_size = (2 * rescale_factor - rescale_factor % 2) #Corresponding bilinear filter size >>> num_in_channels = 3 >>> num_out_channels = 3 >>> deconv_filter_shape = [filter_size, filter_size, num_out_channels, num_in_channels] >>> x = tf.placeholder(tf.float32, [1, imsize, imsize, num_channels]) >>> network = tl.layers.InputLayer(x, name='input_layer') >>> bilinear_init = deconv2d_bilinear_upsampling_initializer(shape=filter_shape) >>> network = tl.layers.DeConv2dLayer(network, shape = filter_shape, output_shape = [1, imsize*rescale_factor, imsize*rescale_factor, num_out_channels], strides=[1, rescale_factor, rescale_factor, 1], W_init=bilinear_init, padding='SAME', act=tf.identity, name='g/h1/decon2d') """ if shape[0] != shape[1]: raise Exception('deconv2d_bilinear_upsampling_initializer only supports symmetrical filter sizes') if shape[3] < shape [2]: raise Exception('deconv2d_bilinear_upsampling_initializer behaviour is not defined for num_in_channels < num_out_channels ') filter_size = shape[0] num_out_channels = shape[2] num_in_channels = shape[3] #Create bilinear filter kernel as numpy array bilinear_kernel = np.zeros([filter_size, filter_size], dtype=np.float32) scale_factor = (filter_size + 1) // 2 if filter_size % 2 == 1: center = scale_factor - 1 else: center = scale_factor - 0.5 for x in range(filter_size): for y in range(filter_size): bilinear_kernel[x,y] = (1 - abs(x - center) / scale_factor) * \ (1 - abs(y - center) / scale_factor) weights = np.zeros((filter_size, filter_size, num_out_channels, num_in_channels)) for i in range(num_out_channels): weights[:, :, i, i] = bilinear_kernel #assign numpy array to constant_initalizer and pass to get_variable bilinear_weights_init = tf.constant_initializer(value=weights, dtype=tf.float32) return bilinear_weights_init ## Convolutional layer (Simplified)
[docs]def Conv1d(net, n_filter=32, filter_size=5, stride=1, dilation_rate=1, act=None, padding='SAME', use_cudnn_on_gpu=None, data_format="NWC", W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='conv1d',): """Wrapper for :class:`Conv1dLayer`, if you don't understand how to use :class:`Conv1dLayer`, this function may be easier. Parameters ---------- net : TensorLayer layer. n_filter : number of filter. filter_size : an int. stride : an int. dilation_rate : As it is 1D conv, the default is "NWC". act : None or activation function. others : see :class:`Conv1dLayer`. """ if act is None: act = tf.identity net = Conv1dLayer(layer = net, act = act, shape = [filter_size, int(net.outputs.get_shape()[-1]), n_filter], stride = stride, dilation_rate = dilation_rate, padding = padding, use_cudnn_on_gpu = use_cudnn_on_gpu, data_format = data_format, W_init = W_init, b_init = b_init, W_init_args = W_init_args, b_init_args = b_init_args, name = name, ) return net
[docs]def Conv2d(net, n_filter=32, filter_size=(3, 3), strides=(1, 1), act = None, padding='SAME', W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, use_cudnn_on_gpu = None, data_format = None,name ='conv2d',): """Wrapper for :class:`Conv2dLayer`, if you don't understand how to use :class:`Conv2dLayer`, this function may be easier. Parameters ---------- net : TensorLayer layer. n_filter : number of filter. filter_size : tuple (height, width) for filter size. strides : tuple (height, width) for strides. act : None or activation function. others : see :class:`Conv2dLayer`. Examples -------- >>> w_init = tf.truncated_normal_initializer(stddev=0.01) >>> b_init = tf.constant_initializer(value=0.0) >>> inputs = InputLayer(x, name='inputs') >>> conv1 = Conv2d(inputs, 64, (3, 3), act=tf.nn.relu, padding='SAME', W_init=w_init, b_init=b_init, name='conv1_1') >>> conv1 = Conv2d(conv1, 64, (3, 3), act=tf.nn.relu, padding='SAME', W_init=w_init, b_init=b_init, name='conv1_2') >>> pool1 = MaxPool2d(conv1, (2, 2), padding='SAME', name='pool1') >>> conv2 = Conv2d(pool1, 128, (3, 3), act=tf.nn.relu, padding='SAME', W_init=w_init, b_init=b_init, name='conv2_1') >>> conv2 = Conv2d(conv2, 128, (3, 3), act=tf.nn.relu, padding='SAME', W_init=w_init, b_init=b_init, name='conv2_2') >>> pool2 = MaxPool2d(conv2, (2, 2), padding='SAME', name='pool2') """ assert len(strides) == 2, "len(strides) should be 2, Conv2d and Conv2dLayer are different." if act is None: act = tf.identity try: pre_channel = int(net.outputs.get_shape()[-1]) except: # if pre_channel is ?, it happens when using Spatial Transformer Net pre_channel = 1 print("[warnings] unknow input channels, set to 1") net = Conv2dLayer(net, act = act, shape = [filter_size[0], filter_size[1], pre_channel, n_filter], # 32 features for each 5x5 patch strides = [1, strides[0], strides[1], 1], padding = padding, W_init = W_init, W_init_args = W_init_args, b_init = b_init, b_init_args = b_init_args, use_cudnn_on_gpu = use_cudnn_on_gpu, data_format = data_format, name = name) return net
[docs]def DeConv2d(net, n_out_channel = 32, filter_size=(3, 3), out_size = (30, 30), strides = (2, 2), padding = 'SAME', batch_size = None, act = None, W_init = tf.truncated_normal_initializer(stddev=0.02), b_init = tf.constant_initializer(value=0.0), W_init_args = {}, b_init_args = {}, name ='decnn2d'): """Wrapper for :class:`DeConv2dLayer`, if you don't understand how to use :class:`DeConv2dLayer`, this function may be easier. Parameters ---------- net : TensorLayer layer. n_out_channel : int, number of output channel. filter_size : tuple of (height, width) for filter size. out_size : tuple of (height, width) of output. batch_size : int or None, batch_size. If None, try to find the batch_size from the first dim of net.outputs (you should tell the batch_size when define the input placeholder). strides : tuple of (height, width) for strides. act : None or activation function. others : see :class:`DeConv2dLayer`. """ assert len(strides) == 2, "len(strides) should be 2, DeConv2d and DeConv2dLayer are different." if act is None: act = tf.identity # if batch_size is None: # batch_size = tf.shape(net.outputs)[0] fixed_batch_size = net.outputs.get_shape().with_rank_at_least(1)[0] if fixed_batch_size.value: batch_size = fixed_batch_size.value else: from tensorflow.python.ops import array_ops batch_size = array_ops.shape(net.outputs)[0] net = DeConv2dLayer(layer = net, act = act, shape = [filter_size[0], filter_size[1], n_out_channel, int(net.outputs.get_shape()[-1])], output_shape = [batch_size, int(out_size[0]), int(out_size[1]), n_out_channel], strides = [1, strides[0], strides[1], 1], padding = padding, W_init = W_init, b_init = b_init, W_init_args = W_init_args, b_init_args = b_init_args, name = name) return net
[docs]def MaxPool1d(net, filter_size, strides, padding='valid', data_format='channels_last', name=None): #Untested """Wrapper for `tf.layers.max_pooling1d <https://www.tensorflow.org/api_docs/python/tf/layers/max_pooling1d>`_ . Parameters ------------ net : TensorLayer layer, the tensor over which to pool. Must have rank 3. filter_size (pool_size) : An integer or tuple/list of a single integer, representing the size of the pooling window. strides : An integer or tuple/list of a single integer, specifying the strides of the pooling operation. padding : A string. The padding method, either 'valid' or 'same'. Case-insensitive. data_format : A string, one of channels_last (default) or channels_first. The ordering of the dimensions in the inputs. channels_last corresponds to inputs with shape (batch, length, channels) while channels_first corresponds to inputs with shape (batch, channels, length). name : A string, the name of the layer. Returns -------- - A :class:`Layer` which the output tensor, of rank 3. """ print(" [TL] MaxPool1d %s: filter_size:%s strides:%s padding:%s" % (name, str(filter_size), str(strides), str(padding))) outputs = tf.layers.max_pooling1d(net.outputs, filter_size, strides, padding=padding, data_format=data_format, name=name) net_new = copy.copy(net) net_new.outputs = outputs net_new.all_layers.extend( [outputs] ) return net_new
[docs]def MeanPool1d(net, filter_size, strides, padding='valid', data_format='channels_last', name=None): #Untested """Wrapper for `tf.layers.average_pooling1d <https://www.tensorflow.org/api_docs/python/tf/layers/average_pooling1d>`_ . Parameters ------------ net : TensorLayer layer, the tensor over which to pool. Must have rank 3. filter_size (pool_size) : An integer or tuple/list of a single integer, representing the size of the pooling window. strides : An integer or tuple/list of a single integer, specifying the strides of the pooling operation. padding : A string. The padding method, either 'valid' or 'same'. Case-insensitive. data_format : A string, one of channels_last (default) or channels_first. The ordering of the dimensions in the inputs. channels_last corresponds to inputs with shape (batch, length, channels) while channels_first corresponds to inputs with shape (batch, channels, length). name : A string, the name of the layer. Returns -------- - A :class:`Layer` which the output tensor, of rank 3. """ print(" [TL] MeanPool1d %s: filter_size:%s strides:%s padding:%s" % (name, str(filter_size), str(strides), str(padding))) outputs = tf.layers.average_pooling1d(net.outputs, filter_size, strides, padding=padding, data_format=data_format, name=name) net_new = copy.copy(net) net_new.outputs = outputs net_new.all_layers.extend( [outputs] ) return net_new
[docs]def MaxPool2d(net, filter_size=(2, 2), strides=None, padding='SAME', name='maxpool'): """Wrapper for :class:`PoolLayer`. Parameters ----------- net : TensorLayer layer. filter_size : tuple of (height, width) for filter size. strides : tuple of (height, width). Default is the same with filter_size. others : see :class:`PoolLayer`. """ if strides is None: strides = filter_size assert len(strides) == 2, "len(strides) should be 2, MaxPool2d and PoolLayer are different." net = PoolLayer(net, ksize=[1, filter_size[0], filter_size[1], 1], strides=[1, strides[0], strides[1], 1], padding=padding, pool = tf.nn.max_pool, name = name) return net
[docs]def MeanPool2d(net, filter_size=(2, 2), strides=None, padding='SAME', name='meanpool'): """Wrapper for :class:`PoolLayer`. Parameters ----------- net : TensorLayer layer. filter_size : tuple of (height, width) for filter size. strides : tuple of (height, width). Default is the same with filter_size. others : see :class:`PoolLayer`. """ if strides is None: strides = filter_size assert len(strides) == 2, "len(strides) should be 2, MeanPool2d and PoolLayer are different." net = PoolLayer(net, ksize=[1, filter_size[0], filter_size[1], 1], strides=[1, strides[0], strides[1], 1], padding=padding, pool = tf.nn.avg_pool, name = name) return net
[docs]def MaxPool3d(net, filter_size, strides, padding='valid', data_format='channels_last', name=None): #Untested """Wrapper for `tf.layers.max_pooling3d <https://www.tensorflow.org/api_docs/python/tf/layers/max_pooling3d>`_ . Parameters ------------ net : TensorLayer layer, the tensor over which to pool. Must have rank 5. filter_size (pool_size) : An integer or tuple/list of 3 integers: (pool_depth, pool_height, pool_width) specifying the size of the pooling window. Can be a single integer to specify the same value for all spatial dimensions. strides : An integer or tuple/list of 3 integers, specifying the strides of the pooling operation. Can be a single integer to specify the same value for all spatial dimensions. padding : A string. The padding method, either 'valid' or 'same'. Case-insensitive. data_format : A string. The ordering of the dimensions in the inputs. channels_last (default) and channels_first are supported. channels_last corresponds to inputs with shape (batch, depth, height, width, channels) while channels_first corresponds to inputs with shape (batch, channels, depth, height, width). name : A string, the name of the layer. """ print(" [TL] MaxPool3d %s: filter_size:%s strides:%s padding:%s" % (name, str(filter_size), str(strides), str(padding))) outputs = tf.layers.max_pooling3d(net.outputs, filter_size, strides, padding=padding, data_format=data_format, name=name) net_new = copy.copy(net) net_new.outputs = outputs net_new.all_layers.extend( [outputs] ) return net_new
[docs]def MeanPool3d(net, filter_size, strides, padding='valid', data_format='channels_last', name=None): #Untested """Wrapper for `tf.layers.average_pooling3d <https://www.tensorflow.org/api_docs/python/tf/layers/average_pooling3d>`_ Parameters ------------ net : TensorLayer layer, the tensor over which to pool. Must have rank 5. filter_size (pool_size) : An integer or tuple/list of 3 integers: (pool_depth, pool_height, pool_width) specifying the size of the pooling window. Can be a single integer to specify the same value for all spatial dimensions. strides : An integer or tuple/list of 3 integers, specifying the strides of the pooling operation. Can be a single integer to specify the same value for all spatial dimensions. padding : A string. The padding method, either 'valid' or 'same'. Case-insensitive. data_format : A string. The ordering of the dimensions in the inputs. channels_last (default) and channels_first are supported. channels_last corresponds to inputs with shape (batch, depth, height, width, channels) while channels_first corresponds to inputs with shape (batch, channels, depth, height, width). name : A string, the name of the layer. """ print(" [TL] MeanPool3d %s: filter_size:%s strides:%s padding:%s" % (name, str(filter_size), str(strides), str(padding))) outputs = tf.layers.average_pooling3d(net.outputs, filter_size, strides, padding=padding, data_format=data_format, name=name) net_new = copy.copy(net) net_new.outputs = outputs net_new.all_layers.extend( [outputs] ) return net_new
## Super resolution
[docs]def SubpixelConv2d(net, scale=2, n_out_channel=None, act=tf.identity, name='subpixel_conv2d'): """The :class:`SubpixelConv2d` class is a sub-pixel 2d convolutional ayer, usually be used for Super-Resolution applications, `example code <https://github.com/zsdonghao/SRGAN/>`_. Parameters ------------ net : TensorLayer layer. scale : int, upscaling ratio, a wrong setting will lead to Dimension size error. n_out_channel : int or None, the number of output channels. Note that, the number of input channels == (scale x scale) x The number of output channels. If None, automatically set n_out_channel == the number of input channels / (scale x scale). act : activation function. name : string. An optional name to attach to this layer. Examples --------- >>> # examples here just want to tell you how to set the n_out_channel. >>> x = np.random.rand(2, 16, 16, 4) >>> X = tf.placeholder("float32", shape=(2, 16, 16, 4), name="X") >>> net = InputLayer(X, name='input') >>> net = SubpixelConv2d(net, scale=2, n_out_channel=1, name='subpixel_conv2d') >>> y = sess.run(net.outputs, feed_dict={X: x}) >>> print(x.shape, y.shape) ... (2, 16, 16, 4) (2, 32, 32, 1) >>> >>> x = np.random.rand(2, 16, 16, 4*10) >>> X = tf.placeholder("float32", shape=(2, 16, 16, 4*10), name="X") >>> net = InputLayer(X, name='input2') >>> net = SubpixelConv2d(net, scale=2, n_out_channel=10, name='subpixel_conv2d2') >>> y = sess.run(net.outputs, feed_dict={X: x}) >>> print(x.shape, y.shape) ... (2, 16, 16, 40) (2, 32, 32, 10) >>> >>> x = np.random.rand(2, 16, 16, 25*10) >>> X = tf.placeholder("float32", shape=(2, 16, 16, 25*10), name="X") >>> net = InputLayer(X, name='input3') >>> net = SubpixelConv2d(net, scale=5, n_out_channel=None, name='subpixel_conv2d3') >>> y = sess.run(net.outputs, feed_dict={X: x}) >>> print(x.shape, y.shape) ... (2, 16, 16, 250) (2, 80, 80, 10) References ------------ - `Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network <https://arxiv.org/pdf/1609.05158.pdf>`_ """ # github/Tetrachrome/subpixel https://github.com/Tetrachrome/subpixel/blob/master/subpixel.py _err_log = "SubpixelConv2d: The number of input channels == (scale x scale) x The number of output channels" scope_name = tf.get_variable_scope().name if scope_name: name = scope_name + '/' + name def _PS(X, r, n_out_channel): if n_out_channel >= 1: assert int(X.get_shape()[-1]) == (r ** 2) * n_out_channel, _err_log bsize, a, b, c = X.get_shape().as_list() bsize = tf.shape(X)[0] # Handling Dimension(None) type for undefined batch dim Xs=tf.split(X,r,3) #b*h*w*r*r Xr=tf.concat(Xs,2) #b*h*(r*w)*r X=tf.reshape(Xr,(bsize,r*a,r*b,n_out_channel)) # b*(r*h)*(r*w)*c else: print(_err_log) return X inputs = net.outputs if n_out_channel is None: assert int(inputs.get_shape()[-1])/ (scale ** 2) % 1 == 0, _err_log n_out_channel = int(int(inputs.get_shape()[-1])/ (scale ** 2)) print(" [TL] SubpixelConv2d %s: scale: %d n_out_channel: %s act: %s" % (name, scale, n_out_channel, act.__name__)) net_new = Layer(inputs, name=name) # with tf.name_scope(name): with tf.variable_scope(name) as vs: net_new.outputs = act(_PS(inputs, r=scale, n_out_channel=n_out_channel)) net_new.all_layers = list(net.all_layers) net_new.all_params = list(net.all_params) net_new.all_drop = dict(net.all_drop) net_new.all_layers.extend( [net_new.outputs] ) return net_new
def SubpixelConv2d_old(net, scale=2, n_out_channel=None, act=tf.identity, name='subpixel_conv2d'): """The :class:`SubpixelConv2d` class is a sub-pixel 2d convolutional ayer, usually be used for Super-Resolution applications, `example code <https://github.com/zsdonghao/SRGAN/>`_. Parameters ------------ net : TensorLayer layer. scale : int, upscaling ratio, a wrong setting will lead to Dimension size error. n_out_channel : int or None, the number of output channels. Note that, the number of input channels == (scale x scale) x The number of output channels. If None, automatically set n_out_channel == the number of input channels / (scale x scale). act : activation function. name : string. An optional name to attach to this layer. Examples --------- >>> # examples here just want to tell you how to set the n_out_channel. >>> x = np.random.rand(2, 16, 16, 4) >>> X = tf.placeholder("float32", shape=(2, 16, 16, 4), name="X") >>> net = InputLayer(X, name='input') >>> net = SubpixelConv2d(net, scale=2, n_out_channel=1, name='subpixel_conv2d') >>> y = sess.run(net.outputs, feed_dict={X: x}) >>> print(x.shape, y.shape) ... (2, 16, 16, 4) (2, 32, 32, 1) >>> >>> x = np.random.rand(2, 16, 16, 4*10) >>> X = tf.placeholder("float32", shape=(2, 16, 16, 4*10), name="X") >>> net = InputLayer(X, name='input2') >>> net = SubpixelConv2d(net, scale=2, n_out_channel=10, name='subpixel_conv2d2') >>> y = sess.run(net.outputs, feed_dict={X: x}) >>> print(x.shape, y.shape) ... (2, 16, 16, 40) (2, 32, 32, 10) >>> >>> x = np.random.rand(2, 16, 16, 25*10) >>> X = tf.placeholder("float32", shape=(2, 16, 16, 25*10), name="X") >>> net = InputLayer(X, name='input3') >>> net = SubpixelConv2d(net, scale=5, n_out_channel=None, name='subpixel_conv2d3') >>> y = sess.run(net.outputs, feed_dict={X: x}) >>> print(x.shape, y.shape) ... (2, 16, 16, 250) (2, 80, 80, 10) References ------------ - `Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network <https://arxiv.org/pdf/1609.05158.pdf>`_ """ # github/Tetrachrome/subpixel https://github.com/Tetrachrome/subpixel/blob/master/subpixel.py _err_log = "SubpixelConv2d: The number of input channels == (scale x scale) x The number of output channels" scope_name = tf.get_variable_scope().name if scope_name: name = scope_name + '/' + name def _phase_shift(I, r): if tf.__version__ < '1.0': raise Exception("Only support TF1.0+") bsize, a, b, c = I.get_shape().as_list() bsize = tf.shape(I)[0] # Handling Dimension(None) type for undefined batch dim X = tf.reshape(I, (bsize, a, b, r, r)) X = tf.transpose(X, (0, 1, 2, 4, 3)) # bsize, a, b, 1, 1 # tf 0.12 # X = tf.split(1, a, X) # a, [bsize, b, r, r] # tf 0.12 X = tf.split(X, a, 1) # X = tf.concat(2, [tf.squeeze(x, axis=1) for x in X]) # bsize, b, a*r, r # tf 0.12 X = tf.concat([tf.squeeze(x, axis=1) for x in X], 2) # X = tf.split(1, b, X) # b, [bsize, a*r, r] # tf 0.12 X = tf.split(X, b, 1) # X = tf.concat(2, [tf.squeeze(x, axis=1) for x in X]) # bsize, a*r, b*r # tf 0.12 X = tf.concat([tf.squeeze(x, axis=1) for x in X], 2) return tf.reshape(X, (bsize, a*r, b*r, 1)) def _PS(X, r, n_out_channel): if n_out_channel > 1: assert int(X.get_shape()[-1]) == (r ** 2) * n_out_channel, _err_log Xc = tf.split(X, n_out_channel, 3) X = tf.concat([_phase_shift(x, r) for x in Xc], 3) elif n_out_channel == 1: assert int(X.get_shape()[-1]) == (r ** 2), _err_log X = _phase_shift(X, r) else: print(_err_log) return X inputs = net.outputs if n_out_channel is None: assert int(inputs.get_shape()[-1])/ (scale ** 2) % 1 == 0, _err_log n_out_channel = int(int(inputs.get_shape()[-1])/ (scale ** 2)) print(" [TL] SubpixelConv2d %s: scale: %d n_out_channel: %s act: %s" % (name, scale, n_out_channel, act.__name__)) net_new = Layer(inputs, name=name) # with tf.name_scope(name): with tf.variable_scope(name) as vs: net_new.outputs = act(_PS(inputs, r=scale, n_out_channel=n_out_channel)) net_new.all_layers = list(net.all_layers) net_new.all_params = list(net.all_params) net_new.all_drop = dict(net.all_drop) net_new.all_layers.extend( [net_new.outputs] ) return net_new
[docs]def SubpixelConv1d(net, scale=2, act=tf.identity, name='subpixel_conv1d'): """One-dimensional subpixel upsampling layer. Calls a tensorflow function that directly implements this functionality. We assume input has dim (batch, width, r) Parameters ------------ net : TensorLayer layer. scale : int, upscaling ratio, a wrong setting will lead to Dimension size error. act : activation function. name : string. An optional name to attach to this layer. Examples ---------- >>> t_signal = tf.placeholder('float32', [10, 100, 4], name='x') >>> n = InputLayer(t_signal, name='in') >>> n = SubpixelConv1d(n, scale=2, name='s') >>> print(n.outputs.shape) ... (10, 200, 2) References ----------- - `Audio Super Resolution Implementation <https://github.com/kuleshov/audio-super-res/blob/master/src/models/layers/subpixel.py>`_. """ def _PS(I, r): X = tf.transpose(I, [2,1,0]) # (r, w, b) X = tf.batch_to_space_nd(X, [r], [[0,0]]) # (1, r*w, b) X = tf.transpose(X, [2,1,0]) return X print(" [TL] SubpixelConv1d %s: scale: %d act: %s" % (name, scale, act.__name__)) inputs = net.outputs net_new = Layer(inputs, name=name) with tf.name_scope(name): net_new.outputs = act(_PS(inputs, r=scale)) net_new.all_layers = list(net.all_layers) net_new.all_params = list(net.all_params) net_new.all_drop = dict(net.all_drop) net_new.all_layers.extend( [net_new.outputs] ) return net_new
## Spatial Transformer Nets
[docs]def transformer(U, theta, out_size, name='SpatialTransformer2dAffine', **kwargs): """Spatial Transformer Layer for `2D Affine Transformation <https://en.wikipedia.org/wiki/Affine_transformation>`_ , see :class:`SpatialTransformer2dAffineLayer` class. Parameters ---------- U : float The output of a convolutional net should have the shape [num_batch, height, width, num_channels]. theta: float The output of the localisation network should be [num_batch, 6], value range should be [0, 1] (via tanh). out_size: tuple of two ints The size of the output of the network (height, width) References ---------- - `Spatial Transformer Networks <https://arxiv.org/abs/1506.02025>`_ - `TensorFlow/Models <https://github.com/tensorflow/models/tree/master/transformer>`_ Notes ----- - To initialize the network to the identity transform init. >>> ``theta`` to >>> identity = np.array([[1., 0., 0.], ... [0., 1., 0.]]) >>> identity = identity.flatten() >>> theta = tf.Variable(initial_value=identity) """ def _repeat(x, n_repeats): with tf.variable_scope('_repeat'): rep = tf.transpose( tf.expand_dims(tf.ones(shape=tf.stack([n_repeats, ])), 1), [1, 0]) rep = tf.cast(rep, 'int32') x = tf.matmul(tf.reshape(x, (-1, 1)), rep) return tf.reshape(x, [-1]) def _interpolate(im, x, y, out_size): with tf.variable_scope('_interpolate'): # constants num_batch = tf.shape(im)[0] height = tf.shape(im)[1] width = tf.shape(im)[2] channels = tf.shape(im)[3] x = tf.cast(x, 'float32') y = tf.cast(y, 'float32') height_f = tf.cast(height, 'float32') width_f = tf.cast(width, 'float32') out_height = out_size[0] out_width = out_size[1] zero = tf.zeros([], dtype='int32') max_y = tf.cast(tf.shape(im)[1] - 1, 'int32') max_x = tf.cast(tf.shape(im)[2] - 1, 'int32') # scale indices from [-1, 1] to [0, width/height] x = (x + 1.0)*(width_f) / 2.0 y = (y + 1.0)*(height_f) / 2.0 # do sampling x0 = tf.cast(tf.floor(x), 'int32') x1 = x0 + 1 y0 = tf.cast(tf.floor(y), 'int32') y1 = y0 + 1 x0 = tf.clip_by_value(x0, zero, max_x) x1 = tf.clip_by_value(x1, zero, max_x) y0 = tf.clip_by_value(y0, zero, max_y) y1 = tf.clip_by_value(y1, zero, max_y) dim2 = width dim1 = width*height base = _repeat(tf.range(num_batch)*dim1, out_height*out_width) base_y0 = base + y0*dim2 base_y1 = base + y1*dim2 idx_a = base_y0 + x0 idx_b = base_y1 + x0 idx_c = base_y0 + x1 idx_d = base_y1 + x1 # use indices to lookup pixels in the flat image and restore # channels dim im_flat = tf.reshape(im, tf.stack([-1, channels])) im_flat = tf.cast(im_flat, 'float32') Ia = tf.gather(im_flat, idx_a) Ib = tf.gather(im_flat, idx_b) Ic = tf.gather(im_flat, idx_c) Id = tf.gather(im_flat, idx_d) # and finally calculate interpolated values x0_f = tf.cast(x0, 'float32') x1_f = tf.cast(x1, 'float32') y0_f = tf.cast(y0, 'float32') y1_f = tf.cast(y1, 'float32') wa = tf.expand_dims(((x1_f-x) * (y1_f-y)), 1) wb = tf.expand_dims(((x1_f-x) * (y-y0_f)), 1) wc = tf.expand_dims(((x-x0_f) * (y1_f-y)), 1) wd = tf.expand_dims(((x-x0_f) * (y-y0_f)), 1) output = tf.add_n([wa*Ia, wb*Ib, wc*Ic, wd*Id]) return output def _meshgrid(height, width): with tf.variable_scope('_meshgrid'): # This should be equivalent to: # x_t, y_t = np.meshgrid(np.linspace(-1, 1, width), # np.linspace(-1, 1, height)) # ones = np.ones(np.prod(x_t.shape)) # grid = np.vstack([x_t.flatten(), y_t.flatten(), ones]) x_t = tf.matmul(tf.ones(shape=tf.stack([height, 1])), tf.transpose(tf.expand_dims(tf.linspace(-1.0, 1.0, width), 1), [1, 0])) y_t = tf.matmul(tf.expand_dims(tf.linspace(-1.0, 1.0, height), 1), tf.ones(shape=tf.stack([1, width]))) x_t_flat = tf.reshape(x_t, (1, -1)) y_t_flat = tf.reshape(y_t, (1, -1)) ones = tf.ones_like(x_t_flat) grid = tf.concat(axis=0, values=[x_t_flat, y_t_flat, ones]) return grid def _transform(theta, input_dim, out_size): with tf.variable_scope('_transform'): num_batch = tf.shape(input_dim)[0] height = tf.shape(input_dim)[1] width = tf.shape(input_dim)[2] num_channels = tf.shape(input_dim)[3] theta = tf.reshape(theta, (-1, 2, 3)) theta = tf.cast(theta, 'float32') # grid of (x_t, y_t, 1), eq (1) in ref [1] height_f = tf.cast(height, 'float32') width_f = tf.cast(width, 'float32') out_height = out_size[0] out_width = out_size[1] grid = _meshgrid(out_height, out_width) grid = tf.expand_dims(grid, 0) grid = tf.reshape(grid, [-1]) grid = tf.tile(grid, tf.stack([num_batch])) grid = tf.reshape(grid, tf.stack([num_batch, 3, -1])) # Transform A x (x_t, y_t, 1)^T -> (x_s, y_s) T_g = tf.matmul(theta, grid) x_s = tf.slice(T_g, [0, 0, 0], [-1, 1, -1]) y_s = tf.slice(T_g, [0, 1, 0], [-1, 1, -1]) x_s_flat = tf.reshape(x_s, [-1]) y_s_flat = tf.reshape(y_s, [-1]) input_transformed = _interpolate( input_dim, x_s_flat, y_s_flat, out_size) output = tf.reshape( input_transformed, tf.stack([num_batch, out_height, out_width, num_channels])) return output with tf.variable_scope(name): output = _transform(theta, U, out_size) return output
[docs]def batch_transformer(U, thetas, out_size, name='BatchSpatialTransformer2dAffine'): """Batch Spatial Transformer function for `2D Affine Transformation <https://en.wikipedia.org/wiki/Affine_transformation>`_. Parameters ---------- U : float tensor of inputs [batch, height, width, num_channels] thetas : float a set of transformations for each input [batch, num_transforms, 6] out_size : int the size of the output [out_height, out_width] Returns: float Tensor of size [batch * num_transforms, out_height, out_width, num_channels] """ with tf.variable_scope(name): num_batch, num_transforms = map(int, thetas.get_shape().as_list()[:2]) indices = [[i]*num_transforms for i in xrange(num_batch)] input_repeated = tf.gather(U, tf.reshape(indices, [-1])) return transformer(input_repeated, thetas, out_size)
[docs]class SpatialTransformer2dAffineLayer(Layer): """The :class:`SpatialTransformer2dAffineLayer` class is a `Spatial Transformer Layer <https://arxiv.org/abs/1506.02025>`_ for `2D Affine Transformation <https://en.wikipedia.org/wiki/Affine_transformation>`_. Parameters ----------- layer : a layer class with 4-D Tensor of shape [batch, height, width, channels] theta_layer : a layer class for the localisation network. In this layer, we will use a :class:`DenseLayer` to make the theta size to [batch, 6], value range to [0, 1] (via tanh). out_size : tuple of two ints. The size of the output of the network (height, width), the feature maps will be resized by this. References ----------- - `Spatial Transformer Networks <https://arxiv.org/abs/1506.02025>`_ - `TensorFlow/Models <https://github.com/tensorflow/models/tree/master/transformer>`_ """ def __init__( self, layer = None, theta_layer = None, out_size = [40, 40], name ='sapatial_trans_2d_affine', ): Layer.__init__(self, name=name) self.inputs = layer.outputs self.theta_layer = theta_layer print(" [TL] SpatialTransformer2dAffineLayer %s: in_size:%s out_size:%s" % (name, self.inputs.get_shape().as_list(), out_size)) with tf.variable_scope(name) as vs: ## 1. make the localisation network to [batch, 6] via Flatten and Dense. if self.theta_layer.outputs.get_shape().ndims > 2: self.theta_layer.outputs = flatten_reshape(self.theta_layer.outputs, 'flatten') ## 2. To initialize the network to the identity transform init. # 2.1 W n_in = int(self.theta_layer.outputs.get_shape()[-1]) shape = (n_in, 6) W = tf.get_variable(name='W', initializer=tf.zeros(shape)) # 2.2 b identity = tf.constant(np.array([[1., 0, 0], [0, 1., 0]]).astype('float32').flatten()) b = tf.get_variable(name='b', initializer=identity) # 2.3 transformation matrix self.theta = tf.nn.tanh(tf.matmul(self.theta_layer.outputs, W) + b) ## 3. Spatial Transformer Sampling # 3.1 transformation self.outputs = transformer(self.inputs, self.theta, out_size=out_size) # 3.2 automatically set batch_size and channels # e.g. [?, 40, 40, ?] --> [64, 40, 40, 1] or [64, 20, 20, 4]/ Hao Dong # fixed_batch_size = self.inputs.get_shape().with_rank_at_least(1)[0] if fixed_batch_size.value: batch_size = fixed_batch_size.value else: from tensorflow.python.ops import array_ops batch_size = array_ops.shape(self.inputs)[0] size = self.inputs.get_shape().as_list() n_channels = self.inputs.get_shape().as_list()[-1] # print(self.outputs) self.outputs = tf.reshape(self.outputs, shape=[batch_size, out_size[0], out_size[1], n_channels]) # print(self.outputs) # exit() ## 4. Get all parameters variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) ## fixed self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) ## theta_layer self.all_layers.extend(theta_layer.all_layers) self.all_params.extend(theta_layer.all_params) self.all_drop.update(theta_layer.all_drop) ## this layer self.all_layers.extend( [self.outputs] ) self.all_params.extend( variables )
# ## Normalization layer
[docs]class LocalResponseNormLayer(Layer): """The :class:`LocalResponseNormLayer` class is for Local Response Normalization, see ``tf.nn.local_response_normalization`` or ``tf.nn.lrn`` for new TF version. The 4-D input tensor is treated as a 3-D array of 1-D vectors (along the last dimension), and each vector is normalized independently. Within a given vector, each component is divided by the weighted, squared sum of inputs within depth_radius. Parameters ----------- layer : a layer class. Must be one of the following types: float32, half. 4-D. depth_radius : An optional int. Defaults to 5. 0-D. Half-width of the 1-D normalization window. bias : An optional float. Defaults to 1. An offset (usually positive to avoid dividing by 0). alpha : An optional float. Defaults to 1. A scale factor, usually positive. beta : An optional float. Defaults to 0.5. An exponent. name : A string or None, an optional name to attach to this layer. """ def __init__( self, layer = None, depth_radius = None, bias = None, alpha = None, beta = None, name ='lrn_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] LocalResponseNormLayer %s: depth_radius: %d, bias: %f, alpha: %f, beta: %f" % (self.name, depth_radius, bias, alpha, beta)) with tf.variable_scope(name) as vs: self.outputs = tf.nn.lrn(self.inputs, depth_radius=depth_radius, bias=bias, alpha=alpha, beta=beta) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
[docs]class BatchNormLayer(Layer): """ The :class:`BatchNormLayer` class is a normalization layer, see ``tf.nn.batch_normalization`` and ``tf.nn.moments``. Batch normalization on fully-connected or convolutional maps. Parameters ----------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. decay : float, default is 0.9. A decay factor for ExponentialMovingAverage, use larger value for large dataset. epsilon : float A small float number to avoid dividing by 0. act : activation function. is_train : boolean Whether train or inference. beta_init : beta initializer The initializer for initializing beta gamma_init : gamma initializer The initializer for initializing gamma name : a string or None An optional name to attach to this layer. References ---------- - `Source <https://github.com/ry/tensorflow-resnet/blob/master/resnet.py>`_ - `stackoverflow <http://stackoverflow.com/questions/38312668/how-does-one-do-inference-with-batch-normalization-with-tensor-flow>`_ """ def __init__( self, layer = None, decay = 0.9, epsilon = 0.00001, act = tf.identity, is_train = False, beta_init = tf.zeros_initializer, gamma_init = tf.random_normal_initializer(mean=1.0, stddev=0.002), # tf.ones_initializer, name ='batchnorm_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] BatchNormLayer %s: decay:%f epsilon:%f act:%s is_train:%s" % (self.name, decay, epsilon, act.__name__, is_train)) x_shape = self.inputs.get_shape() params_shape = x_shape[-1:] from tensorflow.python.training import moving_averages from tensorflow.python.ops import control_flow_ops with tf.variable_scope(name) as vs: axis = list(range(len(x_shape) - 1)) ## 1. beta, gamma if tf.__version__ > '0.12.1' and beta_init == tf.zeros_initializer: beta_init = beta_init() beta = tf.get_variable('beta', shape=params_shape, initializer=beta_init, trainable=is_train)#, restore=restore) gamma = tf.get_variable('gamma', shape=params_shape, initializer=gamma_init, trainable=is_train, )#restore=restore) ## 2. if tf.__version__ > '0.12.1': moving_mean_init = tf.zeros_initializer() else: moving_mean_init = tf.zeros_initializer moving_mean = tf.get_variable('moving_mean', params_shape, initializer=moving_mean_init, trainable=False,)# restore=restore) moving_variance = tf.get_variable('moving_variance', params_shape, initializer=tf.constant_initializer(1.), trainable=False,)# restore=restore) ## 3. # These ops will only be preformed when training. mean, variance = tf.nn.moments(self.inputs, axis) try: # TF12 update_moving_mean = moving_averages.assign_moving_average( moving_mean, mean, decay, zero_debias=False) # if zero_debias=True, has bias update_moving_variance = moving_averages.assign_moving_average( moving_variance, variance, decay, zero_debias=False) # if zero_debias=True, has bias # print("TF12 moving") except Exception as e: # TF11 update_moving_mean = moving_averages.assign_moving_average( moving_mean, mean, decay) update_moving_variance = moving_averages.assign_moving_average( moving_variance, variance, decay) # print("TF11 moving") def mean_var_with_update(): with tf.control_dependencies([update_moving_mean, update_moving_variance]): return tf.identity(mean), tf.identity(variance) if is_train: mean, var = mean_var_with_update() self.outputs = act( tf.nn.batch_normalization(self.inputs, mean, var, beta, gamma, epsilon) ) else: self.outputs = act( tf.nn.batch_normalization(self.inputs, moving_mean, moving_variance, beta, gamma, epsilon) ) variables = [beta, gamma, moving_mean, moving_variance] # print(len(variables)) # for idx, v in enumerate(variables): # print(" var {:3}: {:15} {}".format(idx, str(v.get_shape()), v)) # exit() self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( variables )
# class BatchNormLayer_TF(Layer): # Work well TF contrib https://github.com/tensorflow/tensorflow/blob/b826b79718e3e93148c3545e7aa3f90891744cc0/tensorflow/contrib/layers/python/layers/layers.py#L100 # """ # The :class:`BatchNormLayer` class is a normalization layer, see ``tf.nn.batch_normalization`` and ``tf.nn.moments``. # # Batch normalization on fully-connected or convolutional maps. # # Parameters # ----------- # layer : a :class:`Layer` instance # The `Layer` class feeding into this layer. # decay : float # A decay factor for ExponentialMovingAverage. # center: If True, subtract `beta`. If False, `beta` is ignored. # scale: If True, multiply by `gamma`. If False, `gamma` is # not used. When the next layer is linear (also e.g. `nn.relu`), this can be # disabled since the scaling can be done by the next layer. # epsilon : float # A small float number to avoid dividing by 0. # act : activation function. # is_train : boolean # Whether train or inference. # beta_init : beta initializer # The initializer for initializing beta # gamma_init : gamma initializer # The initializer for initializing gamma # name : a string or None # An optional name to attach to this layer. # # References # ---------- # - `Source <https://github.com/ry/tensorflow-resnet/blob/master/resnet.py>`_ # - `stackoverflow <http://stackoverflow.com/questions/38312668/how-does-one-do-inference-with-batch-normalization-with-tensor-flow>`_ # """ # def __init__( # self, # layer = None, # decay = 0.95,#.999, # center = True, # scale = True, # epsilon = 0.00001, # act = tf.identity, # is_train = False, # beta_init = tf.zeros_initializer, # # gamma_init = tf.ones_initializer, # gamma_init = tf.random_normal_initializer(mean=1.0, stddev=0.002), # name ='batchnorm_layer', # ): # Layer.__init__(self, name=name) # self.inputs = layer.outputs # print(" [TL] BatchNormLayer %s: decay: %f, epsilon: %f, act: %s, is_train: %s" % # (self.name, decay, epsilon, act.__name__, is_train)) # from tensorflow.contrib.layers.python.layers import utils # from tensorflow.contrib.framework.python.ops import variables # from tensorflow.python.ops import init_ops # from tensorflow.python.ops import nn # from tensorflow.python.training import moving_averages # from tensorflow.python.framework import ops # from tensorflow.python.ops import variable_scope # variables_collections = None # outputs_collections=None # updates_collections=None#ops.GraphKeys.UPDATE_OPS # # with variable_scope.variable_op_scope([inputs], # # scope, 'BatchNorm', reuse=reuse) as sc: # # with variable_scope.variable_op_scope([self.inputs], None, name) as vs: # with tf.variable_scope(name) as vs: # inputs_shape = self.inputs.get_shape() # dtype = self.inputs.dtype.base_dtype # axis = list(range(len(inputs_shape) - 1)) # [0, 1, 2] # params_shape = inputs_shape[-1:] # # Allocate parameters for the beta and gamma of the normalization. # beta, gamma = None, None # if center: # beta_collections = utils.get_variable_collections(variables_collections, # 'beta') # beta = variables.model_variable('beta', # shape=params_shape, # dtype=dtype, # # initializer=init_ops.zeros_initializer, # initializer=beta_init, # collections=beta_collections,) # # trainable=trainable) # if scale: # gamma_collections = utils.get_variable_collections(variables_collections, # 'gamma') # gamma = variables.model_variable('gamma', # shape=params_shape, # dtype=dtype, # # initializer=init_ops.ones_initializer, # initializer=gamma_init, # collections=gamma_collections,) # # trainable=trainable) # # Create moving_mean and moving_variance variables and add them to the # # appropiate collections. # moving_mean_collections = utils.get_variable_collections( # variables_collections, # 'moving_mean') # moving_mean = variables.model_variable( # 'moving_mean', # shape=params_shape, # dtype=dtype, # # initializer=init_ops.zeros_initializer, # initializer=tf.zeros_initializer, # trainable=False, # collections=moving_mean_collections) # moving_variance_collections = utils.get_variable_collections( # variables_collections, # 'moving_variance') # moving_variance = variables.model_variable( # 'moving_variance', # shape=params_shape, # dtype=dtype, # # initializer=init_ops.ones_initializer, # initializer=tf.constant_initializer(1.), # trainable=False, # collections=moving_variance_collections) # if is_train: # # Calculate the moments based on the individual batch. # mean, variance = nn.moments(self.inputs, axis, shift=moving_mean) # # Update the moving_mean and moving_variance moments. # # update_moving_mean = moving_averages.assign_moving_average( # # moving_mean, mean, decay) # # update_moving_variance = moving_averages.assign_moving_average( # # moving_variance, variance, decay) # # if updates_collections is None: # # # Make sure the updates are computed here. # # with ops.control_dependencies([update_moving_mean, # # update_moving_variance]): # # outputs = nn.batch_normalization( # # self.inputs, mean, variance, beta, gamma, epsilon) # # update_moving_mean = tf.assign(moving_mean, # moving_mean * decay + mean * (1 - decay)) # update_moving_variance = tf.assign(moving_variance, # moving_variance * decay + variance * (1 - decay)) # with tf.control_dependencies([update_moving_mean, update_moving_variance]): # outputs = nn.batch_normalization( # self.inputs, mean, variance, beta, gamma, epsilon) # # else: # # # Collect the updates to be computed later. # # ops.add_to_collections(updates_collections, update_moving_mean) # # ops.add_to_collections(updates_collections, update_moving_variance) # # outputs = nn.batch_normalization( # # self.inputs, mean, variance, beta, gamma, epsilon) # else: # # mean, variance = nn.moments(self.inputs, axis, shift=moving_mean) # outputs = nn.batch_normalization( # self.inputs, moving_mean, moving_variance, beta, gamma, epsilon) # # self.inputs, mean, variance, beta, gamma, epsilon) # outputs.set_shape(self.inputs.get_shape()) # # if activation_fn: # self.outputs = act(outputs) # # # variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) # # return utils.collect_named_outputs(outputs_collections, sc.name, outputs) # variables = [beta, gamma, moving_mean, moving_variance] # # mean, variance = nn.moments(self.inputs, axis, shift=moving_mean) # self.check_mean = mean # self.check_variance = variance # # self.all_layers = list(layer.all_layers) # self.all_params = list(layer.all_params) # self.all_drop = dict(layer.all_drop) # self.all_layers.extend( [self.outputs] ) # self.all_params.extend( variables ) # # class BatchNormLayer5(Layer): # Akara Work well # """ # The :class:`BatchNormLayer` class is a normalization layer, see ``tf.nn.batch_normalization`` and ``tf.nn.moments``. # # Batch normalization on fully-connected or convolutional maps. # # Parameters # ----------- # layer : a :class:`Layer` instance # The `Layer` class feeding into this layer. # decay : float # A decay factor for ExponentialMovingAverage. # epsilon : float # A small float number to avoid dividing by 0. # act : activation function. # is_train : boolean # Whether train or inference. # beta_init : beta initializer # The initializer for initializing beta # gamma_init : gamma initializer # The initializer for initializing gamma # name : a string or None # An optional name to attach to this layer. # # References # ---------- # - `Source <https://github.com/ry/tensorflow-resnet/blob/master/resnet.py>`_ # - `stackoverflow <http://stackoverflow.com/questions/38312668/how-does-one-do-inference-with-batch-normalization-with-tensor-flow>`_ # """ # def __init__( # self, # layer = None, # decay = 0.9, # epsilon = 0.00001, # act = tf.identity, # is_train = False, # beta_init = tf.zeros_initializer, # # gamma_init = tf.ones_initializer, # gamma_init = tf.random_normal_initializer(mean=1.0, stddev=0.002), # name ='batchnorm_layer', # ): # Layer.__init__(self, name=name) # self.inputs = layer.outputs # print(" [TL] BatchNormLayer %s: decay: %f, epsilon: %f, act: %s, is_train: %s" % # (self.name, decay, epsilon, act.__name__, is_train)) # x_shape = self.inputs.get_shape() # params_shape = x_shape[-1:] # # from tensorflow.python.training import moving_averages # from tensorflow.python.ops import control_flow_ops # # with tf.variable_scope(name) as vs: # axis = list(range(len(x_shape) - 1)) # # ## 1. beta, gamma # beta = tf.get_variable('beta', shape=params_shape, # initializer=beta_init, # trainable=is_train)#, restore=restore) # # gamma = tf.get_variable('gamma', shape=params_shape, # initializer=gamma_init, trainable=is_train, # )#restore=restore) # # ## 2. moving variables during training (not update by gradient!) # moving_mean = tf.get_variable('moving_mean', # params_shape, # initializer=tf.zeros_initializer, # trainable=False,)# restore=restore) # moving_variance = tf.get_variable('moving_variance', # params_shape, # initializer=tf.constant_initializer(1.), # trainable=False,)# restore=restore) # # batch_mean, batch_var = tf.nn.moments(self.inputs, axis) # ## 3. # # These ops will only be preformed when training. # def mean_var_with_update(): # try: # TF12 # update_moving_mean = moving_averages.assign_moving_average( # moving_mean, batch_mean, decay, zero_debias=False) # if zero_debias=True, has bias # update_moving_variance = moving_averages.assign_moving_average( # moving_variance, batch_var, decay, zero_debias=False) # if zero_debias=True, has bias # # print("TF12 moving") # except Exception as e: # TF11 # update_moving_mean = moving_averages.assign_moving_average( # moving_mean, batch_mean, decay) # update_moving_variance = moving_averages.assign_moving_average( # moving_variance, batch_var, decay) # # print("TF11 moving") # # # def mean_var_with_update(): # with tf.control_dependencies([update_moving_mean, update_moving_variance]): # # return tf.identity(update_moving_mean), tf.identity(update_moving_variance) # return tf.identity(batch_mean), tf.identity(batch_var) # # # if not is_train: # if is_train: # mean, var = mean_var_with_update() # else: # mean, var = (moving_mean, moving_variance) # # normed = tf.nn.batch_normalization( # x=self.inputs, # mean=mean, # variance=var, # offset=beta, # scale=gamma, # variance_epsilon=epsilon, # name="tf_bn" # ) # self.outputs = act( normed ) # # variables = [beta, gamma, moving_mean, moving_variance] # # print(len(variables)) # # for idx, v in enumerate(variables): # # print(" var {:3}: {:15} {}".format(idx, str(v.get_shape()), v)) # # exit() # # self.all_layers = list(layer.all_layers) # self.all_params = list(layer.all_params) # self.all_drop = dict(layer.all_drop) # self.all_layers.extend( [self.outputs] ) # self.all_params.extend( variables ) # # self.all_params.extend( [beta, gamma] ) # # class BatchNormLayer4(Layer): # work TFlearn https://github.com/tflearn/tflearn/blob/master/tflearn/layers/normalization.py # """ # The :class:`BatchNormLayer` class is a normalization layer, see ``tf.nn.batch_normalization`` and ``tf.nn.moments``. # # Batch normalization on fully-connected or convolutional maps. # # Parameters # ----------- # layer : a :class:`Layer` instance # The `Layer` class feeding into this layer. # decay : float # A decay factor for ExponentialMovingAverage. # epsilon : float # A small float number to avoid dividing by 0. # act : activation function. # is_train : boolean # Whether train or inference. # beta_init : beta initializer # The initializer for initializing beta # gamma_init : gamma initializer # The initializer for initializing gamma # name : a string or None # An optional name to attach to this layer. # # References # ---------- # - `Source <https://github.com/ry/tensorflow-resnet/blob/master/resnet.py>`_ # - `stackoverflow <http://stackoverflow.com/questions/38312668/how-does-one-do-inference-with-batch-normalization-with-tensor-flow>`_ # """ # def __init__( # self, # layer = None, # decay = 0.999, # epsilon = 0.00001, # act = tf.identity, # is_train = None, # beta_init = tf.zeros_initializer, # # gamma_init = tf.ones_initializer, # gamma_init = tf.random_normal_initializer(mean=1.0, stddev=0.002), # name ='batchnorm_layer', # ): # Layer.__init__(self, name=name) # self.inputs = layer.outputs # print(" [TL] BatchNormLayer %s: decay: %f, epsilon: %f, act: %s, is_train: %s" % # (self.name, decay, epsilon, act.__name__, is_train)) # input_shape = self.inputs.get_shape() # # params_shape = input_shape[-1:] # input_ndim = len(input_shape) # from tensorflow.python.training import moving_averages # from tensorflow.python.ops import control_flow_ops # # # gamma_init = tf.random_normal_initializer(mean=gamma, stddev=stddev) # # # Variable Scope fix for older TF # scope = name # try: # vscope = tf.variable_scope(scope, default_name=name, values=[self.inputs],) # # reuse=reuse) # except Exception: # vscope = tf.variable_op_scope([self.inputs], scope, name)#, reuse=reuse) # # with vscope as scope: # name = scope.name # # with tf.variable_scope(name) as vs: # beta = tf.get_variable('beta', shape=[input_shape[-1]], # initializer=beta_init,) # # initializer=tf.constant_initializer(beta),) # # trainable=trainable, )#restore=restore) # gamma = tf.get_variable('gamma', shape=[input_shape[-1]], # initializer=gamma_init, )#trainable=trainable,) # # restore=restore) # # axis = list(range(input_ndim - 1)) # moving_mean = tf.get_variable('moving_mean', # input_shape[-1:], # initializer=tf.zeros_initializer, # trainable=False,) # # restore=restore) # moving_variance = tf.get_variable('moving_variance', # input_shape[-1:], # initializer=tf.constant_initializer(1.), # trainable=False,) # # restore=restore) # # # Define a function to update mean and variance # def update_mean_var(): # mean, variance = tf.nn.moments(self.inputs, axis) # # # Fix TF 0.12 # try: # update_moving_mean = moving_averages.assign_moving_average( # moving_mean, mean, decay, zero_debias=False) # if zero_debias=True, accuracy is high .. # update_moving_variance = moving_averages.assign_moving_average( # moving_variance, variance, decay, zero_debias=False) # except Exception as e: # TF 11 # update_moving_mean = moving_averages.assign_moving_average( # moving_mean, mean, decay) # update_moving_variance = moving_averages.assign_moving_average( # moving_variance, variance, decay) # # with tf.control_dependencies( # [update_moving_mean, update_moving_variance]): # return tf.identity(mean), tf.identity(variance) # # # Retrieve variable managing training mode # # is_training = tflearn.get_training_mode() # if not is_train: # test : mean=0, std=1 # # if is_train: # train : mean=0, std=1 # is_training = tf.cast(tf.ones([]), tf.bool) # else: # is_training = tf.cast(tf.zeros([]), tf.bool) # mean, var = tf.cond( # is_training, update_mean_var, lambda: (moving_mean, moving_variance)) # # ones zeros # try: # inference = tf.nn.batch_normalization( # self.inputs, mean, var, beta, gamma, epsilon) # inference.set_shape(input_shape) # # Fix for old Tensorflow # except Exception as e: # inference = tf.nn.batch_norm_with_global_normalization( # self.inputs, mean, var, beta, gamma, epsilon, # scale_after_normalization=True, # ) # inference.set_shape(input_shape) # # variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope.name) # 2 params beta, gamma # # variables = [beta, gamma, moving_mean, moving_variance] # # # print(len(variables)) # # for idx, v in enumerate(variables): # # print(" var {:3}: {:15} {}".format(idx, str(v.get_shape()), v.name)) # # exit() # # # Add attributes for easy access # # inference.scope = scope # inference.scope = name # inference.beta = beta # inference.gamma = gamma # # self.outputs = act( inference ) # # self.all_layers = list(layer.all_layers) # self.all_params = list(layer.all_params) # self.all_drop = dict(layer.all_drop) # self.all_layers.extend( [self.outputs] ) # self.all_params.extend( variables ) # class BatchNormLayer2(Layer): # don't work http://r2rt.com/implementing-batch-normalization-in-tensorflow.html # """ # The :class:`BatchNormLayer` class is a normalization layer, see ``tf.nn.batch_normalization`` and ``tf.nn.moments``. # # Batch normalization on fully-connected or convolutional maps. # # Parameters # ----------- # layer : a :class:`Layer` instance # The `Layer` class feeding into this layer. # decay : float # A decay factor for ExponentialMovingAverage. # epsilon : float # A small float number to avoid dividing by 0. # act : activation function. # is_train : boolean # Whether train or inference. # beta_init : beta initializer # The initializer for initializing beta # gamma_init : gamma initializer # The initializer for initializing gamma # name : a string or None # An optional name to attach to this layer. # # References # ---------- # - `Source <https://github.com/ry/tensorflow-resnet/blob/master/resnet.py>`_ # - `stackoverflow <http://stackoverflow.com/questions/38312668/how-does-one-do-inference-with-batch-normalization-with-tensor-flow>`_ # """ # def __init__( # self, # layer = None, # decay = 0.999, # epsilon = 0.00001, # act = tf.identity, # is_train = None, # beta_init = tf.zeros_initializer, # # gamma_init = tf.ones_initializer, # gamma_init = tf.random_normal_initializer(mean=1.0, stddev=0.002), # name ='batchnorm_layer', # ): # Layer.__init__(self, name=name) # self.inputs = layer.outputs # print(" [TL] BatchNormLayer %s: decay: %f, epsilon: %f, act: %s, is_train: %s" % # (self.name, decay, epsilon, act.__name__, is_train)) # x_shape = self.inputs.get_shape() # params_shape = x_shape[-1:] # # with tf.variable_scope(name) as vs: # gamma = tf.get_variable("gamma", shape=params_shape, # initializer=gamma_init) # beta = tf.get_variable("beta", shape=params_shape, # initializer=beta_init) # pop_mean = tf.get_variable("pop_mean", shape=params_shape, # initializer=tf.zeros_initializer, trainable=False) # pop_var = tf.get_variable("pop_var", shape=params_shape, # initializer=tf.constant_initializer(1.), trainable=False) # # if is_train: # batch_mean, batch_var = tf.nn.moments(self.inputs, list(range(len(x_shape) - 1))) # train_mean = tf.assign(pop_mean, # pop_mean * decay + batch_mean * (1 - decay)) # train_var = tf.assign(pop_var, # pop_var * decay + batch_var * (1 - decay)) # with tf.control_dependencies([train_mean, train_var]): # self.outputs = act(tf.nn.batch_normalization(self.inputs, # batch_mean, batch_var, beta, gamma, epsilon)) # else: # self.outputs = act(tf.nn.batch_normalization(self.inputs, # pop_mean, pop_var, beta, gamma, epsilon)) # # self.outputs = act( tf.nn.batch_normalization(self.inputs, mean, variance, beta, gamma, epsilon) ) # # variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) # 8 params in TF12 if zero_debias=True # variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=vs.name) # 2 params beta, gamma # # variables = [beta, gamma, moving_mean, moving_variance] # # # print(len(variables)) # # for idx, v in enumerate(variables): # # print(" var {:3}: {:15} {}".format(idx, str(v.get_shape()), v.name)) # # exit() # # self.all_layers = list(layer.all_layers) # self.all_params = list(layer.all_params) # self.all_drop = dict(layer.all_drop) # self.all_layers.extend( [self.outputs] ) # self.all_params.extend( variables ) # class BatchNormLayer3(Layer): # don't work http://r2rt.com/implementing-batch-normalization-in-tensorflow.html # """ # The :class:`BatchNormLayer` class is a normalization layer, see ``tf.nn.batch_normalization`` and ``tf.nn.moments``. # # Batch normalization on fully-connected or convolutional maps. # # Parameters # ----------- # layer : a :class:`Layer` instance # The `Layer` class feeding into this layer. # decay : float # A decay factor for ExponentialMovingAverage. # epsilon : float # A small float number to avoid dividing by 0. # act : activation function. # is_train : boolean # Whether train or inference. # beta_init : beta initializer # The initializer for initializing beta # gamma_init : gamma initializer # The initializer for initializing gamma # name : a string or None # An optional name to attach to this layer. # # References # ---------- # - `Source <https://github.com/ry/tensorflow-resnet/blob/master/resnet.py>`_ # - `stackoverflow <http://stackoverflow.com/questions/38312668/how-does-one-do-inference-with-batch-normalization-with-tensor-flow>`_ # """ # def __init__( # self, # layer = None, # decay = 0.999, # epsilon = 0.00001, # act = tf.identity, # is_train = None, # beta_init = tf.zeros_initializer, # # gamma_init = tf.ones_initializer, # gamma_init = tf.random_normal_initializer(mean=1.0, stddev=0.002), # name ='batchnorm_layer', # ): # """ # Batch normalization on convolutional maps. # Ref.: http://stackoverflow.com/questions/33949786/how-could-i-use-batch-normalization-in-tensorflow # Args: # x: Tensor, 4D BHWD input maps # n_out: integer, depth of input maps # phase_train: boolean tf.Varialbe, true indicates training phase # scope: string, variable scope # Return: # normed: batch-normalized maps # """ # Layer.__init__(self, name=name) # self.inputs = layer.outputs # print(" [TL] BatchNormLayer %s: decay: %f, epsilon: %f, act: %s, is_train: %s" % # (self.name, decay, epsilon, act.__name__, is_train)) # x_shape = self.inputs.get_shape() # params_shape = x_shape[-1:] # # if is_train: # phase_train = tf.cast(tf.ones([]), tf.bool) # else: # phase_train = tf.cast(tf.zeros([]), tf.bool) # # with tf.variable_scope(name) as vs: # gamma = tf.get_variable("gamma", shape=params_shape, # initializer=gamma_init) # beta = tf.get_variable("beta", shape=params_shape, # initializer=beta_init) # batch_mean, batch_var = tf.nn.moments(self.inputs, list(range(len(x_shape) - 1)),#[0,1,2], # name='moments') # ema = tf.train.ExponentialMovingAverage(decay=decay) # # def mean_var_with_update(): # ema_apply_op = ema.apply([batch_mean, batch_var]) # with tf.control_dependencies([ema_apply_op]): # return tf.identity(batch_mean), tf.identity(batch_var) # # mean, var = tf.cond(phase_train, # mean_var_with_update, # lambda: (ema.average(batch_mean), ema.average(batch_var))) # normed = tf.nn.batch_normalization(self.inputs, mean, var, beta, gamma, epsilon) # self.outputs = act( normed ) # variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=vs.name) # 2 params beta, gamma # # variables = [beta, gamma, moving_mean, moving_variance] # # # print(len(variables)) # # for idx, v in enumerate(variables): # # print(" var {:3}: {:15} {}".format(idx, str(v.get_shape()), v.name)) # # exit() # # self.all_layers = list(layer.all_layers) # self.all_params = list(layer.all_params) # self.all_drop = dict(layer.all_drop) # self.all_layers.extend( [self.outputs] ) # self.all_params.extend( variables ) # class BatchNormLayer_old(Layer): # don't work # """ # The :class:`BatchNormLayer` class is a normalization layer, see ``tf.nn.batch_normalization``. # # Batch normalization on fully-connected or convolutional maps. # # Parameters # ----------- # layer : a :class:`Layer` instance # The `Layer` class feeding into this layer. # decay : float # A decay factor for ExponentialMovingAverage. # epsilon : float # A small float number to avoid dividing by 0. # is_train : boolean # Whether train or inference. # name : a string or None # An optional name to attach to this layer. # # References # ---------- # - `tf.nn.batch_normalization <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/g3doc/api_docs/python/functions_and_classes/shard8/tf.nn.batch_normalization.md>`_ # - `stackoverflow <http://stackoverflow.com/questions/33949786/how-could-i-use-batch-normalization-in-tensorflow>`_ # - `tensorflow.contrib <https://github.com/tensorflow/tensorflow/blob/b826b79718e3e93148c3545e7aa3f90891744cc0/tensorflow/contrib/layers/python/layers/layers.py#L100>`_ # """ # def __init__( # self, # layer = None, # act = tf.identity, # decay = 0.999, # epsilon = 0.001, # is_train = None, # name ='batchnorm_layer', # ): # Layer.__init__(self, name=name) # self.inputs = layer.outputs # print(" [TL] BatchNormLayer %s: decay: %f, epsilon: %f, is_train: %s" % # (self.name, decay, epsilon, is_train)) # if is_train == None: # raise Exception("is_train must be True or False") # # # (name, input_var, decay, epsilon, is_train) # inputs_shape = self.inputs.get_shape() # axis = list(range(len(inputs_shape) - 1)) # params_shape = inputs_shape[-1:] # # with tf.variable_scope(name) as vs: # beta = tf.get_variable(name='beta', shape=params_shape, # initializer=tf.constant_initializer(0.0)) # gamma = tf.get_variable(name='gamma', shape=params_shape, # initializer=tf.constant_initializer(1.0)) # batch_mean, batch_var = tf.nn.moments(self.inputs, # axis, # name='moments') # ema = tf.train.ExponentialMovingAverage(decay=decay) # # def mean_var_with_update(): # ema_apply_op = ema.apply([batch_mean, batch_var]) # with tf.control_dependencies([ema_apply_op]): # return tf.identity(batch_mean), tf.identity(batch_var) # # if is_train: # is_train = tf.cast(tf.ones(1), tf.bool) # else: # is_train = tf.cast(tf.zeros(1), tf.bool) # # is_train = tf.reshape(is_train, []) # # # print(is_train) # # exit() # # mean, var = tf.cond( # is_train, # mean_var_with_update, # lambda: (ema.average(batch_mean), ema.average(batch_var)) # ) # normed = tf.nn.batch_normalization( # x=self.inputs, # mean=mean, # variance=var, # offset=beta, # scale=gamma, # variance_epsilon=epsilon, # name='tf_bn' # ) # self.outputs = act( normed ) # # self.all_layers = list(layer.all_layers) # self.all_params = list(layer.all_params) # self.all_drop = dict(layer.all_drop) # self.all_layers.extend( [self.outputs] ) # self.all_params.extend( [beta, gamma] ) ## Pooling layer
[docs]class PoolLayer(Layer): """ The :class:`PoolLayer` class is a Pooling layer, you can choose ``tf.nn.max_pool`` and ``tf.nn.avg_pool`` for 2D or ``tf.nn.max_pool3d`` and ``tf.nn.avg_pool3d`` for 3D. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. ksize : a list of ints that has length >= 4. The size of the window for each dimension of the input tensor. strides : a list of ints that has length >= 4. The stride of the sliding window for each dimension of the input tensor. padding : a string from: "SAME", "VALID". The type of padding algorithm to use. pool : a pooling function - see `TensorFlow pooling APIs <https://www.tensorflow.org/versions/master/api_docs/python/nn.html#pooling>`_ - class ``tf.nn.max_pool`` - class ``tf.nn.avg_pool`` - class ``tf.nn.max_pool3d`` - class ``tf.nn.avg_pool3d`` name : a string or None An optional name to attach to this layer. Examples -------- - see :class:`Conv2dLayer`. """ def __init__( self, layer = None, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', pool = tf.nn.max_pool, name ='pool_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] PoolLayer %s: ksize:%s strides:%s padding:%s pool:%s" % (self.name, str(ksize), str(strides), padding, pool.__name__)) self.outputs = pool(self.inputs, ksize=ksize, strides=strides, padding=padding, name=name) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
## Padding layer
[docs]class PadLayer(Layer): """ The :class:`PadLayer` class is a Padding layer for any modes and dimensions. Please see `tf.pad <https://www.tensorflow.org/api_docs/python/tf/pad>`_ for usage. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. padding : a Tensor of type int32. mode : one of "CONSTANT", "REFLECT", or "SYMMETRIC" (case-insensitive) name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, paddings = None, mode = 'CONSTANT', name = 'pad_layer', ): Layer.__init__(self, name=name) assert paddings is not None, "paddings should be a Tensor of type int32. see https://www.tensorflow.org/api_docs/python/tf/pad" self.inputs = layer.outputs print(" [TL] PadLayer %s: paddings:%s mode:%s" % (self.name, list(paddings.get_shape()), mode)) self.outputs = tf.pad(self.inputs, paddings=paddings, mode=mode, name=name) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
## TimeDistributedLayer
[docs]class TimeDistributedLayer(Layer): """ The :class:`TimeDistributedLayer` class that applies a function to every timestep of the input tensor. For example, if using :class:`DenseLayer` as the ``layer_class``, inputs [batch_size , length, dim] outputs [batch_size , length, new_dim]. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer, [batch_size , length, dim] layer_class : a :class:`Layer` class args : dictionary The arguments for the ``layer_class``. name : a string or None An optional name to attach to this layer. Examples -------- >>> batch_size = 32 >>> timestep = 20 >>> input_dim = 100 >>> x = tf.placeholder(dtype=tf.float32, shape=[batch_size, timestep, input_dim], name="encode_seqs") >>> net = InputLayer(x, name='input') >>> net = TimeDistributedLayer(net, layer_class=DenseLayer, args={'n_units':50, 'name':'dense'}, name='time_dense') ... [TL] InputLayer input: (32, 20, 100) ... [TL] TimeDistributedLayer time_dense: layer_class:DenseLayer >>> print(net.outputs._shape) ... (32, 20, 50) >>> net.print_params(False) ... param 0: (100, 50) time_dense/dense/W:0 ... param 1: (50,) time_dense/dense/b:0 ... num of params: 5050 """ def __init__( self, layer = None, layer_class = None, args = {}, name ='time_distributed', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] TimeDistributedLayer %s: layer_class:%s args:%s" % (self.name, layer_class.__name__, args)) if not args: args = dict() assert isinstance(args, dict), "'args' must be a dict." if not isinstance(self.inputs, tf.Tensor): self.inputs = tf.transpose(tf.stack(self.inputs), [1, 0, 2]) input_shape = self.inputs.get_shape() timestep = input_shape[1] x = tf.unstack(self.inputs, axis=1) with ops.suppress_stdout(): for i in range(0, timestep): with tf.variable_scope(name, reuse=(False if i==0 else True)) as vs: set_name_reuse((False if i==0 else True)) net = layer_class(InputLayer(x[i], name=args['name']+str(i)), **args) # net = layer_class(InputLayer(x[i], name="input_"+args['name']), **args) x[i] = net.outputs variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) self.outputs = tf.stack(x, axis=1, name=name) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( variables )
## Recurrent layer
[docs]class RNNLayer(Layer): """ The :class:`RNNLayer` class is a RNN layer, you can implement vanilla RNN, LSTM and GRU with it. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. cell_fn : a TensorFlow's core RNN cell as follow (Note TF1.0+ and TF1.0- are different). - see `RNN Cells in TensorFlow <https://www.tensorflow.org/api_docs/python/>`_ cell_init_args : a dictionary The arguments for the cell initializer. n_hidden : a int The number of hidden units in the layer. initializer : initializer The initializer for initializing the parameters. n_steps : a int The sequence length. initial_state : None or RNN State If None, initial_state is zero_state. return_last : boolean - If True, return the last output, "Sequence input and single output" - If False, return all outputs, "Synced sequence input and output" - In other word, if you want to apply one or more RNN(s) on this layer, set to False. return_seq_2d : boolean - When return_last = False - If True, return 2D Tensor [n_example, n_hidden], for stacking DenseLayer after it. - If False, return 3D Tensor [n_example/n_steps, n_steps, n_hidden], for stacking multiple RNN after it. name : a string or None An optional name to attach to this layer. Variables -------------- outputs : a tensor The output of this RNN. return_last = False, outputs = all cell_output, which is the hidden state. cell_output.get_shape() = (?, n_hidden) final_state : a tensor or StateTuple When state_is_tuple = False, it is the final hidden and cell states, states.get_shape() = [?, 2 * n_hidden].\n When state_is_tuple = True, it stores two elements: (c, h), in that order. You can get the final state after each iteration during training, then feed it to the initial state of next iteration. initial_state : a tensor or StateTuple It is the initial state of this RNN layer, you can use it to initialize your state at the begining of each epoch or iteration according to your training procedure. batch_size : int or tensor Is int, if able to compute the batch_size, otherwise, tensor for ``?``. Examples -------- - For words >>> input_data = tf.placeholder(tf.int32, [batch_size, num_steps]) >>> network = tl.layers.EmbeddingInputlayer( ... inputs = input_data, ... vocabulary_size = vocab_size, ... embedding_size = hidden_size, ... E_init = tf.random_uniform_initializer(-init_scale, init_scale), ... name ='embedding_layer') >>> if is_training: >>> network = tl.layers.DropoutLayer(network, keep=keep_prob, name='drop1') >>> network = tl.layers.RNNLayer(network, ... cell_fn=tf.nn.rnn_cell.BasicLSTMCell, ... cell_init_args={'forget_bias': 0.0},# 'state_is_tuple': True}, ... n_hidden=hidden_size, ... initializer=tf.random_uniform_initializer(-init_scale, init_scale), ... n_steps=num_steps, ... return_last=False, ... name='basic_lstm_layer1') >>> lstm1 = network >>> if is_training: >>> network = tl.layers.DropoutLayer(network, keep=keep_prob, name='drop2') >>> network = tl.layers.RNNLayer(network, ... cell_fn=tf.nn.rnn_cell.BasicLSTMCell, ... cell_init_args={'forget_bias': 0.0}, # 'state_is_tuple': True}, ... n_hidden=hidden_size, ... initializer=tf.random_uniform_initializer(-init_scale, init_scale), ... n_steps=num_steps, ... return_last=False, ... return_seq_2d=True, ... name='basic_lstm_layer2') >>> lstm2 = network >>> if is_training: >>> network = tl.layers.DropoutLayer(network, keep=keep_prob, name='drop3') >>> network = tl.layers.DenseLayer(network, ... n_units=vocab_size, ... W_init=tf.random_uniform_initializer(-init_scale, init_scale), ... b_init=tf.random_uniform_initializer(-init_scale, init_scale), ... act = tl.activation.identity, name='output_layer') - For CNN+LSTM >>> x = tf.placeholder(tf.float32, shape=[batch_size, image_size, image_size, 1]) >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.Conv2dLayer(network, ... act = tf.nn.relu, ... shape = [5, 5, 1, 32], # 32 features for each 5x5 patch ... strides=[1, 2, 2, 1], ... padding='SAME', ... name ='cnn_layer1') >>> network = tl.layers.PoolLayer(network, ... ksize=[1, 2, 2, 1], ... strides=[1, 2, 2, 1], ... padding='SAME', ... pool = tf.nn.max_pool, ... name ='pool_layer1') >>> network = tl.layers.Conv2dLayer(network, ... act = tf.nn.relu, ... shape = [5, 5, 32, 10], # 10 features for each 5x5 patch ... strides=[1, 2, 2, 1], ... padding='SAME', ... name ='cnn_layer2') >>> network = tl.layers.PoolLayer(network, ... ksize=[1, 2, 2, 1], ... strides=[1, 2, 2, 1], ... padding='SAME', ... pool = tf.nn.max_pool, ... name ='pool_layer2') >>> network = tl.layers.FlattenLayer(network, name='flatten_layer') >>> network = tl.layers.ReshapeLayer(network, shape=[-1, num_steps, int(network.outputs._shape[-1])]) >>> rnn1 = tl.layers.RNNLayer(network, ... cell_fn=tf.nn.rnn_cell.LSTMCell, ... cell_init_args={}, ... n_hidden=200, ... initializer=tf.random_uniform_initializer(-0.1, 0.1), ... n_steps=num_steps, ... return_last=False, ... return_seq_2d=True, ... name='rnn_layer') >>> network = tl.layers.DenseLayer(rnn1, n_units=3, ... act = tl.activation.identity, name='output_layer') Notes ----- Input dimension should be rank 3 : [batch_size, n_steps, n_features], if no, please see :class:`ReshapeLayer`. References ---------- - `Neural Network RNN Cells in TensorFlow <https://www.tensorflow.org/api_docs/python/rnn_cell/>`_ - `tensorflow/python/ops/rnn.py <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/rnn.py>`_ - `tensorflow/python/ops/rnn_cell.py <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/rnn_cell.py>`_ - see TensorFlow tutorial ``ptb_word_lm.py``, TensorLayer tutorials ``tutorial_ptb_lstm*.py`` and ``tutorial_generate_text.py`` """ def __init__( self, layer = None, cell_fn = None,#tf.nn.rnn_cell.BasicRNNCell, cell_init_args = {}, n_hidden = 100, initializer = tf.random_uniform_initializer(-0.1, 0.1), n_steps = 5, initial_state = None, return_last = False, # is_reshape = True, return_seq_2d = False, name = 'rnn_layer', ): Layer.__init__(self, name=name) if cell_fn is None: raise Exception("Please put in cell_fn") if 'GRU' in cell_fn.__name__: try: cell_init_args.pop('state_is_tuple') except: pass self.inputs = layer.outputs print(" [TL] RNNLayer %s: n_hidden:%d n_steps:%d in_dim:%d in_shape:%s cell_fn:%s " % (self.name, n_hidden, n_steps, self.inputs.get_shape().ndims, self.inputs.get_shape(), cell_fn.__name__)) # You can get the dimension by .get_shape() or ._shape, and check the # dimension by .with_rank() as follow. # self.inputs.get_shape().with_rank(2) # self.inputs.get_shape().with_rank(3) # Input dimension should be rank 3 [batch_size, n_steps(max), n_features] try: self.inputs.get_shape().with_rank(3) except: raise Exception("RNN : Input dimension should be rank 3 : [batch_size, n_steps, n_features]") # is_reshape : boolean (deprecate) # Reshape the inputs to 3 dimension tensor.\n # If input is´╝╗batch_size, n_steps, n_features], we do not need to reshape it.\n # If input is [batch_size * n_steps, n_features], we need to reshape it. # if is_reshape: # self.inputs = tf.reshape(self.inputs, shape=[-1, n_steps, int(self.inputs._shape[-1])]) fixed_batch_size = self.inputs.get_shape().with_rank_at_least(1)[0] if fixed_batch_size.value: batch_size = fixed_batch_size.value print(" RNN batch_size (concurrent processes): %d" % batch_size) else: from tensorflow.python.ops import array_ops batch_size = array_ops.shape(self.inputs)[0] print(" non specified batch_size, uses a tensor instead.") self.batch_size = batch_size # Simplified version of tensorflow.models.rnn.rnn.py's rnn(). # This builds an unrolled LSTM for tutorial purposes only. # In general, use the rnn() or state_saving_rnn() from rnn.py. # # The alternative version of the code below is: # # from tensorflow.models.rnn import rnn # inputs = [tf.squeeze(input_, [1]) # for input_ in tf.split(1, num_steps, inputs)] # outputs, state = rnn.rnn(cell, inputs, initial_state=self._initial_state) outputs = [] if 'reuse' in inspect.getargspec(cell_fn.__init__).args: self.cell = cell = cell_fn(num_units=n_hidden, reuse=tf.get_variable_scope().reuse, **cell_init_args) else: self.cell = cell = cell_fn(num_units=n_hidden, **cell_init_args) if initial_state is None: self.initial_state = cell.zero_state(batch_size, dtype=tf.float32) # 1.2.3 state = self.initial_state # with tf.variable_scope("model", reuse=None, initializer=initializer): with tf.variable_scope(name, initializer=initializer) as vs: for time_step in range(n_steps): if time_step > 0: tf.get_variable_scope().reuse_variables() (cell_output, state) = cell(self.inputs[:, time_step, :], state) outputs.append(cell_output) # Retrieve just the RNN variables. # rnn_variables = [v for v in tf.all_variables() if v.name.startswith(vs.name)] rnn_variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) print(" n_params : %d" % (len(rnn_variables))) if return_last: # 2D Tensor [batch_size, n_hidden] self.outputs = outputs[-1] else: if return_seq_2d: # PTB tutorial: stack dense layer after that, or compute the cost from the output # 2D Tensor [n_example, n_hidden] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs, 1), [-1, n_hidden]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [-1, n_hidden]) else: # <akara>: stack more RNN layer after that # 3D Tensor [n_example/n_steps, n_steps, n_hidden] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs, 1), [-1, n_steps, n_hidden]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [-1, n_steps, n_hidden]) self.final_state = state self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) # print(type(self.outputs)) self.all_layers.extend( [self.outputs] ) self.all_params.extend( rnn_variables )
[docs]class BiRNNLayer(Layer): """ The :class:`BiRNNLayer` class is a Bidirectional RNN layer. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. cell_fn : a TensorFlow's core RNN cell as follow (Note TF1.0+ and TF1.0- are different). - see `RNN Cells in TensorFlow <https://www.tensorflow.org/api_docs/python/>`_ cell_init_args : a dictionary The arguments for the cell initializer. n_hidden : a int The number of hidden units in the layer. initializer : initializer The initializer for initializing the parameters. n_steps : a int The sequence length. fw_initial_state : None or forward RNN State If None, initial_state is zero_state. bw_initial_state : None or backward RNN State If None, initial_state is zero_state. dropout : `tuple` of `float`: (input_keep_prob, output_keep_prob). The input and output keep probability. n_layer : a int, default is 1. The number of RNN layers. return_last : boolean - If True, return the last output, "Sequence input and single output" - If False, return all outputs, "Synced sequence input and output" - In other word, if you want to apply one or more RNN(s) on this layer, set to False. return_seq_2d : boolean - When return_last = False - If True, return 2D Tensor [n_example, n_hidden], for stacking DenseLayer after it. - If False, return 3D Tensor [n_example/n_steps, n_steps, n_hidden], for stacking multiple RNN after it. name : a string or None An optional name to attach to this layer. Variables -------------- outputs : a tensor The output of this RNN. return_last = False, outputs = all cell_output, which is the hidden state. cell_output.get_shape() = (?, n_hidden) fw(bw)_final_state : a tensor or StateTuple When state_is_tuple = False, it is the final hidden and cell states, states.get_shape() = [?, 2 * n_hidden].\n When state_is_tuple = True, it stores two elements: (c, h), in that order. You can get the final state after each iteration during training, then feed it to the initial state of next iteration. fw(bw)_initial_state : a tensor or StateTuple It is the initial state of this RNN layer, you can use it to initialize your state at the begining of each epoch or iteration according to your training procedure. batch_size : int or tensor Is int, if able to compute the batch_size, otherwise, tensor for ``?``. Notes ----- - Input dimension should be rank 3 : [batch_size, n_steps, n_features], if no, please see :class:`ReshapeLayer`. - For predicting, the sequence length has to be the same with the sequence length of training, while, for normal RNN, we can use sequence length of 1 for predicting. References ---------- - `Source <https://github.com/akaraspt/deepsleep/blob/master/deepsleep/model.py>`_ """ def __init__( self, layer = None, cell_fn = None, #tf.nn.rnn_cell.LSTMCell, cell_init_args = {'use_peepholes':True, 'state_is_tuple':True}, n_hidden = 100, initializer = tf.random_uniform_initializer(-0.1, 0.1), n_steps = 5, fw_initial_state = None, bw_initial_state = None, dropout = None, n_layer = 1, return_last = False, return_seq_2d = False, name = 'birnn_layer', ): Layer.__init__(self, name=name) if cell_fn is None: raise Exception("Please put in cell_fn") if 'GRU' in cell_fn.__name__: try: cell_init_args.pop('state_is_tuple') except: pass self.inputs = layer.outputs print(" [TL] BiRNNLayer %s: n_hidden:%d n_steps:%d in_dim:%d in_shape:%s cell_fn:%s dropout:%s n_layer:%d " % (self.name, n_hidden, n_steps, self.inputs.get_shape().ndims, self.inputs.get_shape(), cell_fn.__name__, dropout, n_layer)) fixed_batch_size = self.inputs.get_shape().with_rank_at_least(1)[0] if fixed_batch_size.value: self.batch_size = fixed_batch_size.value print(" RNN batch_size (concurrent processes): %d" % self.batch_size) else: from tensorflow.python.ops import array_ops self.batch_size = array_ops.shape(self.inputs)[0] print(" non specified batch_size, uses a tensor instead.") # Input dimension should be rank 3 [batch_size, n_steps(max), n_features] try: self.inputs.get_shape().with_rank(3) except: raise Exception("RNN : Input dimension should be rank 3 : [batch_size, n_steps, n_features]") with tf.variable_scope(name, initializer=initializer) as vs: rnn_creator = lambda: cell_fn(num_units=n_hidden, **cell_init_args) # Apply dropout if dropout: if type(dropout) in [tuple, list]: in_keep_prob = dropout[0] out_keep_prob = dropout[1] elif isinstance(dropout, float): in_keep_prob, out_keep_prob = dropout, dropout else: raise Exception("Invalid dropout type (must be a 2-D tuple of " "float)") try: # TF 1.0 DropoutWrapper_fn = tf.contrib.rnn.DropoutWrapper except: DropoutWrapper_fn = tf.nn.rnn_cell.DropoutWrapper cell_creator = lambda: DropoutWrapper_fn(rnn_creator(), input_keep_prob=in_keep_prob, output_keep_prob=1.0) # out_keep_prob) else: cell_creator = rnn_creator self.fw_cell = cell_creator() self.bw_cell = cell_creator() # Apply multiple layers if n_layer > 1: try: # TF1.0 MultiRNNCell_fn = tf.contrib.rnn.MultiRNNCell except: MultiRNNCell_fn = tf.nn.rnn_cell.MultiRNNCell try: self.fw_cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)], state_is_tuple=True) self.bw_cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)], state_is_tuple=True) except: self.fw_cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)]) self.bw_cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)]) # Initial state of RNN if fw_initial_state is None: self.fw_initial_state = self.fw_cell.zero_state(self.batch_size, dtype=tf.float32) else: self.fw_initial_state = fw_initial_state if bw_initial_state is None: self.bw_initial_state = self.bw_cell.zero_state(self.batch_size, dtype=tf.float32) else: self.bw_initial_state = bw_initial_state # exit() # Feedforward to MultiRNNCell try: ## TF1.0 list_rnn_inputs = tf.unstack(self.inputs, axis=1) except: ## TF0.12 list_rnn_inputs = tf.unpack(self.inputs, axis=1) try: # TF1.0 bidirectional_rnn_fn = tf.contrib.rnn.static_bidirectional_rnn except: bidirectional_rnn_fn = tf.nn.bidirectional_rnn outputs, fw_state, bw_state = bidirectional_rnn_fn( # outputs, fw_state, bw_state = tf.contrib.rnn.static_bidirectional_rnn( cell_fw=self.fw_cell, cell_bw=self.bw_cell, inputs=list_rnn_inputs, initial_state_fw=self.fw_initial_state, initial_state_bw=self.bw_initial_state ) if return_last: self.outputs = outputs[-1] else: self.outputs = outputs if return_seq_2d: # 2D Tensor [n_example, n_hidden] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs, 1), [-1, n_hidden*2]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [-1, n_hidden*2]) else: # <akara>: stack more RNN layer after that # 3D Tensor [n_example/n_steps, n_steps, n_hidden] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs,1), [-1, n_steps, n_hidden*2]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [-1, n_steps, n_hidden*2]) self.fw_final_state = fw_state self.bw_final_state = bw_state # Retrieve just the RNN variables. rnn_variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) print(" n_params : %d" % (len(rnn_variables))) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( rnn_variables )
# Advanced Ops for Dynamic RNN
[docs]def advanced_indexing_op(input, index): """Advanced Indexing for Sequences, returns the outputs by given sequence lengths. When return the last output :class:`DynamicRNNLayer` uses it to get the last outputs with the sequence lengths. Parameters ----------- input : tensor for data [batch_size, n_step(max), n_features] index : tensor for indexing, i.e. sequence_length in Dynamic RNN. [batch_size] Examples --------- >>> batch_size, max_length, n_features = 3, 5, 2 >>> z = np.random.uniform(low=-1, high=1, size=[batch_size, max_length, n_features]).astype(np.float32) >>> b_z = tf.constant(z) >>> sl = tf.placeholder(dtype=tf.int32, shape=[batch_size]) >>> o = advanced_indexing_op(b_z, sl) >>> >>> sess = tf.InteractiveSession() >>> tl.layers.initialize_global_variables(sess) >>> >>> order = np.asarray([1,1,2]) >>> print("real",z[0][order[0]-1], z[1][order[1]-1], z[2][order[2]-1]) >>> y = sess.run([o], feed_dict={sl:order}) >>> print("given",order) >>> print("out", y) ... real [-0.93021595 0.53820813] [-0.92548317 -0.77135968] [ 0.89952248 0.19149846] ... given [1 1 2] ... out [array([[-0.93021595, 0.53820813], ... [-0.92548317, -0.77135968], ... [ 0.89952248, 0.19149846]], dtype=float32)] References ----------- - Modified from TFlearn (the original code is used for fixed length rnn), `references <https://github.com/tflearn/tflearn/blob/master/tflearn/layers/recurrent.py>`_. """ batch_size = tf.shape(input)[0] # max_length = int(input.get_shape()[1]) # for fixed length rnn, length is given max_length = tf.shape(input)[1] # for dynamic_rnn, length is unknown dim_size = int(input.get_shape()[2]) index = tf.range(0, batch_size) * max_length + (index - 1) flat = tf.reshape(input, [-1, dim_size]) relevant = tf.gather(flat, index) return relevant
[docs]def retrieve_seq_length_op(data): """An op to compute the length of a sequence from input shape of [batch_size, n_step(max), n_features], it can be used when the features of padding (on right hand side) are all zeros. Parameters ----------- data : tensor [batch_size, n_step(max), n_features] with zero padding on right hand side. Examples --------- >>> data = [[[1],[2],[0],[0],[0]], ... [[1],[2],[3],[0],[0]], ... [[1],[2],[6],[1],[0]]] >>> data = np.asarray(data) >>> print(data.shape) ... (3, 5, 1) >>> data = tf.constant(data) >>> sl = retrieve_seq_length_op(data) >>> sess = tf.InteractiveSession() >>> tl.layers.initialize_global_variables(sess) >>> y = sl.eval() ... [2 3 4] - Multiple features >>> data = [[[1,2],[2,2],[1,2],[1,2],[0,0]], ... [[2,3],[2,4],[3,2],[0,0],[0,0]], ... [[3,3],[2,2],[5,3],[1,2],[0,0]]] >>> sl ... [4 3 4] References ------------ - Borrow from `TFlearn <https://github.com/tflearn/tflearn/blob/master/tflearn/layers/recurrent.py>`_. """ with tf.name_scope('GetLength'): ## TF 1.0 change reduction_indices to axis used = tf.sign(tf.reduce_max(tf.abs(data), 2)) length = tf.reduce_sum(used, 1) ## TF < 1.0 # used = tf.sign(tf.reduce_max(tf.abs(data), reduction_indices=2)) # length = tf.reduce_sum(used, reduction_indices=1) length = tf.cast(length, tf.int32) return length
[docs]def retrieve_seq_length_op2(data): """An op to compute the length of a sequence, from input shape of [batch_size, n_step(max)], it can be used when the features of padding (on right hand side) are all zeros. Parameters ----------- data : tensor [batch_size, n_step(max)] with zero padding on right hand side. Examples -------- >>> data = [[1,2,0,0,0], ... [1,2,3,0,0], ... [1,2,6,1,0]] >>> o = retrieve_seq_length_op2(data) >>> sess = tf.InteractiveSession() >>> tl.layers.initialize_global_variables(sess) >>> print(o.eval()) ... [2 3 4] """ return tf.reduce_sum(tf.cast(tf.greater(data, tf.zeros_like(data)), tf.int32), 1)
def retrieve_seq_length_op3(data, pad_val=0): # HangSheng: return tensor for sequence length, if input is tf.string data_shape_size = data.get_shape().ndims if data_shape_size == 3: return tf.reduce_sum(tf.cast(tf.reduce_any(tf.not_equal(data, pad_val), axis=2), dtype=tf.int32), 1) elif data_shape_size == 2: return tf.reduce_sum(tf.cast(tf.not_equal(data, pad_val), dtype=tf.int32), 1) elif data_shape_size == 1: raise ValueError("retrieve_seq_length_op3: data has wrong shape!") else: raise ValueError("retrieve_seq_length_op3: handling data_shape_size %s hasn't been implemented!" % (data_shape_size)) def target_mask_op(data, pad_val=0): # HangSheng: return tensor for mask,if input is tf.string data_shape_size = data.get_shape().ndims if data_shape_size == 3: return tf.cast(tf.reduce_any(tf.not_equal(data, pad_val), axis=2), dtype=tf.int32) elif data_shape_size == 2: return tf.cast(tf.not_equal(data, pad_val), dtype=tf.int32) elif data_shape_size == 1: raise ValueError("target_mask_op: data has wrong shape!") else: raise ValueError("target_mask_op: handling data_shape_size %s hasn't been implemented!" % (data_shape_size)) # Dynamic RNN
[docs]class DynamicRNNLayer(Layer): """ The :class:`DynamicRNNLayer` class is a Dynamic RNN layer, see ``tf.nn.dynamic_rnn``. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. cell_fn : a TensorFlow's core RNN cell as follow (Note TF1.0+ and TF1.0- are different). - see `RNN Cells in TensorFlow <https://www.tensorflow.org/api_docs/python/>`_ cell_init_args : a dictionary The arguments for the cell initializer. n_hidden : a int The number of hidden units in the layer. initializer : initializer The initializer for initializing the parameters. sequence_length : a tensor, array or None. The sequence length of each row of input data, see ``Advanced Ops for Dynamic RNN``. - If None, it uses ``retrieve_seq_length_op`` to compute the sequence_length, i.e. when the features of padding (on right hand side) are all zeros. - If using word embedding, you may need to compute the sequence_length from the ID array (the integer features before word embedding) by using ``retrieve_seq_length_op2`` or ``retrieve_seq_length_op``. - You can also input an numpy array. - More details about TensorFlow dynamic_rnn in `Wild-ML Blog <http://www.wildml.com/2016/08/rnns-in-tensorflow-a-practical-guide-and-undocumented-features/>`_. initial_state : None or RNN State If None, initial_state is zero_state. dropout : `tuple` of `float`: (input_keep_prob, output_keep_prob). The input and output keep probability. n_layer : a int, default is 1. The number of RNN layers. return_last : boolean - If True, return the last output, "Sequence input and single output" - If False, return all outputs, "Synced sequence input and output" - In other word, if you want to apply one or more RNN(s) on this layer, set to False. return_seq_2d : boolean - When return_last = False - If True, return 2D Tensor [n_example, n_hidden], for stacking DenseLayer or computing cost after it. - If False, return 3D Tensor [n_example/n_steps(max), n_steps(max), n_hidden], for stacking multiple RNN after it. name : a string or None An optional name to attach to this layer. Variables ------------ outputs : a tensor The output of this RNN. return_last = False, outputs = all cell_output, which is the hidden state. cell_output.get_shape() = (?, n_hidden) final_state : a tensor or StateTuple When state_is_tuple = False, it is the final hidden and cell states, states.get_shape() = [?, 2 * n_hidden].\n When state_is_tuple = True, it stores two elements: (c, h), in that order. You can get the final state after each iteration during training, then feed it to the initial state of next iteration. initial_state : a tensor or StateTuple It is the initial state of this RNN layer, you can use it to initialize your state at the begining of each epoch or iteration according to your training procedure. sequence_length : a tensor or array, shape = [batch_size] The sequence lengths computed by Advanced Opt or the given sequence lengths. Notes ----- Input dimension should be rank 3 : [batch_size, n_steps(max), n_features], if no, please see :class:`ReshapeLayer`. Examples -------- >>> input_seqs = tf.placeholder(dtype=tf.int64, shape=[batch_size, None], name="input_seqs") >>> network = tl.layers.EmbeddingInputlayer( ... inputs = input_seqs, ... vocabulary_size = vocab_size, ... embedding_size = embedding_size, ... name = 'seq_embedding') >>> network = tl.layers.DynamicRNNLayer(network, ... cell_fn = tf.contrib.rnn.BasicLSTMCell, # for TF0.2 tf.nn.rnn_cell.BasicLSTMCell, ... n_hidden = embedding_size, ... dropout = 0.7, ... sequence_length = tl.layers.retrieve_seq_length_op2(input_seqs), ... return_seq_2d = True, # stack denselayer or compute cost after it ... name = 'dynamic_rnn') ... network = tl.layers.DenseLayer(network, n_units=vocab_size, ... act=tf.identity, name="output") References ---------- - `Wild-ML Blog <http://www.wildml.com/2016/08/rnns-in-tensorflow-a-practical-guide-and-undocumented-features/>`_ - `dynamic_rnn.ipynb <https://github.com/dennybritz/tf-rnn/blob/master/dynamic_rnn.ipynb>`_ - `tf.nn.dynamic_rnn <https://github.com/tensorflow/tensorflow/blob/master/tensorflow/g3doc/api_docs/python/functions_and_classes/shard8/tf.nn.dynamic_rnn.md>`_ - `tflearn rnn <https://github.com/tflearn/tflearn/blob/master/tflearn/layers/recurrent.py>`_ - ``tutorial_dynamic_rnn.py`` """ def __init__( self, layer = None, cell_fn = None,#tf.nn.rnn_cell.LSTMCell, cell_init_args = {'state_is_tuple' : True}, n_hidden = 256, initializer = tf.random_uniform_initializer(-0.1, 0.1), sequence_length = None, initial_state = None, dropout = None, n_layer = 1, return_last = False, return_seq_2d = False, dynamic_rnn_init_args={}, name = 'dyrnn_layer', ): Layer.__init__(self, name=name) if cell_fn is None: raise Exception("Please put in cell_fn") if 'GRU' in cell_fn.__name__: try: cell_init_args.pop('state_is_tuple') except: pass self.inputs = layer.outputs print(" [TL] DynamicRNNLayer %s: n_hidden:%d, in_dim:%d in_shape:%s cell_fn:%s dropout:%s n_layer:%d" % (self.name, n_hidden, self.inputs.get_shape().ndims, self.inputs.get_shape(), cell_fn.__name__, dropout, n_layer)) # Input dimension should be rank 3 [batch_size, n_steps(max), n_features] try: self.inputs.get_shape().with_rank(3) except: raise Exception("RNN : Input dimension should be rank 3 : [batch_size, n_steps(max), n_features]") # Get the batch_size fixed_batch_size = self.inputs.get_shape().with_rank_at_least(1)[0] if fixed_batch_size.value: batch_size = fixed_batch_size.value print(" batch_size (concurrent processes): %d" % batch_size) else: from tensorflow.python.ops import array_ops batch_size = array_ops.shape(self.inputs)[0] print(" non specified batch_size, uses a tensor instead.") self.batch_size = batch_size # Creats the cell function # cell_instance_fn=lambda: cell_fn(num_units=n_hidden, **cell_init_args) # HanSheng rnn_creator = lambda: cell_fn(num_units=n_hidden, **cell_init_args) # Apply dropout if dropout: if type(dropout) in [tuple, list]: in_keep_prob = dropout[0] out_keep_prob = dropout[1] elif isinstance(dropout, float): in_keep_prob, out_keep_prob = dropout, dropout else: raise Exception("Invalid dropout type (must be a 2-D tuple of " "float)") try: # TF1.0 DropoutWrapper_fn = tf.contrib.rnn.DropoutWrapper except: DropoutWrapper_fn = tf.nn.rnn_cell.DropoutWrapper # cell_instance_fn1=cell_instance_fn # HanSheng # cell_instance_fn=DropoutWrapper_fn( # cell_instance_fn1(), # input_keep_prob=in_keep_prob, # output_keep_prob=out_keep_prob) cell_creator = lambda: DropoutWrapper_fn(rnn_creator(), input_keep_prob=in_keep_prob, output_keep_prob=1.0)#out_keep_prob) else: cell_creator = rnn_creator self.cell = cell_creator() # Apply multiple layers if n_layer > 1: try: MultiRNNCell_fn = tf.contrib.rnn.MultiRNNCell except: MultiRNNCell_fn = tf.nn.rnn_cell.MultiRNNCell # cell_instance_fn2=cell_instance_fn # HanSheng try: # cell_instance_fn=lambda: MultiRNNCell_fn([cell_instance_fn2() for _ in range(n_layer)], state_is_tuple=True) # HanSheng self.cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)], state_is_tuple=True) except: # when GRU # cell_instance_fn=lambda: MultiRNNCell_fn([cell_instance_fn2() for _ in range(n_layer)]) # HanSheng self.cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)]) if dropout: self.cell = DropoutWrapper_fn(self.cell, input_keep_prob=1.0, output_keep_prob=out_keep_prob) # self.cell=cell_instance_fn() # HanSheng # Initialize initial_state if initial_state is None: self.initial_state = self.cell.zero_state(batch_size, dtype=tf.float32) else: self.initial_state = initial_state # Computes sequence_length if sequence_length is None: try: ## TF1.0 sequence_length = retrieve_seq_length_op( self.inputs if isinstance(self.inputs, tf.Tensor) else tf.stack(self.inputs)) except: ## TF0.12 sequence_length = retrieve_seq_length_op( self.inputs if isinstance(self.inputs, tf.Tensor) else tf.pack(self.inputs)) # Main - Computes outputs and last_states with tf.variable_scope(name, initializer=initializer) as vs: outputs, last_states = tf.nn.dynamic_rnn( cell=self.cell, # inputs=X inputs = self.inputs, # dtype=tf.float64, sequence_length=sequence_length, initial_state = self.initial_state, **dynamic_rnn_init_args ) rnn_variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) # print(" n_params : %d" % (len(rnn_variables))) # Manage the outputs if return_last: # [batch_size, n_hidden] # outputs = tf.transpose(tf.pack(outputs), [1, 0, 2]) # TF1.0 tf.pack --> tf.stack self.outputs = advanced_indexing_op(outputs, sequence_length) else: # [batch_size, n_step(max), n_hidden] # self.outputs = result[0]["outputs"] # self.outputs = outputs # it is 3d, but it is a list if return_seq_2d: # PTB tutorial: # 2D Tensor [n_example, n_hidden] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs, 1), [-1, n_hidden]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [-1, n_hidden]) else: # <akara>: # 3D Tensor [batch_size, n_steps(max), n_hidden] max_length = tf.shape(outputs)[1] batch_size = tf.shape(outputs)[0] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs, 1), [batch_size, max_length, n_hidden]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [batch_size, max_length, n_hidden]) # self.outputs = tf.reshape(tf.concat(1, outputs), [-1, max_length, n_hidden]) # Final state self.final_state = last_states self.sequence_length = sequence_length self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( rnn_variables )
# Bidirectional Dynamic RNN
[docs]class BiDynamicRNNLayer(Layer): """ The :class:`BiDynamicRNNLayer` class is a RNN layer, you can implement vanilla RNN, LSTM and GRU with it. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. cell_fn : a TensorFlow's core RNN cell as follow (Note TF1.0+ and TF1.0- are different). - see `RNN Cells in TensorFlow <https://www.tensorflow.org/api_docs/python/>`_ cell_init_args : a dictionary The arguments for the cell initializer. n_hidden : a int The number of hidden units in the layer. initializer : initializer The initializer for initializing the parameters. sequence_length : a tensor, array or None The sequence length of each row of input data, see ``Advanced Ops for Dynamic RNN``. - If None, it uses ``retrieve_seq_length_op`` to compute the sequence_length, i.e. when the features of padding (on right hand side) are all zeros. - If using word embedding, you may need to compute the sequence_length from the ID array (the integer features before word embedding) by using ``retrieve_seq_length_op2`` or ``retrieve_seq_length_op``. - You can also input an numpy array. - More details about TensorFlow dynamic_rnn in `Wild-ML Blog <http://www.wildml.com/2016/08/rnns-in-tensorflow-a-practical-guide-and-undocumented-features/>`_. fw_initial_state : None or forward RNN State If None, initial_state is zero_state. bw_initial_state : None or backward RNN State If None, initial_state is zero_state. dropout : `tuple` of `float`: (input_keep_prob, output_keep_prob). The input and output keep probability. n_layer : a int, default is 1. The number of RNN layers. return_last : boolean If True, return the last output, "Sequence input and single output"\n If False, return all outputs, "Synced sequence input and output"\n In other word, if you want to apply one or more RNN(s) on this layer, set to False. return_seq_2d : boolean - When return_last = False - If True, return 2D Tensor [n_example, 2 * n_hidden], for stacking DenseLayer or computing cost after it. - If False, return 3D Tensor [n_example/n_steps(max), n_steps(max), 2 * n_hidden], for stacking multiple RNN after it. name : a string or None An optional name to attach to this layer. Variables ----------------------- outputs : a tensor The output of this RNN. return_last = False, outputs = all cell_output, which is the hidden state. cell_output.get_shape() = (?, 2 * n_hidden) fw(bw)_final_state : a tensor or StateTuple When state_is_tuple = False, it is the final hidden and cell states, states.get_shape() = [?, 2 * n_hidden].\n When state_is_tuple = True, it stores two elements: (c, h), in that order. You can get the final state after each iteration during training, then feed it to the initial state of next iteration. fw(bw)_initial_state : a tensor or StateTuple It is the initial state of this RNN layer, you can use it to initialize your state at the begining of each epoch or iteration according to your training procedure. sequence_length : a tensor or array, shape = [batch_size] The sequence lengths computed by Advanced Opt or the given sequence lengths. Notes ----- Input dimension should be rank 3 : [batch_size, n_steps(max), n_features], if no, please see :class:`ReshapeLayer`. References ---------- - `Wild-ML Blog <http://www.wildml.com/2016/08/rnns-in-tensorflow-a-practical-guide-and-undocumented-features/>`_ - `bidirectional_rnn.ipynb <https://github.com/dennybritz/tf-rnn/blob/master/bidirectional_rnn.ipynb>`_ """ def __init__( self, layer = None, cell_fn = None,#tf.nn.rnn_cell.LSTMCell, cell_init_args = {'state_is_tuple':True}, n_hidden = 256, initializer = tf.random_uniform_initializer(-0.1, 0.1), sequence_length = None, fw_initial_state = None, bw_initial_state = None, dropout = None, n_layer = 1, return_last = False, return_seq_2d = False, dynamic_rnn_init_args={}, name = 'bi_dyrnn_layer', ): Layer.__init__(self, name=name) if cell_fn is None: raise Exception("Please put in cell_fn") if 'GRU' in cell_fn.__name__: try: cell_init_args.pop('state_is_tuple') except: pass self.inputs = layer.outputs print(" [TL] BiDynamicRNNLayer %s: n_hidden:%d in_dim:%d in_shape:%s cell_fn:%s dropout:%s n_layer:%d" % (self.name, n_hidden, self.inputs.get_shape().ndims, self.inputs.get_shape(), cell_fn.__name__, dropout, n_layer)) # Input dimension should be rank 3 [batch_size, n_steps(max), n_features] try: self.inputs.get_shape().with_rank(3) except: raise Exception("RNN : Input dimension should be rank 3 : [batch_size, n_steps(max), n_features]") # Get the batch_size fixed_batch_size = self.inputs.get_shape().with_rank_at_least(1)[0] if fixed_batch_size.value: batch_size = fixed_batch_size.value print(" batch_size (concurrent processes): %d" % batch_size) else: from tensorflow.python.ops import array_ops batch_size = array_ops.shape(self.inputs)[0] print(" non specified batch_size, uses a tensor instead.") self.batch_size = batch_size with tf.variable_scope(name, initializer=initializer) as vs: # Creats the cell function # cell_instance_fn=lambda: cell_fn(num_units=n_hidden, **cell_init_args) # HanSheng rnn_creator = lambda: cell_fn(num_units=n_hidden, **cell_init_args) # Apply dropout if dropout: if type(dropout) in [tuple, list]: in_keep_prob = dropout[0] out_keep_prob = dropout[1] elif isinstance(dropout, float): in_keep_prob, out_keep_prob = dropout, dropout else: raise Exception("Invalid dropout type (must be a 2-D tuple of " "float)") try: DropoutWrapper_fn = tf.contrib.rnn.DropoutWrapper except: DropoutWrapper_fn = tf.nn.rnn_cell.DropoutWrapper # cell_instance_fn1=cell_instance_fn # HanSheng # cell_instance_fn=lambda: DropoutWrapper_fn( # cell_instance_fn1(), # input_keep_prob=in_keep_prob, # output_keep_prob=out_keep_prob) cell_creator = lambda: DropoutWrapper_fn(rnn_creator(), input_keep_prob=in_keep_prob, output_keep_prob=1.0) # out_keep_prob) else: cell_creator = rnn_creator self.fw_cell = cell_creator() self.bw_cell = cell_creator() # Apply multiple layers if n_layer > 1: try: MultiRNNCell_fn = tf.contrib.rnn.MultiRNNCell except: MultiRNNCell_fn = tf.nn.rnn_cell.MultiRNNCell # cell_instance_fn2=cell_instance_fn # HanSheng # cell_instance_fn=lambda: MultiRNNCell_fn([cell_instance_fn2() for _ in range(n_layer)]) self.fw_cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)]) self.bw_cell = MultiRNNCell_fn([cell_creator() for _ in range(n_layer)]) # self.fw_cell=cell_instance_fn() # self.bw_cell=cell_instance_fn() # Initial state of RNN if fw_initial_state is None: self.fw_initial_state = self.fw_cell.zero_state(self.batch_size, dtype=tf.float32) else: self.fw_initial_state = fw_initial_state if bw_initial_state is None: self.bw_initial_state = self.bw_cell.zero_state(self.batch_size, dtype=tf.float32) else: self.bw_initial_state = bw_initial_state # Computes sequence_length if sequence_length is None: try: ## TF1.0 sequence_length = retrieve_seq_length_op( self.inputs if isinstance(self.inputs, tf.Tensor) else tf.stack(self.inputs)) except: ## TF0.12 sequence_length = retrieve_seq_length_op( self.inputs if isinstance(self.inputs, tf.Tensor) else tf.pack(self.inputs)) outputs, (states_fw, states_bw) = tf.nn.bidirectional_dynamic_rnn( cell_fw=self.fw_cell, cell_bw=self.bw_cell, inputs=self.inputs, sequence_length=sequence_length, initial_state_fw=self.fw_initial_state, initial_state_bw=self.bw_initial_state, **dynamic_rnn_init_args ) rnn_variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) print(" n_params : %d" % (len(rnn_variables))) # Manage the outputs try: # TF1.0 outputs = tf.concat(outputs, 2) except: # TF0.12 outputs = tf.concat(2, outputs) if return_last: # [batch_size, 2 * n_hidden] self.outputs = advanced_indexing_op(outputs, sequence_length) else: # [batch_size, n_step(max), 2 * n_hidden] if return_seq_2d: # PTB tutorial: # 2D Tensor [n_example, 2 * n_hidden] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs, 1), [-1, 2 * n_hidden]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [-1, 2 * n_hidden]) else: # <akara>: # 3D Tensor [batch_size, n_steps(max), 2 * n_hidden] max_length = tf.shape(outputs)[1] batch_size = tf.shape(outputs)[0] try: # TF1.0 self.outputs = tf.reshape(tf.concat(outputs, 1), [batch_size, max_length, 2 * n_hidden]) except: # TF0.12 self.outputs = tf.reshape(tf.concat(1, outputs), [batch_size, max_length, 2 * n_hidden]) # self.outputs = tf.reshape(tf.concat(1, outputs), [-1, max_length, 2 * n_hidden]) # Final state self.fw_final_states = states_fw self.bw_final_states = states_bw self.sequence_length = sequence_length self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( rnn_variables )
# Seq2seq
[docs]class Seq2Seq(Layer): """ The :class:`Seq2Seq` class is a simple :class:`DynamicRNNLayer` based Seq2seq layer, both encoder and decoder are :class:`DynamicRNNLayer`, network details see `Model <https://camo.githubusercontent.com/242210d7d0151cae91107ee63bff364a860db5dd/687474703a2f2f6936342e74696e797069632e636f6d2f333031333674652e706e67>`_ and `Sequence to Sequence Learning with Neural Networks <https://arxiv.org/abs/1409.3215>`_ . Parameters ---------- net_encode_in : a :class:`Layer` instance Encode sequences, [batch_size, None, n_features]. net_decode_in : a :class:`Layer` instance Decode sequences, [batch_size, None, n_features]. cell_fn : a TensorFlow's core RNN cell as follow (Note TF1.0+ and TF1.0- are different). - see `RNN Cells in TensorFlow <https://www.tensorflow.org/api_docs/python/>`_ cell_init_args : a dictionary The arguments for the cell initializer. n_hidden : a int The number of hidden units in the layer. initializer : initializer The initializer for initializing the parameters. encode_sequence_length : tensor for encoder sequence length, see :class:`DynamicRNNLayer` . decode_sequence_length : tensor for decoder sequence length, see :class:`DynamicRNNLayer` . initial_state : None or forward RNN State If None, initial_state is of encoder zero_state. dropout : `tuple` of `float`: (input_keep_prob, output_keep_prob). The input and output keep probability. n_layer : a int, default is 1. The number of RNN layers. return_seq_2d : boolean - When return_last = False - If True, return 2D Tensor [n_example, n_hidden], for stacking DenseLayer or computing cost after it. - If False, return 3D Tensor [n_example/n_steps(max), n_steps(max), n_hidden], for stacking multiple RNN after it. name : a string or None An optional name to attach to this layer. Variables ------------ outputs : a tensor The output of RNN decoder. final_state : a tensor or StateTuple Final state of decoder, see :class:`DynamicRNNLayer` . Examples ---------- >>> from tensorlayer.layers import * >>> batch_size = 32 >>> encode_seqs = tf.placeholder(dtype=tf.int64, shape=[batch_size, None], name="encode_seqs") >>> decode_seqs = tf.placeholder(dtype=tf.int64, shape=[batch_size, None], name="decode_seqs") >>> target_seqs = tf.placeholder(dtype=tf.int64, shape=[batch_size, None], name="target_seqs") >>> target_mask = tf.placeholder(dtype=tf.int64, shape=[batch_size, None], name="target_mask") # tl.prepro.sequences_get_mask() >>> with tf.variable_scope("model"): ... # for chatbot, you can use the same embedding layer, ... # for translation, you may want to use 2 seperated embedding layers >>> with tf.variable_scope("embedding") as vs: >>> net_encode = EmbeddingInputlayer( ... inputs = encode_seqs, ... vocabulary_size = 10000, ... embedding_size = 200, ... name = 'seq_embedding') >>> vs.reuse_variables() >>> tl.layers.set_name_reuse(True) >>> net_decode = EmbeddingInputlayer( ... inputs = decode_seqs, ... vocabulary_size = 10000, ... embedding_size = 200, ... name = 'seq_embedding') >>> net = Seq2Seq(net_encode, net_decode, ... cell_fn = tf.contrib.rnn.BasicLSTMCell, ... n_hidden = 200, ... initializer = tf.random_uniform_initializer(-0.1, 0.1), ... encode_sequence_length = retrieve_seq_length_op2(encode_seqs), ... decode_sequence_length = retrieve_seq_length_op2(decode_seqs), ... initial_state = None, ... dropout = None, ... n_layer = 1, ... return_seq_2d = True, ... name = 'seq2seq') >>> net_out = DenseLayer(net, n_units=10000, act=tf.identity, name='output') >>> e_loss = tl.cost.cross_entropy_seq_with_mask(logits=net_out.outputs, target_seqs=target_seqs, input_mask=target_mask, return_details=False, name='cost') >>> y = tf.nn.softmax(net_out.outputs) >>> net_out.print_params(False) Notes -------- - How to feed data: `Sequence to Sequence Learning with Neural Networks <https://arxiv.org/pdf/1409.3215v3.pdf>`_ - input_seqs : ``['how', 'are', 'you', '<PAD_ID'>]`` - decode_seqs : ``['<START_ID>', 'I', 'am', 'fine', '<PAD_ID'>]`` - target_seqs : ``['I', 'am', 'fine', '<END_ID']`` - target_mask : ``[1, 1, 1, 1, 0]`` - related functions : tl.prepro <pad_sequences, precess_sequences, sequences_add_start_id, sequences_get_mask> """ def __init__( self, net_encode_in = None, net_decode_in = None, cell_fn = None,#tf.nn.rnn_cell.LSTMCell, cell_init_args = {'state_is_tuple':True}, n_hidden = 256, initializer = tf.random_uniform_initializer(-0.1, 0.1), encode_sequence_length = None, decode_sequence_length = None, initial_state = None, dropout = None, n_layer = 1, # return_last = False, return_seq_2d = False, name = 'seq2seq', ): Layer.__init__(self, name=name) if cell_fn is None: raise Exception("Please put in cell_fn") if 'GRU' in cell_fn.__name__: try: cell_init_args.pop('state_is_tuple') except: pass # self.inputs = layer.outputs print(" [**] Seq2Seq %s: n_hidden:%d cell_fn:%s dropout:%s n_layer:%d" % (self.name, n_hidden, cell_fn.__name__, dropout, n_layer)) with tf.variable_scope(name) as vs:#, reuse=reuse): # tl.layers.set_name_reuse(reuse) # network = InputLayer(self.inputs, name=name+'/input') network_encode = DynamicRNNLayer(net_encode_in, cell_fn = cell_fn, cell_init_args = cell_init_args, n_hidden = n_hidden, initial_state = initial_state, dropout = dropout, n_layer = n_layer, sequence_length = encode_sequence_length, return_last = False, return_seq_2d = True, name = name+'_encode') # vs.reuse_variables() # tl.layers.set_name_reuse(True) network_decode = DynamicRNNLayer(net_decode_in, cell_fn = cell_fn, cell_init_args = cell_init_args, n_hidden = n_hidden, initial_state = network_encode.final_state, dropout = dropout, n_layer = n_layer, sequence_length = decode_sequence_length, return_last = False, return_seq_2d = return_seq_2d, name = name+'_decode') self.outputs = network_decode.outputs rnn_variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) # Final state self.final_state = network_decode.final_state # self.sequence_length = sequence_length self.all_layers = list(network_decode.all_layers) self.all_params = list(network_decode.all_params) self.all_drop = dict(network_decode.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( rnn_variables ) self.all_layers = list_remove_repeat(self.all_layers) self.all_params = list_remove_repeat(self.all_params)
[docs]class PeekySeq2Seq(Layer): """ Waiting for contribution. The :class:`PeekySeq2Seq` class, see `Model <https://camo.githubusercontent.com/7f690d451036938a51e62feb77149c8bb4be6675/687474703a2f2f6936342e74696e797069632e636f6d2f333032617168692e706e67>`_ and `Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation <https://arxiv.org/abs/1406.1078>`_ . """ def __init__( self, net_encode_in = None, net_decode_in = None, cell_fn = None,#tf.nn.rnn_cell.LSTMCell, cell_init_args = {'state_is_tuple':True}, n_hidden = 256, initializer = tf.random_uniform_initializer(-0.1, 0.1), in_sequence_length = None, out_sequence_length = None, initial_state = None, dropout = None, n_layer = 1, # return_last = False, return_seq_2d = False, name = 'peeky_seq2seq', ): Layer.__init__(self, name=name) if cell_fn is None: raise Exception("Please put in cell_fn") # self.inputs = layer.outputs print(" [TL] PeekySeq2seq %s: n_hidden:%d cell_fn:%s dropout:%s n_layer:%d" % (self.name, n_hidden, cell_fn.__name__, dropout, n_layer))
[docs]class AttentionSeq2Seq(Layer): """ Waiting for contribution. The :class:`AttentionSeq2Seq` class, see `Model <https://camo.githubusercontent.com/0e2e4e5fb2dd47846c2fe027737a5df5e711df1b/687474703a2f2f6936342e74696e797069632e636f6d2f6132727733642e706e67>`_ and `Neural Machine Translation by Jointly Learning to Align and Translate <https://arxiv.org/pdf/1409.0473v6.pdf>`_ . """ def __init__( self, net_encode_in = None, net_decode_in = None, cell_fn = None,#tf.nn.rnn_cell.LSTMCell, cell_init_args = {'state_is_tuple':True}, n_hidden = 256, initializer = tf.random_uniform_initializer(-0.1, 0.1), in_sequence_length = None, out_sequence_length = None, initial_state = None, dropout = None, n_layer = 1, # return_last = False, return_seq_2d = False, name = 'attention_seq2seq', ): Layer.__init__(self, name=name) if cell_fn is None: raise Exception("Please put in cell_fn") # self.inputs = layer.outputs print(" [TL] PeekySeq2seq %s: n_hidden:%d cell_fn:%s dropout:%s n_layer:%d" % (self.name, n_hidden, cell_fn.__name__, dropout, n_layer))
## Shape layer
[docs]class FlattenLayer(Layer): """ The :class:`FlattenLayer` class is layer which reshape high-dimension input to a vector. Then we can apply DenseLayer, RNNLayer, ConcatLayer and etc on the top of it. [batch_size, mask_row, mask_col, n_mask] ---> [batch_size, mask_row * mask_col * n_mask] Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. name : a string or None An optional name to attach to this layer. Examples -------- >>> x = tf.placeholder(tf.float32, shape=[None, 28, 28, 1]) >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = tl.layers.Conv2dLayer(network, ... act = tf.nn.relu, ... shape = [5, 5, 32, 64], ... strides=[1, 1, 1, 1], ... padding='SAME', ... name ='cnn_layer') >>> network = tl.layers.Pool2dLayer(network, ... ksize=[1, 2, 2, 1], ... strides=[1, 2, 2, 1], ... padding='SAME', ... pool = tf.nn.max_pool, ... name ='pool_layer',) >>> network = tl.layers.FlattenLayer(network, name='flatten_layer') """ def __init__( self, layer = None, name ='flatten_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs self.outputs = flatten_reshape(self.inputs, name=name) self.n_units = int(self.outputs.get_shape()[-1]) print(" [TL] FlattenLayer %s: %d" % (self.name, self.n_units)) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
[docs]class ReshapeLayer(Layer): """ The :class:`ReshapeLayer` class is layer which reshape the tensor. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. shape : a list The output shape. name : a string or None An optional name to attach to this layer. Examples -------- - The core of this layer is ``tf.reshape``. - Use TensorFlow only : >>> x = tf.placeholder(tf.float32, shape=[None, 3]) >>> y = tf.reshape(x, shape=[-1, 3, 3]) >>> sess = tf.InteractiveSession() >>> print(sess.run(y, feed_dict={x:[[1,1,1],[2,2,2],[3,3,3],[4,4,4],[5,5,5],[6,6,6]]})) ... [[[ 1. 1. 1.] ... [ 2. 2. 2.] ... [ 3. 3. 3.]] ... [[ 4. 4. 4.] ... [ 5. 5. 5.] ... [ 6. 6. 6.]]] """ def __init__( self, layer = None, shape = [], name ='reshape_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs self.outputs = tf.reshape(self.inputs, shape=shape, name=name) print(" [TL] ReshapeLayer %s: %s" % (self.name, self.outputs.get_shape())) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
[docs]class LambdaLayer(Layer): """ The :class:`LambdaLayer` class is a layer which is able to use the provided function. Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. fn : a function The function that applies to the outputs of previous layer. fn_args : a dictionary The arguments for the function (option). name : a string or None An optional name to attach to this layer. Examples --------- >>> x = tf.placeholder(tf.float32, shape=[None, 1], name='x') >>> network = tl.layers.InputLayer(x, name='input_layer') >>> network = LambdaLayer(network, lambda x: 2*x, name='lambda_layer') >>> y = network.outputs >>> sess = tf.InteractiveSession() >>> out = sess.run(y, feed_dict={x : [[1],[2]]}) ... [[2],[4]] """ def __init__( self, layer = None, fn = None, fn_args = {}, name = 'lambda_layer', ): Layer.__init__(self, name=name) assert layer is not None assert fn is not None self.inputs = layer.outputs print(" [TL] LambdaLayer %s" % self.name) with tf.variable_scope(name) as vs: self.outputs = fn(self.inputs, **fn_args) variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( variables )
## Merge layer
[docs]class ConcatLayer(Layer): """ The :class:`ConcatLayer` class is layer which concat (merge) two or more :class:`DenseLayer` to a single class:`DenseLayer`. Parameters ---------- layer : a list of :class:`Layer` instances The `Layer` class feeding into this layer. concat_dim : int Dimension along which to concatenate. name : a string or None An optional name to attach to this layer. Examples -------- >>> sess = tf.InteractiveSession() >>> x = tf.placeholder(tf.float32, shape=[None, 784]) >>> inputs = tl.layers.InputLayer(x, name='input_layer') >>> net1 = tl.layers.DenseLayer(inputs, n_units=800, act = tf.nn.relu, name='relu1_1') >>> net2 = tl.layers.DenseLayer(inputs, n_units=300, act = tf.nn.relu, name='relu2_1') >>> network = tl.layers.ConcatLayer(layer = [net1, net2], name ='concat_layer') ... [TL] InputLayer input_layer (?, 784) ... [TL] DenseLayer relu1_1: 800, <function relu at 0x1108e41e0> ... [TL] DenseLayer relu2_1: 300, <function relu at 0x1108e41e0> ... [TL] ConcatLayer concat_layer, 1100 ... >>> tl.layers.initialize_global_variables(sess) >>> network.print_params() ... param 0: (784, 800) (mean: 0.000021, median: -0.000020 std: 0.035525) ... param 1: (800,) (mean: 0.000000, median: 0.000000 std: 0.000000) ... param 2: (784, 300) (mean: 0.000000, median: -0.000048 std: 0.042947) ... param 3: (300,) (mean: 0.000000, median: 0.000000 std: 0.000000) ... num of params: 863500 >>> network.print_layers() ... layer 0: Tensor("Relu:0", shape=(?, 800), dtype=float32) ... layer 1: Tensor("Relu_1:0", shape=(?, 300), dtype=float32) ... """ def __init__( self, layer = [], concat_dim = 1, name ='concat_layer', ): Layer.__init__(self, name=name) self.inputs = [] for l in layer: self.inputs.append(l.outputs) try: # TF1.0 self.outputs = tf.concat(self.inputs, concat_dim, name=name) except: # TF0.12 self.outputs = tf.concat(concat_dim, self.inputs, name=name) self.n_units = int(self.outputs.get_shape()[-1]) print(" [TL] ConcatLayer %s: %d" % (self.name, self.n_units)) self.all_layers = list(layer[0].all_layers) self.all_params = list(layer[0].all_params) self.all_drop = dict(layer[0].all_drop) for i in range(1, len(layer)): self.all_layers.extend(list(layer[i].all_layers)) self.all_params.extend(list(layer[i].all_params)) self.all_drop.update(dict(layer[i].all_drop)) self.all_layers = list_remove_repeat(self.all_layers) self.all_params = list_remove_repeat(self.all_params)
#self.all_drop = list_remove_repeat(self.all_drop) # it is a dict
[docs]class ElementwiseLayer(Layer): """ The :class:`ElementwiseLayer` class combines multiple :class:`Layer` which have the same output shapes by a given elemwise-wise operation. Parameters ---------- layer : a list of :class:`Layer` instances The `Layer` class feeding into this layer. combine_fn : a TensorFlow elemwise-merge function e.g. AND is ``tf.minimum`` ; OR is ``tf.maximum`` ; ADD is ``tf.add`` ; MUL is ``tf.multiply`` and so on. See `TensorFlow Math API <https://www.tensorflow.org/versions/master/api_docs/python/math_ops.html#math>`_ . name : a string or None An optional name to attach to this layer. Examples -------- - AND Logic >>> net_0 = tl.layers.DenseLayer(net_0, n_units=500, ... act = tf.nn.relu, name='net_0') >>> net_1 = tl.layers.DenseLayer(net_1, n_units=500, ... act = tf.nn.relu, name='net_1') >>> net_com = tl.layers.ElementwiseLayer(layer = [net_0, net_1], ... combine_fn = tf.minimum, ... name = 'combine_layer') """ def __init__( self, layer = [], combine_fn = tf.minimum, name ='elementwise_layer', ): Layer.__init__(self, name=name) print(" [TL] ElementwiseLayer %s: size:%s fn:%s" % (self.name, layer[0].outputs.get_shape(), combine_fn.__name__)) self.outputs = layer[0].outputs # print(self.outputs._shape, type(self.outputs._shape)) for l in layer[1:]: assert str(self.outputs.get_shape()) == str(l.outputs.get_shape()), "Hint: the input shapes should be the same. %s != %s" % (self.outputs.get_shape() , str(l.outputs.get_shape())) self.outputs = combine_fn(self.outputs, l.outputs, name=name) self.all_layers = list(layer[0].all_layers) self.all_params = list(layer[0].all_params) self.all_drop = dict(layer[0].all_drop) for i in range(1, len(layer)): self.all_layers.extend(list(layer[i].all_layers)) self.all_params.extend(list(layer[i].all_params)) self.all_drop.update(dict(layer[i].all_drop)) self.all_layers = list_remove_repeat(self.all_layers) self.all_params = list_remove_repeat(self.all_params)
# self.all_drop = list_remove_repeat(self.all_drop) # Extend
[docs]class ExpandDimsLayer(Layer): """ The :class:`ExpandDimsLayer` class inserts a dimension of 1 into a tensor's shape, see `tf.expand_dims() <https://www.tensorflow.org/api_docs/python/array_ops/shapes_and_shaping#expand_dims>`_ . Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. axis : int, 0-D (scalar). Specifies the dimension index at which to expand the shape of input. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, axis = None, name = 'expand_dims', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] ExpandDimsLayer %s: axis:%d" % (self.name, axis)) with tf.variable_scope(name) as vs: try: # TF12 TF1.0 self.outputs = tf.expand_dims(self.inputs, axis=axis) except: # TF11 self.outputs = tf.expand_dims(self.inputs, dim=axis) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
# self.all_params.extend( variables )
[docs]class TileLayer(Layer): """ The :class:`TileLayer` class constructs a tensor by tiling a given tensor, see `tf.tile() <https://www.tensorflow.org/api_docs/python/array_ops/slicing_and_joining#tile>`_ . Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. multiples: a list of int Must be one of the following types: int32, int64. 1-D. Length must be the same as the number of dimensions in input name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, multiples = None, name = 'tile', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] TileLayer %s: multiples:%s" % (self.name, multiples)) with tf.variable_scope(name) as vs: self.outputs = tf.tile(self.inputs, multiples=multiples) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] )
# self.all_params.extend( variables ) ## TF-Slim layer
[docs]class SlimNetsLayer(Layer): """ The :class:`SlimNetsLayer` class can be used to merge all TF-Slim nets into TensorLayer. Model can be found in `slim-model <https://github.com/tensorflow/models/tree/master/slim#pre-trained-models>`_ , more about slim see `slim-git <https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/slim>`_ . Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. slim_layer : a slim network function The network you want to stack onto, end with ``return net, end_points``. slim_args : dictionary The arguments for the slim model. name : a string or None An optional name to attach to this layer. Examples -------- - see Inception V3 example on `Github <https://github.com/zsdonghao/tensorlayer>`_ Notes ----- The due to TF-Slim stores the layers as dictionary, the ``all_layers`` in this network is not in order ! Fortunately, the ``all_params`` are in order. """ def __init__( self, layer = None, slim_layer = None, slim_args = {}, name ='tfslim_layer', ): Layer.__init__(self, name=name) assert slim_layer is not None assert slim_args is not None self.inputs = layer.outputs print(" [TL] SlimNetsLayer %s: %s" % (self.name, slim_layer.__name__)) # with tf.variable_scope(name) as vs: # net, end_points = slim_layer(self.inputs, **slim_args) # slim_variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) net, end_points = slim_layer(self.inputs, **slim_args) slim_variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=name) if slim_variables == []: print("No variables found under %s : the name of SlimNetsLayer should be matched with the begining of the ckpt file, see tutorial_inceptionV3_tfslim.py for more details" % name) self.outputs = net slim_layers = [] for v in end_points.values(): # tf.contrib.layers.summaries.summarize_activation(v) slim_layers.append(v) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( slim_layers ) self.all_params.extend( slim_variables )
## Keras layer
[docs]class KerasLayer(Layer): """ The :class:`KerasLayer` class can be used to merge all Keras layers into TensorLayer. Example can be found here `tutorial_keras.py <https://github.com/zsdonghao/tensorlayer/blob/master/example/tutorial_keras.py>`_ Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. keras_layer : a keras network function keras_args : dictionary The arguments for the keras model. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, keras_layer = None, keras_args = {}, name ='keras_layer', ): Layer.__init__(self, name=name) assert layer is not None assert keras_layer is not None self.inputs = layer.outputs print(" [TL] KerasLayer %s: %s" % (self.name, keras_layer)) print(" This API will be removed, please use LambdaLayer instead.") with tf.variable_scope(name) as vs: self.outputs = keras_layer(self.inputs, **keras_args) variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( variables )
## Estimator layer
[docs]class EstimatorLayer(Layer): """ The :class:`EstimatorLayer` class accepts ``model_fn`` that described the model. It is similar with :class:`KerasLayer`, see `tutorial_keras.py <https://github.com/zsdonghao/tensorlayer/blob/master/example/tutorial_keras.py>`_ Parameters ---------- layer : a :class:`Layer` instance The `Layer` class feeding into this layer. model_fn : a function that described the model. args : dictionary The arguments for the model_fn. name : a string or None An optional name to attach to this layer. """ def __init__( self, layer = None, model_fn = None, args = {}, name ='estimator_layer', ): Layer.__init__(self, name=name) assert layer is not None assert model_fn is not None self.inputs = layer.outputs print(" [TL] EstimatorLayer %s: %s" % (self.name, model_fn)) print(" This API will be removed, please use LambdaLayer instead.") with tf.variable_scope(name) as vs: self.outputs = model_fn(self.inputs, **args) variables = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( variables )
## Special activation
[docs]class PReluLayer(Layer): """ The :class:`PReluLayer` class is Parametric Rectified Linear layer. Parameters ---------- x : A `Tensor` with type `float`, `double`, `int32`, `int64`, `uint8`, `int16`, or `int8`. channel_shared : `bool`. Single weight is shared by all channels a_init : alpha initializer, default zero constant. The initializer for initializing the alphas. a_init_args : dictionary The arguments for the weights initializer. name : A name for this activation op (optional). References ----------- - `Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification <http://arxiv.org/pdf/1502.01852v1.pdf>`_ """ def __init__( self, layer = None, channel_shared = False, a_init = tf.constant_initializer(value=0.0), a_init_args = {}, # restore = True, name="prelu_layer" ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] PReluLayer %s: channel_shared:%s" % (self.name, channel_shared)) if channel_shared: w_shape = (1,) else: w_shape = int(self.inputs.get_shape()[-1]) # with tf.name_scope(name) as scope: with tf.variable_scope(name) as vs: alphas = tf.get_variable(name='alphas', shape=w_shape, initializer=a_init, **a_init_args ) try: ## TF 1.0 self.outputs = tf.nn.relu(self.inputs) + tf.multiply(alphas, (self.inputs - tf.abs(self.inputs))) * 0.5 except: ## TF 0.12 self.outputs = tf.nn.relu(self.inputs) + tf.mul(alphas, (self.inputs - tf.abs(self.inputs))) * 0.5 self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( [alphas] )
## Flow control layer
[docs]class MultiplexerLayer(Layer): """ The :class:`MultiplexerLayer` selects one of several input and forwards the selected input into the output, see `tutorial_mnist_multiplexer.py`. Parameters ---------- layer : a list of :class:`Layer` instances The `Layer` class feeding into this layer. name : a string or None An optional name to attach to this layer. Variables ----------------------- sel : a placeholder Input an int [0, inf], which input is the output Examples -------- >>> x = tf.placeholder(tf.float32, shape=[None, 784], name='x') >>> y_ = tf.placeholder(tf.int64, shape=[None, ], name='y_') >>> # define the network >>> net_in = tl.layers.InputLayer(x, name='input_layer') >>> net_in = tl.layers.DropoutLayer(net_in, keep=0.8, name='drop1') >>> # net 0 >>> net_0 = tl.layers.DenseLayer(net_in, n_units=800, ... act = tf.nn.relu, name='net0/relu1') >>> net_0 = tl.layers.DropoutLayer(net_0, keep=0.5, name='net0/drop2') >>> net_0 = tl.layers.DenseLayer(net_0, n_units=800, ... act = tf.nn.relu, name='net0/relu2') >>> # net 1 >>> net_1 = tl.layers.DenseLayer(net_in, n_units=800, ... act = tf.nn.relu, name='net1/relu1') >>> net_1 = tl.layers.DropoutLayer(net_1, keep=0.8, name='net1/drop2') >>> net_1 = tl.layers.DenseLayer(net_1, n_units=800, ... act = tf.nn.relu, name='net1/relu2') >>> net_1 = tl.layers.DropoutLayer(net_1, keep=0.8, name='net1/drop3') >>> net_1 = tl.layers.DenseLayer(net_1, n_units=800, ... act = tf.nn.relu, name='net1/relu3') >>> # multiplexer >>> net_mux = tl.layers.MultiplexerLayer(layer = [net_0, net_1], name='mux_layer') >>> network = tl.layers.ReshapeLayer(net_mux, shape=[-1, 800], name='reshape_layer') # >>> network = tl.layers.DropoutLayer(network, keep=0.5, name='drop3') >>> # output layer >>> network = tl.layers.DenseLayer(network, n_units=10, ... act = tf.identity, name='output_layer') References ------------ - See ``tf.pack() for TF0.12 or tf.stack() for TF1.0`` and ``tf.gather()`` at `TensorFlow - Slicing and Joining <https://www.tensorflow.org/versions/master/api_docs/python/array_ops.html#slicing-and-joining>`_ """ def __init__(self, layer = [], name='mux_layer'): Layer.__init__(self, name=name) self.n_inputs = len(layer) self.inputs = [] for l in layer: self.inputs.append(l.outputs) try: ## TF1.0 all_inputs = tf.stack(self.inputs, name=name) # pack means concat a list of tensor in a new dim # 1.2 except: all_inputs = tf.pack(self.inputs, name=name) # pack means concat a list of tensor in a new dim # 1.2 print(" [TL] MultiplexerLayer %s: n_inputs:%d" % (self.name, self.n_inputs)) self.sel = tf.placeholder(tf.int32) self.outputs = tf.gather(all_inputs, self.sel, name=name) # [sel, :, : ...] # 1.2 # print(self.outputs, vars(self.outputs)) # # tf.reshape(self.outputs, shape=) # exit() # the same with ConcatLayer self.all_layers = list(layer[0].all_layers) self.all_params = list(layer[0].all_params) self.all_drop = dict(layer[0].all_drop) for i in range(1, len(layer)): self.all_layers.extend(list(layer[i].all_layers)) self.all_params.extend(list(layer[i].all_params)) self.all_drop.update(dict(layer[i].all_drop)) self.all_layers = list_remove_repeat(self.all_layers) self.all_params = list_remove_repeat(self.all_params)
# self.all_drop = list_remove_repeat(self.all_drop) ## We can Duplicate the network instead of DemultiplexerLayer # class DemultiplexerLayer(Layer): # """ # The :class:`DemultiplexerLayer` takes a single input and select one of many output lines, which is connected to the input. # # Parameters # ---------- # layer : a list of :class:`Layer` instances # The `Layer` class feeding into this layer. # n_outputs : a int # The number of output # name : a string or None # An optional name to attach to this layer. # # Field (Class Variables) # ----------------------- # sel : a placeholder # Input int [0, inf], the # outputs : a list of Tensor # A list of outputs # # Examples # -------- # >>> # """ # def __init__(self, # layer = None, # name='demux_layer'): # Layer.__init__(self, name=name) # self.outputs = [] ## Wrapper
[docs]class EmbeddingAttentionSeq2seqWrapper(Layer): """Sequence-to-sequence model with attention and for multiple buckets (Deprecated after TF0.12). This example implements a multi-layer recurrent neural network as encoder, and an attention-based decoder. This is the same as the model described in this paper: - `Grammar as a Foreign Language <http://arxiv.org/abs/1412.7449>`_ please look there for details, or into the seq2seq library for complete model implementation. This example also allows to use GRU cells in addition to LSTM cells, and sampled softmax to handle large output vocabulary size. A single-layer version of this model, but with bi-directional encoder, was presented in - `Neural Machine Translation by Jointly Learning to Align and Translate <http://arxiv.org/abs/1409.0473>`_ The sampled softmax is described in Section 3 of the following paper. - `On Using Very Large Target Vocabulary for Neural Machine Translation <http://arxiv.org/abs/1412.2007>`_ Parameters ---------- source_vocab_size : size of the source vocabulary. target_vocab_size : size of the target vocabulary. buckets : a list of pairs (I, O), where I specifies maximum input length that will be processed in that bucket, and O specifies maximum output length. Training instances that have inputs longer than I or outputs longer than O will be pushed to the next bucket and padded accordingly. We assume that the list is sorted, e.g., [(2, 4), (8, 16)]. size : number of units in each layer of the model. num_layers : number of layers in the model. max_gradient_norm : gradients will be clipped to maximally this norm. batch_size : the size of the batches used during training; the model construction is independent of batch_size, so it can be changed after initialization if this is convenient, e.g., for decoding. learning_rate : learning rate to start with. learning_rate_decay_factor : decay learning rate by this much when needed. use_lstm : if true, we use LSTM cells instead of GRU cells. num_samples : number of samples for sampled softmax. forward_only : if set, we do not construct the backward pass in the model. name : a string or None An optional name to attach to this layer. """ def __init__(self, source_vocab_size, target_vocab_size, buckets, size, num_layers, max_gradient_norm, batch_size, learning_rate, learning_rate_decay_factor, use_lstm=False, num_samples=512, forward_only=False, name='wrapper'): Layer.__init__(self)#, name=name) self.source_vocab_size = source_vocab_size self.target_vocab_size = target_vocab_size self.buckets = buckets self.batch_size = batch_size self.learning_rate = tf.Variable(float(learning_rate), trainable=False, name='learning_rate') self.learning_rate_decay_op = self.learning_rate.assign( self.learning_rate * learning_rate_decay_factor) self.global_step = tf.Variable(0, trainable=False, name='global_step') if tf.__version__ >= "0.12": raise Exception("Deprecated after TF0.12 : use other seq2seq layers instead.") # =========== Fake output Layer for compute cost ====== # If we use sampled softmax, we need an output projection. with tf.variable_scope(name) as vs: output_projection = None softmax_loss_function = None # Sampled softmax only makes sense if we sample less than vocabulary size. if num_samples > 0 and num_samples < self.target_vocab_size: w = tf.get_variable("proj_w", [size, self.target_vocab_size]) w_t = tf.transpose(w) b = tf.get_variable("proj_b", [self.target_vocab_size]) output_projection = (w, b) def sampled_loss(inputs, labels): labels = tf.reshape(labels, [-1, 1]) return tf.nn.sampled_softmax_loss(w_t, b, inputs, labels, num_samples, self.target_vocab_size) softmax_loss_function = sampled_loss # ============ Seq Encode Layer ============= # Create the internal multi-layer cell for our RNN. try: # TF1.0 cell_creator = lambda: tf.contrib.rnn.GRUCell(size) except: cell_creator = lambda: tf.nn.rnn_cell.GRUCell(size) if use_lstm: try: # TF1.0 cell_creator = lambda: tf.contrib.rnn.BasicLSTMCell(size) except: cell_creator = lambda: tf.nn.rnn_cell.BasicLSTMCell(size) cell = cell_creator() if num_layers > 1: try: # TF1.0 cell = tf.contrib.rnn.MultiRNNCell([single_cell] * num_layers) except: cell = tf.nn.rnn_cell.MultiRNNCell([single_cell] * num_layers) # ============== Seq Decode Layer ============ # The seq2seq function: we use embedding for the input and attention. def seq2seq_f(encoder_inputs, decoder_inputs, do_decode): return tf.nn.seq2seq.embedding_attention_seq2seq( encoder_inputs, decoder_inputs, cell, num_encoder_symbols=source_vocab_size, num_decoder_symbols=target_vocab_size, embedding_size=size, output_projection=output_projection, feed_previous=do_decode) #============================================================= # Feeds for inputs. self.encoder_inputs = [] self.decoder_inputs = [] self.target_weights = [] for i in xrange(buckets[-1][0]): # Last bucket is the biggest one. self.encoder_inputs.append(tf.placeholder(tf.int32, shape=[None], name="encoder{0}".format(i))) for i in xrange(buckets[-1][1] + 1): self.decoder_inputs.append(tf.placeholder(tf.int32, shape=[None], name="decoder{0}".format(i))) self.target_weights.append(tf.placeholder(tf.float32, shape=[None], name="weight{0}".format(i))) # Our targets are decoder inputs shifted by one. targets = [self.decoder_inputs[i + 1] for i in xrange(len(self.decoder_inputs) - 1)] self.targets = targets # DH add for debug # Training outputs and losses. if forward_only: self.outputs, self.losses = tf.nn.seq2seq.model_with_buckets( self.encoder_inputs, self.decoder_inputs, targets, self.target_weights, buckets, lambda x, y: seq2seq_f(x, y, True), softmax_loss_function=softmax_loss_function) # If we use output projection, we need to project outputs for decoding. if output_projection is not None: for b in xrange(len(buckets)): self.outputs[b] = [ tf.matmul(output, output_projection[0]) + output_projection[1] for output in self.outputs[b] ] else: self.outputs, self.losses = tf.nn.seq2seq.model_with_buckets( self.encoder_inputs, self.decoder_inputs, targets, self.target_weights, buckets, lambda x, y: seq2seq_f(x, y, False), softmax_loss_function=softmax_loss_function) # Gradients and SGD update operation for training the model. params = tf.trainable_variables() if not forward_only: self.gradient_norms = [] self.updates = [] opt = tf.train.GradientDescentOptimizer(self.learning_rate) for b in xrange(len(buckets)): gradients = tf.gradients(self.losses[b], params) clipped_gradients, norm = tf.clip_by_global_norm(gradients, max_gradient_norm) self.gradient_norms.append(norm) self.updates.append(opt.apply_gradients( zip(clipped_gradients, params), global_step=self.global_step)) # if save into npz self.all_params = tf.get_collection(TF_GRAPHKEYS_VARIABLES, scope=vs.name) # if save into ckpt self.saver = tf.train.Saver(tf.all_variables())
[docs] def step(self, session, encoder_inputs, decoder_inputs, target_weights, bucket_id, forward_only): """Run a step of the model feeding the given inputs. Parameters ---------- session : tensorflow session to use. encoder_inputs : list of numpy int vectors to feed as encoder inputs. decoder_inputs : list of numpy int vectors to feed as decoder inputs. target_weights : list of numpy float vectors to feed as target weights. bucket_id : which bucket of the model to use. forward_only : whether to do the backward step or only forward. Returns -------- A triple consisting of gradient norm (or None if we did not do backward), average perplexity, and the outputs. Raises -------- ValueError : if length of encoder_inputs, decoder_inputs, or target_weights disagrees with bucket size for the specified bucket_id. """ # Check if the sizes match. encoder_size, decoder_size = self.buckets[bucket_id] if len(encoder_inputs) != encoder_size: raise ValueError("Encoder length must be equal to the one in bucket," " %d != %d." % (len(encoder_inputs), encoder_size)) if len(decoder_inputs) != decoder_size: raise ValueError("Decoder length must be equal to the one in bucket," " %d != %d." % (len(decoder_inputs), decoder_size)) if len(target_weights) != decoder_size: raise ValueError("Weights length must be equal to the one in bucket," " %d != %d." % (len(target_weights), decoder_size)) # print('in model.step()') # print('a',bucket_id, encoder_size, decoder_size) # Input feed: encoder inputs, decoder inputs, target_weights, as provided. input_feed = {} for l in xrange(encoder_size): input_feed[self.encoder_inputs[l].name] = encoder_inputs[l] for l in xrange(decoder_size): input_feed[self.decoder_inputs[l].name] = decoder_inputs[l] input_feed[self.target_weights[l].name] = target_weights[l] # print(self.encoder_inputs[l].name) # print(self.decoder_inputs[l].name) # print(self.target_weights[l].name) # Since our targets are decoder inputs shifted by one, we need one more. last_target = self.decoder_inputs[decoder_size].name input_feed[last_target] = np.zeros([self.batch_size], dtype=np.int32) # print('last_target', last_target) # Output feed: depends on whether we do a backward step or not. if not forward_only: output_feed = [self.updates[bucket_id], # Update Op that does SGD. self.gradient_norms[bucket_id], # Gradient norm. self.losses[bucket_id]] # Loss for this batch. else: output_feed = [self.losses[bucket_id]] # Loss for this batch. for l in xrange(decoder_size): # Output logits. output_feed.append(self.outputs[bucket_id][l]) outputs = session.run(output_feed, input_feed) if not forward_only: return outputs[1], outputs[2], None # Gradient norm, loss, no outputs. else: return None, outputs[0], outputs[1:] # No gradient norm, loss, outputs.
[docs] def get_batch(self, data, bucket_id, PAD_ID=0, GO_ID=1, EOS_ID=2, UNK_ID=3): """ Get a random batch of data from the specified bucket, prepare for step. To feed data in step(..) it must be a list of batch-major vectors, while data here contains single length-major cases. So the main logic of this function is to re-index data cases to be in the proper format for feeding. Parameters ---------- data : a tuple of size len(self.buckets) in which each element contains lists of pairs of input and output data that we use to create a batch. bucket_id : integer, which bucket to get the batch for. PAD_ID : int Index of Padding in vocabulary GO_ID : int Index of GO in vocabulary EOS_ID : int Index of End of sentence in vocabulary UNK_ID : int Index of Unknown word in vocabulary Returns ------- The triple (encoder_inputs, decoder_inputs, target_weights) for the constructed batch that has the proper format to call step(...) later. """ encoder_size, decoder_size = self.buckets[bucket_id] encoder_inputs, decoder_inputs = [], [] # Get a random batch of encoder and decoder inputs from data, # pad them if needed, reverse encoder inputs and add GO to decoder. for _ in xrange(self.batch_size): encoder_input, decoder_input = random.choice(data[bucket_id]) # Encoder inputs are padded and then reversed. encoder_pad = [PAD_ID] * (encoder_size - len(encoder_input)) encoder_inputs.append(list(reversed(encoder_input + encoder_pad))) # Decoder inputs get an extra "GO" symbol, and are padded then. decoder_pad_size = decoder_size - len(decoder_input) - 1 decoder_inputs.append([GO_ID] + decoder_input + [PAD_ID] * decoder_pad_size) # Now we create batch-major vectors from the data selected above. batch_encoder_inputs, batch_decoder_inputs, batch_weights = [], [], [] # Batch encoder inputs are just re-indexed encoder_inputs. for length_idx in xrange(encoder_size): batch_encoder_inputs.append( np.array([encoder_inputs[batch_idx][length_idx] for batch_idx in xrange(self.batch_size)], dtype=np.int32)) # Batch decoder inputs are re-indexed decoder_inputs, we create weights. for length_idx in xrange(decoder_size): batch_decoder_inputs.append( np.array([decoder_inputs[batch_idx][length_idx] for batch_idx in xrange(self.batch_size)], dtype=np.int32)) # Create target_weights to be 0 for targets that are padding. batch_weight = np.ones(self.batch_size, dtype=np.float32) for batch_idx in xrange(self.batch_size): # We set weight to 0 if the corresponding target is a PAD symbol. # The corresponding target is decoder_input shifted by 1 forward. if length_idx < decoder_size - 1: target = decoder_inputs[batch_idx][length_idx + 1] if length_idx == decoder_size - 1 or target == PAD_ID: batch_weight[batch_idx] = 0.0 batch_weights.append(batch_weight) return batch_encoder_inputs, batch_decoder_inputs, batch_weights
## Developing or Untested class MaxoutLayer(Layer): """ Waiting for contribution Single DenseLayer with Max-out behaviour, work well with Dropout. References ----------- `Goodfellow (2013) Maxout Networks <http://arxiv.org/abs/1302.4389>`_ """ def __init__( self, layer = None, n_units = 100, name ='maxout_layer', ): Layer.__init__(self, name=name) self.inputs = layer.outputs print(" [TL] MaxoutLayer %s: %d" % (self.name, self.n_units)) print(" Waiting for contribution") with tf.variable_scope(name) as vs: pass # W = tf.Variable(init.xavier_init(n_inputs=n_in, n_outputs=n_units, uniform=True), name='W') # b = tf.Variable(tf.zeros([n_units]), name='b') # self.outputs = act(tf.matmul(self.inputs, W) + b) # https://www.tensorflow.org/versions/r0.9/api_docs/python/array_ops.html#pack # http://stackoverflow.com/questions/34362193/how-to-explicitly-broadcast-a-tensor-to-match-anothers-shape-in-tensorflow # tf.concat tf.pack tf.tile self.all_layers = list(layer.all_layers) self.all_params = list(layer.all_params) self.all_drop = dict(layer.all_drop) self.all_layers.extend( [self.outputs] ) self.all_params.extend( [W, b] ) #