def train(self,total_loss, lr, global_step):
# all of them are tensor
#total_sample = 274 yike: ok to comment out?
#num_batches_per_epoch = 274/1 yike: ok to comment out?
loss_averages_op = _add_loss_summaries(total_loss)
# Compute gradients.
with tf.control_dependencies([loss_averages_op]):
#print('try...')
opt = tf.train.AdamOptimizer(lr)
print('toto_loss_shape: '+str(total_loss))
opt.compute_gradients(total_loss)
grads = opt.compute_gradients(total_loss)
#print(grads)
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)
# Add histograms for gradients.
for grad, var in grads:
if grad is not None:
tf.summary.histogram(var.op.name + '/gradients', grad)
# Track the moving averages of all trainable variables.
variable_averages = tf.train.ExponentialMovingAverage(Model.MOVING_AVERAGE_DECAY, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
with tf.control_dependencies([apply_gradient_op, variables_averages_op]):
train_op = tf.no_op(name='train')
return train_op
def train_batch(total_loss, global_step, optimizer, learning_rate, moving_average_decay,
update_gradient_vars):
loss_averages_op = utils._add_loss_summaries(total_loss)
#计算梯度
with tf.control_dependencies([loss_averages_op]):
if optimizer == 'ADAGRAD':
opt = tf.train.AdagradOptimizer(learning_rate)
elif optimizer == 'ADADELTA':
opt = tf.train.AdadeltaOptimizer(learning_rate,rho=0.9,epsilon=1e-6)
elif optimizer == 'ADAM':
opt = tf.train.AdamOptimizer(learning_rate,beta1=0.9,beta2=0.999,epsilon=0.1)
elif optimizer == 'RMSPROP':
opt = tf.train.RMSPropOptimizer(learning_rate,decay=0.9,momentum=0.9,epsilon=1.0)
elif optimizer == 'MOM':
opt = tf.train.MomentumOptimizer(learning_rate,0.9,use_nesterov=True)
else:
raise ValueError('Invalid optimization algorithm')
grads = opt.compute_gradients(total_loss,update_gradient_vars)
apply_gradient_op = opt.apply_gradients(grads,global_step=global_step)
variable_averages = tf.train.ExponentialMovingAverage(moving_average_decay,global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
with tf.control_dependencies([apply_gradient_op,variables_averages_op]):
train_op = tf.no_op(name='train')
return train_op
def captcha_train(total_loss, global_step):
"""
Train captcha model.
Create an optimizer and apply to all trainable variables. Add moving
average for all trainable variables.
Args:
total_loss: Total loss from loss().
global_step: Integer Variable counting the number of training steps
processed.
Returns:
train_op: op for training.
"""
# Variables that affect learning rate.
num_batches_per_epoch = NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / FLAGS.batch_size
decay_steps = int(num_batches_per_epoch * NUM_EPOCHS_PER_DECAY)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(
INITIAL_LEARNING_RATE,
global_step,
decay_steps,
LEARNING_RATE_DECAY_FACTOR,
staircase=True
)
tf.scalar_summary('learning_rate', lr)
# Generate moving averages of all losses and associated summaries.
loss_averages_op = _add_loss_summaries(total_loss)
# Compute gradients.
with tf.control_dependencies([loss_averages_op]):
opt = tf.train.AdamOptimizer(lr)
grads = opt.compute_gradients(total_loss)
# Apply gradients.
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
tf.histogram_summary(var.op.name, var)
# Add histograms for gradients.
for grad, var in grads:
if grad:
tf.histogram_summary(var.op.name + '/gradients', grad)
# Track the moving averages of all trainable variables.
variable_averages = tf.train.ExponentialMovingAverage(
MOVING_AVERAGE_DECAY, global_step)
variables_to_average = list(set(tf.trainable_variables() + filter(lambda v: "_mean" in v.name or "_variance" in v.name, tf.all_variables())))
variables_averages_op = variable_averages.apply(variables_to_average)
with tf.control_dependencies([apply_gradient_op, variables_averages_op]):
train_op = tf.no_op(name='train')
return train_op
def captcha_train(total_loss, global_step):
"""
Train captcha model.
Create an optimizer and apply to all trainable variables. Add moving
average for all trainable variables.
Args:
total_loss: Total loss from loss().
global_step: Integer Variable counting the number of training steps
processed.
Returns:
train_op: op for training.
"""
# Variables that affect learning rate.
num_batches_per_epoch = NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN / FLAGS.batch_size
decay_steps = int(num_batches_per_epoch * NUM_EPOCHS_PER_DECAY)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(INITIAL_LEARNING_RATE,
global_step,
decay_steps,
LEARNING_RATE_DECAY_FACTOR,
staircase=True)
tf.scalar_summary('learning_rate', lr)
# Generate moving averages of all losses and associated summaries.
loss_averages_op = _add_loss_summaries(total_loss)
# Compute gradients.
with tf.control_dependencies([loss_averages_op]):
opt = tf.train.AdamOptimizer(lr)
grads = opt.compute_gradients(total_loss)
# Apply gradients.
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
tf.histogram_summary(var.op.name, var)
# Add histograms for gradients.
for grad, var in grads:
if grad:
tf.histogram_summary(var.op.name + '/gradients', grad)
# Track the moving averages of all trainable variables.
variable_averages = tf.train.ExponentialMovingAverage(
MOVING_AVERAGE_DECAY, global_step)
variables_to_average = list(
set(tf.trainable_variables() +
filter(lambda v: "_mean" in v.name or "_variance" in v.name,
tf.all_variables())))
variables_averages_op = variable_averages.apply(variables_to_average)
with tf.control_dependencies([apply_gradient_op, variables_averages_op]):
train_op = tf.no_op(name='train')
return train_op
def __init__(
self,
q_values,
im_shape, num_channels,
batch_size,
actions, train_dir,
optimizer,
decay_steps, initial_learning_rate, learning_rate_decay_factor=.1,
checkpoint_path=None):
"""
checkpoint_path: path to model checkpoint file
logits is a function taking a tensor -> float array
logits must take a single tensor argument
ind_to_label is object such that ind_to_label[ind] = label
"""
self.im_shape = im_shape
self.num_channels = num_channels
self.batch_size = batch_size
self.train_dir = train_dir
self.g = tf.Graph()
with self.g.as_default():
self.sess = tf.Session()
# variables to represent images, q values, expected q_values, and action inds
self.images_placeholder = tf.placeholder("float", shape=[None, im_shape[0], im_shape[1], num_channels])
self.images = self.images_placeholder #preprocess_images(self.images_placeholder)
# TODO APPLY THIS NORMALIZATION ACROSS THE BATCH
self.q_values = q_values(self.images, num_channels, len(actions))
# TODO FIGURE OUT HOW TO REUSE THE VARIABLES BETWEEN THESE TWO OUTPUTS
self.target_values = tf.placeholder("float", shape=[None, len(actions)])
self.target_inds = tf.placeholder("float", shape=[None, len(actions)])
# generate the loss and training up
self.q_learning_loss = losses.q_learning_loss(self.q_values, self.target_values, self.target_inds, "q_loss")
tf.add_to_collection('losses', self.q_learning_loss)
self.loss = tf.add_n(tf.get_collection('losses'), name='total_loss')
loss_averages_op = _add_loss_summaries(self.loss)
self.global_step = tf.Variable(0, trainable=False)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(
INITIAL_LEARNING_RATE,
self.global_step,
decay_steps,
learning_rate_decay_factor,
staircase=True
)
tf.scalar_summary('learning_rate', lr)
# Compute gradients.
with tf.control_dependencies([loss_averages_op]):
opt = optimizer(lr) # tf.train.AdamOptimizer(lr)
grads = opt.compute_gradients(self.loss)
# Apply gradients.
self.apply_gradient_op = opt.apply_gradients(grads, global_step=self.global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
tf.histogram_summary(var.op.name, var)
# Add histograms for gradients.
for grad, var in grads:
if grad:
tf.histogram_summary(var.op.name + '/gradients', grad)
with tf.control_dependencies([self.apply_gradient_op]):
self.train_op = tf.no_op(name='train')
# Build the summary operation based on the TF collection of Summaries.
self.summary_op = tf.merge_all_summaries()
self.saver = tf.train.Saver(tf.all_variables())
self.actions = actions
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
self.sess.run(init)
self.summary_writer = tf.train.SummaryWriter(train_dir,
graph_def=self.sess.graph_def)
if checkpoint_path is not None:
# restore the model's parameters
self.checkpoint_step = checkpoint_path.split('-')[-1]
self.checkpoint_path = checkpoint_path
self.saver.restore(self.sess, checkpoint_path)
def __init__(self,
q_values,
im_shape,
num_channels,
batch_size,
actions,
train_dir,
optimizer,
decay_steps,
initial_learning_rate,
learning_rate_decay_factor=.1,
checkpoint_path=None):
"""
checkpoint_path: path to model checkpoint file
logits is a function taking a tensor -> float array
logits must take a single tensor argument
ind_to_label is object such that ind_to_label[ind] = label
"""
self.im_shape = im_shape
self.num_channels = num_channels
self.batch_size = batch_size
self.train_dir = train_dir
self.g = tf.Graph()
with self.g.as_default():
self.sess = tf.Session()
# variables to represent images, q values, expected q_values, and action inds
self.images_placeholder = tf.placeholder(
"float", shape=[None, im_shape[0], im_shape[1], num_channels])
self.images = self.images_placeholder #preprocess_images(self.images_placeholder)
# TODO APPLY THIS NORMALIZATION ACROSS THE BATCH
self.q_values = q_values(self.images, num_channels, len(actions))
# TODO FIGURE OUT HOW TO REUSE THE VARIABLES BETWEEN THESE TWO OUTPUTS
self.target_values = tf.placeholder("float",
shape=[None,
len(actions)])
self.target_inds = tf.placeholder("float",
shape=[None, len(actions)])
# generate the loss and training up
self.q_learning_loss = losses.q_learning_loss(
self.q_values, self.target_values, self.target_inds, "q_loss")
tf.add_to_collection('losses', self.q_learning_loss)
self.loss = tf.add_n(tf.get_collection('losses'),
name='total_loss')
loss_averages_op = _add_loss_summaries(self.loss)
self.global_step = tf.Variable(0, trainable=False)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(INITIAL_LEARNING_RATE,
self.global_step,
decay_steps,
learning_rate_decay_factor,
staircase=True)
tf.scalar_summary('learning_rate', lr)
# Compute gradients.
with tf.control_dependencies([loss_averages_op]):
opt = optimizer(lr) # tf.train.AdamOptimizer(lr)
grads = opt.compute_gradients(self.loss)
# Apply gradients.
self.apply_gradient_op = opt.apply_gradients(
grads, global_step=self.global_step)
# Add histograms for trainable variables.
for var in tf.trainable_variables():
tf.histogram_summary(var.op.name, var)
# Add histograms for gradients.
for grad, var in grads:
if grad:
tf.histogram_summary(var.op.name + '/gradients', grad)
with tf.control_dependencies([self.apply_gradient_op]):
self.train_op = tf.no_op(name='train')
# Build the summary operation based on the TF collection of Summaries.
self.summary_op = tf.merge_all_summaries()
self.saver = tf.train.Saver(tf.all_variables())
self.actions = actions
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
self.sess.run(init)
self.summary_writer = tf.train.SummaryWriter(
train_dir, graph_def=self.sess.graph_def)
if checkpoint_path is not None:
# restore the model's parameters
self.checkpoint_step = checkpoint_path.split('-')[-1]
self.checkpoint_path = checkpoint_path
self.saver.restore(self.sess, checkpoint_path)