def test(loss, globalStep, **kwargs):
# Generate moving averages of all losses and associated summaries.
lossAveragesOp = loss_base.add_loss_summaries(loss, kwargs.get('activeBatchSize', None))
with tf.control_dependencies([]):
opTest = tf.no_op(name='test')
return opTest
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