def train(loss, globalStep, **kwargs):
if kwargs.get('optimizer') == 'MomentumOptimizer':
optimizerParams = optimizer_params.get_momentum_optimizer_params(
globalStep, **kwargs)
if kwargs.get('optimizer') == 'AdamOptimizer':
optimizerParams = optimizer_params.get_adam_optimizer_params(
globalStep, **kwargs)
if kwargs.get('optimizer') == 'GradientDescentOptimizer':
optimizerParams = optimizer_params.get_gradient_descent_optimizer_params(
globalStep, **kwargs)
if kwargs.get('optimizer') == 'AdaGrad':
optimizerParams = optimizer_params.get_adagrad_optimizer_params(
globalStep, **kwargs)
# Generate moving averages of all losses and associated summaries.
#lossAveragesOp = loss_base.add_loss_summaries(loss, kwargs.get('activeBatchSize', None))
tf.add_to_collection('losses', loss)
losses = tf.get_collection('losses')
for l in losses:
tf.summary.scalar('loss_' + l.op.name, l)
# Compute gradients.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) # for batchnorm
#tvars = tf.trainable_variables()
#with tf.control_dependencies(update_ops):
# if kwargs.get('optimizer') == 'AdamOptimizer':
# optim = tf.train.AdamOptimizer(learning_rate=optimizerParams['learningRate'], epsilon=optimizerParams['epsilon'])
# if kwargs.get('optimizer') == 'MomentumOptimizer':
# optim = tf.train.MomentumOptimizer(learning_rate=optimizerParams['learningRate'], momentum=optimizerParams['momentum'])
# if kwargs.get('optimizer') == 'GradientDescentOptimizer':
# optim = tf.train.GradientDescentOptimizer(learning_rate=optimizerParams['learningRate'])
# if kwargs.get('optimizer') == 'AdaGrad':
# optim = tf.train.AdamOptimizer(learning_rate=optimizerParams['learningRate'])
learningRate = optimizer_params._get_learning_rate_piecewise_shifted(
globalStep, **kwargs)
optim = tf.train.GradientDescentOptimizer(learningRate)
print(loss)
opTrain = tf.contrib.slim.learning.create_train_op(loss, optimizer=optim)
# grads, norm = tf.clip_by_global_norm(tf.gradients(loss, tvars), kwargs.get('clipNorm'))
#
# # Apply gradients.
# #applyGradientOp = opt.apply_gradients(grads, global_step=globalStep)
# #train_op = opt.apply_gradients(gradsNvars, global_step=globalStep)
# opApplyGradients = optim.apply_gradients(zip(grads, tvars), global_step=globalStep)
# # 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 zip(grads, tvars):
# if grad is not None:
# tf.summary.histogram(var.op.name + '/gradients', grad)
#
# with tf.control_dependencies([opApplyGradients]):
# opTrain = tf.no_op(name='train')
return opTrain
def train(loss, globalStep, **kwargs):
if kwargs.get('optimizer') == 'MomentumOptimizer':
optimizerParams = optimizer_params.get_momentum_optimizer_params(
globalStep, **kwargs)
if kwargs.get('optimizer') == 'AdamOptimizer':
optimizerParams = optimizer_params.get_adam_optimizer_params(
globalStep, **kwargs)
if kwargs.get('optimizer') == 'GradientDescentOptimizer':
optimizerParams = optimizer_params.get_gradient_descent_optimizer_params(
globalStep, **kwargs)
# Generate moving averages of all losses and associated summaries.
lossAveragesOp = loss_base.add_loss_summaries(
loss, kwargs.get('activeBatchSize', None))
# Compute gradients.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) # for batchnorm
tvars = tf.trainable_variables()
with tf.control_dependencies(update_ops):
if kwargs.get('optimizer') == 'AdamOptimizer':
optim = tf.train.AdamOptimizer(
learning_rate=optimizerParams['learningRate'],
epsilon=optimizerParams['epsilon'])
if kwargs.get('optimizer') == 'MomentumOptimizer':
optim = tf.train.MomentumOptimizer(
learning_rate=optimizerParams['learningRate'],
momentum=optimizerParams['momentum'])
if kwargs.get('optimizer') == 'GradientDescentOptimizer':
optim = tf.train.GradientDescentOptimizer(
learning_rate=optimizerParams['learningRate'])
if kwargs.get('optimizer') == 'AdaGrad':
optim = tf.train.AdamOptimizer(
learning_rate=optimizerParams['learningRate'])
grads, norm = tf.clip_by_global_norm(tf.gradients(loss, tvars),
kwargs.get('clipNorm'))
# Apply gradients.
#applyGradientOp = opt.apply_gradients(grads, global_step=globalStep)
#train_op = opt.apply_gradients(gradsNvars, global_step=globalStep)
opApplyGradients = optim.apply_gradients(zip(grads, tvars),
global_step=globalStep)
# 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 zip(grads, tvars):
if grad is not None:
tf.summary.histogram(var.op.name + '/gradients', grad)
with tf.control_dependencies([opApplyGradients]):
opTrain = tf.no_op(name='train')
return opTrain
