def _get_testing(rnn_logits,sequence_length,label,label_length):
"""Create ops for testing (all scalars):
loss: CTC loss function value,
label_error: Batch-normalized edit distance on beam search max
sequence_error: Batch-normalized sequence error rate
"""
with tf.name_scope("train"):
loss = model.ctc_loss_layer(rnn_logits,label,sequence_length)
with tf.name_scope("test"):
predictions, probability = tf.nn.ctc_beam_search_decoder(rnn_logits, # вытаскивание log_probabilities из ctc
sequence_length,
beam_width=128,
top_paths=1,
merge_repeated=False) # если True, то на выходе модели не будет повторяющихся символов
#predictions, probability = tf.nn.ctc_greedy_decoder(rnn_logits, # альтернативный loss (?)
# sequence_length,
# merge_repeated=True)
hypothesis = tf.cast(predictions[0], tf.int32) # for edit_distance
label_errors = tf.edit_distance(hypothesis, label, normalize=False) # расстояние Левенштейна
sequence_errors = tf.count_nonzero(label_errors, axis=0) # подсчет ненулевых значнией, то есть случаев, когда выход модели не совпадает с gt
total_label_error = tf.reduce_sum(label_errors) # рассчет суммы расстояний Левенштейна по батчу
total_labels = tf.reduce_sum(label_length) # количество символов в gt для всего батча
label_error = tf.truediv(total_label_error, tf.cast(total_labels, tf.float32), name='label_error') # нормированное расстояние Левенштейна (деленное на количество символов)
sequence_error = tf.truediv(tf.cast(sequence_errors, tf.int32), tf.shape(label_length)[0], name='sequence_error') # доля неправильных ответов
tf.summary.scalar( 'loss', loss )
tf.summary.scalar( 'label_error', label_error )
tf.summary.scalar( 'sequence_error', sequence_error )
return loss, label_error, sequence_error, predictions[0], probability
def _get_testing(rnn_logits, sequence_length, label, label_length):
"""Create ops for testing (all scalars):
loss: CTC loss function value,
label_error: Batch-normalized edit distance on beam search max
sequence_error: Batch-normalized sequence error rate
"""
with tf.name_scope("train"):
loss = model.ctc_loss_layer(rnn_logits, label, sequence_length)
with tf.name_scope("test"):
predictions, _ = tf.nn.ctc_beam_search_decoder(rnn_logits,
sequence_length,
beam_width=128,
top_paths=1,
merge_repeated=True)
hypothesis = tf.cast(predictions[0], tf.int32) # for edit_distance
label_errors = tf.edit_distance(hypothesis, label, normalize=False)
sequence_errors = tf.count_nonzero(label_errors, axis=0)
total_label_error = tf.reduce_sum(label_errors)
total_labels = tf.reduce_sum(label_length)
label_error = tf.truediv(total_label_error,
tf.cast(total_labels, tf.float32),
name='label_error')
sequence_error = tf.truediv(tf.cast(sequence_errors, tf.int32),
tf.shape(label_length)[0],
name='sequence_error')
tf.summary.scalar('loss', loss)
tf.summary.scalar('label_error', label_error)
tf.summary.scalar('sequence_error', sequence_error)
return loss, label_error, sequence_error
def _get_testing( rnn_logits,sequence_length,label,label_length,
continuous_eval, lexicon, lexicon_prior ):
"""Create ops for testing (all scalars):
loss: CTC loss function value,
label_error: batch level edit distance on beam search max
sequence_error: batch level sequence error rate
"""
with tf.name_scope( "train" ):
# Reduce by mean (rather than sum) if doing continuous evaluation
batch_loss = model.ctc_loss_layer( rnn_logits,label,sequence_length,
reduce_mean=continuous_eval)
with tf.name_scope( "test" ):
predictions,_ = _get_output( rnn_logits, sequence_length,
lexicon, lexicon_prior )
hypothesis = tf.cast( predictions[0], tf.int32 ) # for edit_distance
# Per-sequence statistic
num_label_errors = tf.edit_distance( hypothesis, label,
normalize=False )
# Per-batch summary counts
batch_num_label_errors = tf.reduce_sum( num_label_errors)
batch_num_sequence_errors = tf.count_nonzero( num_label_errors, axis=0 )
batch_num_labels = tf.reduce_sum( label_length )
# Wide integer type casts (prefer unsigned, but truediv dislikes those)
batch_num_label_errors = tf.cast( batch_num_label_errors, tf.int64 )
batch_num_sequence_errors = tf.cast( batch_num_sequence_errors,
tf.int64 )
batch_num_labels = tf.cast( batch_num_labels, tf.int64)
return batch_loss, batch_num_label_errors, batch_num_sequence_errors, \
batch_num_labels, predictions
def _get_testing(rnn_logits, sequence_length, label, label_length):
"""Create ops for testing (all scalars):
loss: CTC loss function value,
label_error: batch level edit distance on beam search max
sequence_error: batch level sequence error rate
"""
with tf.name_scope("train"):
loss = model.ctc_loss_layer(rnn_logits, label, sequence_length)
with tf.name_scope("test"):
predictions, _ = tf.nn.ctc_beam_search_decoder(rnn_logits,
sequence_length,
beam_width=128,
top_paths=1,
merge_repeated=True)
hypothesis = tf.cast(predictions[0], tf.int32) # for edit_distance
# Per-sequence statistic
num_label_errors = tf.edit_distance(hypothesis, label, normalize=False)
# Per-batch summary counts
batch_num_label_errors = tf.reduce_sum(num_label_errors)
batch_num_sequence_errors = tf.count_nonzero(num_label_errors, axis=0)
batch_num_labels = tf.reduce_sum(label_length)
# Wide integer type casts (prefer unsigned, but truediv dislikes those)
batch_num_label_errors = tf.cast(batch_num_label_errors, tf.int64)
batch_num_sequence_errors = tf.cast(batch_num_sequence_errors,
tf.int64)
batch_num_labels = tf.cast(batch_num_labels, tf.int64)
return loss, batch_num_label_errors, batch_num_sequence_errors, \
batch_num_labels, predictions
def _get_training(rnn_logits, label, sequence_length, global_steps):
"""Set up training ops"""
with tf.name_scope("train"):
rnn_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES
) #tf.get_collection：从一个集合中取出全部变量，是一个列
loss = model.ctc_loss_layer(rnn_logits, label, sequence_length)
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops): #控制依赖
learning_rate = tf.train.exponential_decay(
FLAGS.learning_rate,
global_steps,
FLAGS.decay_steps,
FLAGS.decay_rate,
staircase=FLAGS.decay_staircase,
name='learning_rate')
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
beta1=FLAGS.momentum)
train_op = tf.contrib.layers.optimize_loss(
loss=loss,
global_step=global_steps,
learning_rate=learning_rate,
optimizer=optimizer,
variables=rnn_vars)
tf.summary.scalar('learning_rate', learning_rate)
return train_op
def _get_testing(rnn_logits,sequence_length,label,label_length):
"""Create ops for testing (all scalars):
loss: CTC loss function value,
label_error: Batch-normalized edit distance on beam search max
sequence_error: Batch-normalized sequence error rate
"""
with tf.name_scope("train"):
loss = model.ctc_loss_layer(rnn_logits,label,sequence_length)
with tf.name_scope("test"):
predictions,_ = tf.nn.ctc_beam_search_decoder(rnn_logits,
sequence_length,
beam_width=128,
top_paths=1,
merge_repeated=True)
hypothesis = tf.cast(predictions[0], tf.int32) # for edit_distance
label_errors = tf.edit_distance(hypothesis, label, normalize=False)
sequence_errors = tf.count_nonzero(label_errors,axis=0)
total_label_error = tf.reduce_sum( label_errors )
total_labels = tf.reduce_sum( label_length )
label_error = tf.truediv( total_label_error,
tf.cast(total_labels, tf.float32 ),
name='label_error')
sequence_error = tf.truediv( tf.cast( sequence_errors, tf.int32 ),
tf.shape(label_length)[0],
name='sequence_error')
tf.summary.scalar( 'loss', loss )
tf.summary.scalar( 'label_error', label_error )
tf.summary.scalar( 'sequence_error', sequence_error )
return loss, label_error, sequence_error
def _get_training(rnn_logits, label, sequence_length, label_length):
# Set up training ops
with tf.name_scope('train'):
if FLAGS.tune_scope:
scope = FLAGS.tune_scope
else:
scope = 'convnet|rnn'
rnn_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=scope)
loss = model.ctc_loss_layer(rnn_logits, label, sequence_length)
# Update batch norm stats [http://stackoverflow.com/questions/43234667]
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
learning_rate = tf.train.exponential_decay(
FLAGS.learning_rate,
tf.train.get_global_step(),
FLAGS.decay_steps,
FLAGS.decay_rate,
staircase=FLAGS.decay_staircase,
name='learning_rate')
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
beta1=FLAGS.momentum)
train_op = tf.contrib.layers.optimize_loss(
loss=loss,
global_step=tf.train.get_global_step(),
learning_rate=learning_rate,
optimizer=optimizer,
variables=rnn_vars)
with tf.name_scope('test'):
predictions, probability = tf.nn.ctc_beam_search_decoder(
rnn_logits, # вытаскивание log_probabilities из ctc
sequence_length,
beam_width=128,
top_paths=1,
merge_repeated=False
) # если True, то на выходе модели не будет повторяющихся символов
hypothesis = tf.cast(predictions[0], tf.int32) # for edit_distance
label_errors = tf.edit_distance(
hypothesis, label, normalize=False) # расстояние Левенштейна
sequence_errors = tf.count_nonzero(
label_errors, axis=0
) # подсчет ненулевых значнией, то есть случаев, когда выход модели не совпадает с gt
total_label_error = tf.reduce_sum(
label_errors) # рассчет суммы расстояний Левенштейна по батчу
total_labels = tf.reduce_sum(
label_length) # количество символов в gt для всего батча
label_error = tf.truediv(
total_label_error,
tf.cast(total_labels, tf.float32),
name='label_error'
) # нормированное расстояние Левенштейна (деленное на количество символов)
sequence_error = tf.truediv(
tf.cast(sequence_errors, tf.int32),
tf.shape(label_length)[0],
name='sequence_error') # доля неправильных ответов
return train_op, label_error, sequence_error, predictions[0], probability
def _get_training(rnn_logits, label, sequence_length, tune_scope,
learning_rate, decay_steps, decay_rate, decay_staircase,
momentum):
"""Set up training ops"""
with tf.compat.v1.name_scope("train"):
if tune_scope:
scope = tune_scope
else:
scope = "convnet|rnn"
rnn_vars = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope=scope)
loss = model.ctc_loss_layer(rnn_logits, label, sequence_length)
# Update batch norm stats [http://stackoverflow.com/questions/43234667]
extra_update_ops = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
# Calculate the learning rate given the parameters
learning_rate_tensor = tf.compat.v1.train.exponential_decay(
learning_rate,
tf.compat.v1.train.get_global_step(),
decay_steps,
decay_rate,
staircase=decay_staircase,
name='learning_rate')
optimizer = tf.compat.v1.train.AdamOptimizer(
learning_rate=learning_rate_tensor, beta1=momentum)
#train_op = optimizer.minimize(loss = loss, global_step = tf.compat.v1.train.get_global_step(), var_list=rnn_vars)
# train_op = optimizer.minimize(
# loss = loss,
# global_step = tf.compat.v1.train.get_global_step(),
# #learning_rate=learning_rate_tensor,
# var_list=rnn_vars
# )
train_op = layers.optimize_loss(
loss=loss,
global_step=tf.compat.v1.train.get_global_step(),
learning_rate=learning_rate_tensor,
optimizer=optimizer,
variables=rnn_vars)
tf.compat.v1.summary.scalar('learning_rate', learning_rate_tensor)
return train_op, loss
def _get_training(rnn_logits, label, sequence_length):
"""Set up training ops"""
"""
接受RNN层预测的logits，与真实标签label，结合序列长度，进行训练
loss是CTC损失
优化的是CNN和RNN（bi-LSTM）网络的参数
"""
with tf.name_scope("train"):
if FLAGS.tune_scope:
scope = FLAGS.tune_scope
else:
scope = "convnet|rnn"
# scope="convnet|rnn"，即参数包含CNN与RNN两个网络
rnn_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=scope)
loss = model.ctc_loss_layer(rnn_logits, label, sequence_length)
# Update batch norm stats [http://stackoverflow.com/questions/43234667]
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
learning_rate = tf.train.exponential_decay(
FLAGS.learning_rate,
tf.train.get_global_step(),
FLAGS.decay_steps,
FLAGS.decay_rate,
staircase=FLAGS.decay_staircase,
name='learning_rate')
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
beta1=FLAGS.momentum)
train_op = tf.contrib.layers.optimize_loss(
loss=loss,
global_step=tf.train.get_global_step(),
learning_rate=learning_rate,
optimizer=optimizer,
variables=rnn_vars)
tf.summary.scalar('learning_rate', learning_rate)
return train_op
def _get_training(rnn_logits,label,sequence_length):
"""Set up training ops"""
with tf.name_scope("train"):
if FLAGS.tune_scope:
scope=FLAGS.tune_scope
else:
scope="convnet|rnn"
rnn_vars = tf.get_collection( tf.GraphKeys.TRAINABLE_VARIABLES,
scope=scope)
loss = model.ctc_loss_layer(rnn_logits,label,sequence_length)
# Update batch norm stats [http://stackoverflow.com/questions/43234667]
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
learning_rate = tf.train.exponential_decay(
FLAGS.learning_rate,
tf.train.get_global_step(),
FLAGS.decay_steps,
FLAGS.decay_rate,
staircase=FLAGS.decay_staircase,
name='learning_rate')
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate,
beta1=FLAGS.momentum)
#Given loss and parameters for optimizer,returns a training op.
#If not set "summaries",the loss,the learing rate,and the global norm of the gradients will be reported.
train_op = tf.contrib.layers.optimize_loss(
loss=loss,
global_step=tf.train.get_global_step(),
learning_rate=learning_rate,
optimizer=optimizer,
variables=rnn_vars)
tf.summary.scalar( 'learning_rate', learning_rate )
return train_op
def _get_training(rnn_logits,label,sequence_length):
"""Set up training ops"""
with tf.name_scope("train"):
if FLAGS.tune_scope:
scope=FLAGS.tune_scope
else:
scope="convnet|rnn"
rnn_vars = tf.get_collection( tf.GraphKeys.TRAINABLE_VARIABLES,
scope=scope)
loss = model.ctc_loss_layer(rnn_logits,label,sequence_length)
# Update batch norm stats [http://stackoverflow.com/questions/43234667]
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
learning_rate = tf.train.exponential_decay(
FLAGS.learning_rate,
tf.train.get_global_step(),
FLAGS.decay_steps,
FLAGS.decay_rate,
staircase=FLAGS.decay_staircase,
name='learning_rate')
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate,
beta1=FLAGS.momentum)
train_op = tf.contrib.layers.optimize_loss(
loss=loss,
global_step=tf.train.get_global_step(),
learning_rate=learning_rate,
optimizer=optimizer,
variables=rnn_vars)
tf.summary.scalar( 'learning_rate', learning_rate )
return train_op
def _get_training(args, rnn_logits, label, sequence_length):
"""Set up training ops"""
with tf.name_scope("train"):
scope = "convnet|rnn"
rnn_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=scope)
loss = model.ctc_loss_layer(rnn_logits, label, sequence_length)
# Update batch norm stats [http://stackoverflow.com/questions/43234667]
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
learning_rate = tf.train.exponential_decay(
args.learning_rate,
tf.train.get_global_step(),
args.decay_steps,
args.decay_rate,
staircase=args.decay_staircase,
name='learning_rate')
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
beta1=args.momentum)
train_op = tf.contrib.layers.optimize_loss(
loss=loss,
global_step=tf.train.get_global_step(),
learning_rate=learning_rate,
optimizer=optimizer,
variables=rnn_vars)
return train_op, loss
