def compute_loss(self, targets, logits, logit_seq_length,
target_seq_length):
'''
Compute the loss
Creates the operation to compute the cross-enthropy loss for every input
frame (if you want to have a different loss function, overwrite this
method)
Args:
targets: a list that contains a Bx1 tensor containing the targets
for eacht time step where B is the batch size
logits: a list that contains a BxO tensor containing the output
logits for eacht time step where O is the output dimension
logit_seq_length: the length of all the input sequences as a vector
target_seq_length: the length of all the target sequences as a
vector
Returns:
a scalar value containing the loss
'''
with tf.variable_scope('weight_loss'):
trainable_weights = tf.trainable_variables()
weight_loss = 0
for trainable in trainable_weights:
weight_loss += tf.nn.l2_loss(trainable)
weight_loss = weight_loss / len(trainable_weights)
with tf.name_scope('cross_enthropy_loss'):
#training starts at t=1.
targets_t_one = targets[:, 1:, :]
target_seq_length_t_one = target_seq_length - 1
#convert to non sequential data
nonseq_targets = seq_convertors.seq2nonseq(
targets_t_one, target_seq_length_t_one)
nonseq_logits = seq_convertors.seq2nonseq(logits, logit_seq_length)
#make a vector out of the targets
nonseq_targets = tf.reshape(nonseq_targets, [-1])
#one hot encode the targets
#pylint: disable=E1101
nonseq_targets = tf.one_hot(nonseq_targets,
int(nonseq_logits.get_shape()[1]))
#compute the cross-enthropy loss
loss = tf.reduce_sum(
tf.nn.softmax_cross_entropy_with_logits(
nonseq_logits, nonseq_targets))
loss = loss + self.l2_cost_weight * weight_loss
return loss
def compute_loss(self, targets, logits, logit_seq_length,
target_seq_length):
'''
Compute the loss
Creates the operation to compute the cross-enthropy loss for every input
frame (if you want to have a different loss function, overwrite this
method)
Args:
targets: a list that contains a Bx1 tensor containing the targets
for eacht time step where B is the batch size
logits: a list that contains a BxO tensor containing the output
logits for eacht time step where O is the output dimension
logit_seq_length: the length of all the input sequences as a vector
target_seq_length: the length of all the target sequences as a
vector
Returns:
a scalar value containing the loss
'''
with tf.name_scope('cross_enthropy_loss'):
#convert to non sequential data
nonseq_targets = seq_convertors.seq2nonseq(targets,
target_seq_length)
nonseq_logits = seq_convertors.seq2nonseq(logits, logit_seq_length)
#make a vector out of the targets
nonseq_targets = tf.reshape(nonseq_targets, [-1])
#one hot encode the targets
#pylint: disable=E1101
nonseq_targets = tf.one_hot(nonseq_targets,
int(nonseq_logits.get_shape()[1]))
# Evaluate model
# argmax取的是最大值的下标
correct_pred = tf.equal(tf.argmax(nonseq_logits, 1),
tf.argmax(nonseq_targets, 1))
true_count = tf.reduce_sum(tf.cast(correct_pred, tf.float32))
# loss
loss = tf.reduce_sum(
tf.nn.softmax_cross_entropy_with_logits(logits=nonseq_logits,
labels=nonseq_targets))
#compute the cross-enthropy loss
return loss, true_count
def compute_loss(self, targets, logits, logit_seq_length,
target_seq_length):
'''
Compute the loss
Creates the operation to compute the CTC loss for every input
frame (if you want to have a different loss function, overwrite this
method)
Args:
targets: a list that contains a Bx1 tensor containing the targets
for eacht time step where B is the batch size
logits: a list that contains a BxO tensor containing the output
logits for eacht time step where O is the output dimension
logit_seq_length: the length of all the input sequences as a vector
target_seq_length: the length of all the target sequences as a
vector
Returns:
a scalar value containing the loss
'''
#get the batch size
batch_size = int(target_seq_length.get_shape()[0])
#convert the targets into a sparse tensor representation
indices = tf.concat(0, [tf.concat(1, [tf.tile([s], target_seq_length[s])
, tf.range(target_seq_length[s])])
for s in range(len(batch_size))])
values = tf.reshape(seq_convertors.seq2nonseq(logits, logit_seq_length),
[-1])
shape = [batch_size, len(targets)]
sparse_targets = tf.SparseTensor(indices, values, shape)
tf.nn.ctc_loss(tf.pack(logits), sparse_targets, logit_seq_length)
def compute_loss(self, targets, logits, logit_seq_length,
target_seq_length):
'''
Compute the loss
Creates the operation to compute the cross-enthropy loss for every input
frame (if you want to have a different loss function, overwrite this
method)
Args:
targets: a list that contains a Bx1 tensor containing the targets
for eacht time step where B is the batch size
logits: a list that contains a BxO tensor containing the output
logits for eacht time step where O is the output dimension
logit_seq_length: the length of all the input sequences as a vector
target_seq_length: the length of all the target sequences as a
vector
Returns:
a scalar value containing the loss
'''
with tf.name_scope('cross_enthropy_loss'):
#convert to non sequential data
nonseq_targets = seq_convertors.seq2nonseq(targets,
target_seq_length)
nonseq_logits = seq_convertors.seq2nonseq(logits, logit_seq_length)
#make a vector out of the targets
nonseq_targets = tf.reshape(nonseq_targets, [-1])
#one hot encode the targets
#pylint: disable=E1101
nonseq_targets = tf.one_hot(nonseq_targets,
int(nonseq_logits.get_shape()[1]))
#compute the cross-enthropy loss
return tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(
nonseq_logits, nonseq_targets))
def compute_loss(self, targets, logits, logit_seq_length,
target_seq_length):
'''
Compute the loss
Creates the operation to compute the CTC loss for every input
frame (if you want to have a different loss function, overwrite this
method)
Args:
targets: a [batch_size, max_target_length, 1] tensor containing the
targets
logits: a [batch_size, max_input_length, dim] tensor containing the
inputs
logit_seq_length: the length of all the input sequences as a vector
target_seq_length: the length of all the target sequences as a
vector
Returns:
a scalar value containing the loss
'''
with tf.name_scope('CTC_loss'):
#get the batch size
batch_size = int(targets.get_shape()[0])
#convert the targets into a sparse tensor representation
indices = tf.concat([
tf.concat([
tf.expand_dims(tf.tile([s], [target_seq_length[s]]), 1),
tf.expand_dims(tf.range(target_seq_length[s]), 1)
], 1) for s in range(batch_size)
], 0)
values = tf.reshape(
seq_convertors.seq2nonseq(targets, target_seq_length), [-1])
shape = [batch_size, int(targets.get_shape()[1])]
sparse_targets = tf.SparseTensor(tf.cast(indices, tf.int64),
values, shape)
loss = tf.reduce_sum(
tf.nn.ctc_loss(sparse_targets,
logits,
logit_seq_length,
time_major=False))
return loss
def get_outputs(self, logits, logits_seq_length):
'''
Put the classifier output logits through a softmax
Args:
logits: A list containing a 1xO tensor for each timestep where O
is the classifier output dimension
logits_seq_length: the logits sequence length
Returns:
An NxO tensor containing posterior distributions
'''
#convert logits to non sequence for the softmax computation
logits = seq_convertors.seq2nonseq(logits, logits_seq_length)
return tf.nn.softmax(logits)
def __init__(self, classifier, input_dim, max_length):
'''
NnetDecoder constructor, creates the decoding graph
Args:
classifier: the classifier that will be used for decoding
input_dim: the input dimension to the nnnetgraph
'''
self.graph = tf.Graph()
self.max_length = max_length
with self.graph.as_default():
#create the inputs placeholder
self.inputs = tf.placeholder(tf.float32,
shape=[max_length, input_dim],
name='inputs')
#create the sequence length placeholder
self.seq_length = tf.placeholder(tf.int32,
shape=[1],
name='seq_length')
split_inputs = tf.unstack(tf.expand_dims(self.inputs, 1))
#create the decoding graph
logits, _, self.saver, _ = classifier(split_inputs,
self.seq_length,
is_training=False,
reuse=False,
scope='Classifier')
#convert logits to non sequence for the softmax computation
logits = seq_convertors.seq2nonseq(logits, self.seq_length)
#compute the outputs
self.outputs = tf.nn.softmax(logits)
# merge all summary during the decoding
self.merged = tf.summary.merge_all()
self.summarywriter = tf.summary.FileWriter(
logdir="tf-exp/decode_vis", graph=self.graph)
self.decode_visualisation = False
#specify that the graph can no longer be modified after this point
self.graph.finalize()
def get_outputs(self, logits, logits_seq_length):
'''
Put the classifier output logits through a softmax
Args:
logits: A list containing a 1xO tensor for each timestep where O
is the classifier output dimension
logits_seq_length: the logits sequence length
Returns:
An NxO tensor containing posterior distributions
'''
# convert logits to non sequence for the softmax computation
logits = seq_convertors.seq2nonseq(logits, logits_seq_length)
'''
softmax(logits, dim=-1, name=None):
Computes softmax activations.
For each batch `i` and class `j` we have
softmax = exp(logits) / reduce_sum(exp(logits), dim)
'''
return tf.nn.softmax(logits)
def compute_loss(self, targets, logits, logit_seq_length,
target_seq_length):
'''
Compute the loss
Creates the operation to compute the CTC loss for every input
frame (if you want to have a different loss function, overwrite this
method)
Args:
targets: a list that contains a Bx1 tensor containing the targets
for eacht time step where B is the batch size
logits: a list that contains a BxO tensor containing the output
logits for eacht time step where O is the output dimension
logit_seq_length: the length of all the input sequences as a vector
target_seq_length: the length of all the target sequences as a
vector
Returns:
a scalar value containing the loss
'''
#get the batch size
batch_size = int(target_seq_length.get_shape()[0])
#convert the targets into a sparse tensor representation
indices = tf.concat(0, [
tf.concat(1, [
tf.tile([s], target_seq_length[s]),
tf.range(target_seq_length[s])
]) for s in range(len(batch_size))
])
values = tf.reshape(
seq_convertors.seq2nonseq(logits, logit_seq_length), [-1])
shape = [batch_size, len(targets)]
sparse_targets = tf.SparseTensor(indices, values, shape)
tf.nn.ctc_loss(tf.pack(logits), sparse_targets, logit_seq_length)
def __init__(self, classifier, input_dim, max_length):
'''
NnetDecoder constructor, creates the decoding graph
Args:
classifier: the classifier that will be used for decoding
input_dim: the input dimension to the nnnetgraph
'''
self.graph = tf.Graph()
self.max_length = max_length
with self.graph.as_default():
#create the inputs placeholder
self.inputs = tf.placeholder(
tf.float32, shape=[max_length, input_dim], name='inputs')
#create the sequence length placeholder
self.seq_length = tf.placeholder(
tf.int32, shape=[1], name='seq_length')
split_inputs = tf.unpack(tf.expand_dims(self.inputs, 1))
#create the decoding graph
logits, _, self.saver, _ = classifier(split_inputs, self.seq_length,
is_training=False, reuse=False,
scope='Classifier')
#convert logits to non sequence for the softmax computation
logits = seq_convertors.seq2nonseq(logits, self.seq_length)
#compute the outputs
self.outputs = tf.nn.softmax(logits)
#specify that the graph can no longer be modified after this point
self.graph.finalize()
