def resnet_arg_scope(is_training=True,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'is_training': False,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'trainable': False,
'updates_collections': tf.GraphKeys.UPDATE_OPS
}
with arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
weights_initializer=slim.variance_scaling_initializer(),
trainable=is_training,
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with arg_scope([slim.batch_norm], **batch_norm_params) as arg_sc:
return arg_sc
python类variance_scaling_initializer()的实例源码
def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
activation_fn=None,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
activation_fn=None,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS_EXTRA,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
def resnet_arg_scope(is_training=True,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'is_training': False,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'trainable': False,
'updates_collections': tf.GraphKeys.UPDATE_OPS
}
with arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
weights_initializer=slim.variance_scaling_initializer(),
trainable=is_training,
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with arg_scope([slim.batch_norm], **batch_norm_params) as arg_sc:
return arg_sc
def resnet_arg_scope(is_training=True,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'is_training': False,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'trainable': False,
'updates_collections': tf.GraphKeys.UPDATE_OPS
}
with arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(cfg.TRAIN.WEIGHT_DECAY),
weights_initializer=slim.variance_scaling_initializer(),
trainable=is_training,
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with arg_scope([slim.batch_norm], **batch_norm_params) as arg_sc:
return arg_sc
def resnet_arg_scope(is_training=True):
"""Sets up the default arguments for the CIFAR-10 resnet model."""
batch_norm_params = {
'is_training': is_training,
'decay': 0.9,
'epsilon': 0.001,
'scale': True,
# This forces batch_norm to compute the moving averages in-place
# instead of using a global collection which does not work with tf.cond.
# 'updates_collections': None,
}
with slim.arg_scope([slim.conv2d, slim.batch_norm], activation_fn=lrelu):
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(0.0002),
weights_initializer=slim.variance_scaling_initializer(),
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params) as arg_sc:
return arg_sc
def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
activation_fn=None,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
def _create_baseline(self, n_output=1, n_hidden=100,
is_zero_init=False,
collection='BASELINE'):
# center input
h = self._x
if self.mean_xs is not None:
h -= self.mean_xs
if is_zero_init:
initializer = init_ops.zeros_initializer()
else:
initializer = slim.variance_scaling_initializer()
with slim.arg_scope([slim.fully_connected],
variables_collections=[collection, Q_COLLECTION],
trainable=False,
weights_initializer=initializer):
h = slim.fully_connected(h, n_hidden, activation_fn=tf.nn.tanh)
baseline = slim.fully_connected(h, n_output, activation_fn=None)
if n_output == 1:
baseline = tf.reshape(baseline, [-1]) # very important to reshape
return baseline
def _extra_conv_arg_scope_with_bn(weight_decay=0.00001,
activation_fn=None,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS_EXTRA,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
def inference(images, num_classes, is_traiing=True, scope='inception_v3'):
"""Build Inception v3 model architecture.
See here for reference: http://arxiv.org/abs/1512.00567
Args:
images: Images returned from inputs() or distorted_inputs().
num_classes: number of classes
for_training: If set to `True`, build the inference model for training.
Kernels that operate differently for inference during training
e.g. dropout, are appropriately configured.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: optional prefix string identifying the ImageNet tower.
Returns:
Logits. 2-D float Tensor.
Auxiliary Logits. 2-D float Tensor of side-head. Used for training only.
"""
# Parameters for BatchNorm.
batch_norm_params = {
# Decay for the moving averages.
'decay': BATCHNORM_MOVING_AVERAGE_DECAY,
# epsilon to prevent 0s in variance.
'epsilon': 0.001,
# calculate moving average or using exist one
'is_training': is_traiing
}
# Set weight_decay for weights in Conv and FC layers.
with slim.arg_scope([slim.conv2d, slim.fully_connected],
weights_regularizer=slim.l2_regularizer(FLAGS.weight_decay)):
with slim.arg_scope([slim.conv2d],
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
logits, endpoints = inception_v3(
images,
num_classes=num_classes,
dropout_keep_prob=0.8,
is_training=is_traiing,
scope=scope
)
# Add summaries for viewing model statistics on TensorBoard.
_activation_summaries(endpoints)
# Grab the logits associated with the side head. Employed during training.
auxiliary_logits = endpoints['aux_logits']
return logits, auxiliary_logits
def inference(images, num_classes, is_training=True, scope='squeeze'):
"""
Args:
images: Images returned from inputs() or distorted_inputs().
num_classes: number of classes
for_training: If set to `True`, build the inference model for training.
Kernels that operate differently for inference during training
e.g. dropout, are appropriately configured.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: optional prefix string identifying the ImageNet tower.
Returns:
Logits. 2-D float Tensor.
Auxiliary Logits. 2-D float Tensor of side-head. Used for training only.
"""
# Parameters for BatchNorm.
batch_norm_params = {
# Decay for the moving averages.
'decay': BATCHNORM_MOVING_AVERAGE_DECAY,
# epsilon to prevent 0s in variance.
'epsilon': 0.001,
# calculate moving average or using exist one
'is_training': is_training
}
# Set weight_decay for weights in Conv and FC layers.
with slim.arg_scope([slim.conv2d, slim.fully_connected],
weights_regularizer=slim.l2_regularizer(FLAGS.weight_decay)):
with slim.arg_scope([slim.conv2d],
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
logits, endpoints = squeezenet(
images,
num_classes=num_classes,
keep_prob=0.5,
is_training=is_training,
scope=scope
)
# Add summaries for viewing model statistics on TensorBoard.
_activation_summaries(endpoints)
return logits
def inference(images, num_classes, is_training=True,scope='densenet_121'):
"""
Args:
images: Images returned from inputs() or distorted_inputs().
num_classes: number of classes
for_training: If set to `True`, build the inference model for training.
Kernels that operate differently for inference during training
e.g. dropout, are appropriately configured.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: optional prefix string identifying the ImageNet tower.
Returns:
Logits. 2-D float Tensor.
Auxiliary Logits. 2-D float Tensor of side-head. Used for training only.
"""
# Parameters for BatchNorm.
batch_norm_params = {
# Decay for the moving averages.
'decay': BATCHNORM_MOVING_AVERAGE_DECAY,
# epsilon to prevent 0s in variance.
'epsilon': 0.001,
# calculate moving average or using exist one
'is_training': is_training
}
# Set weight_decay for weights in Conv and FC layers.
with slim.arg_scope([slim.conv2d, slim.fully_connected],
weights_regularizer=slim.l2_regularizer(FLAGS.weight_decay)):
with slim.arg_scope([slim.conv2d],
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
logits, endpoints = densenet_a(
images,
num_classes=num_classes,
keep_prob=0.2,
is_training=is_training,
scope=scope
)
# Add summaries for viewing model statistics on TensorBoard.
_activation_summaries(endpoints)
return logits
def inference(images, num_classes, is_training=True, scope='vgg_16'):
"""Build Inception v3 model architecture.
See here for reference: http://arxiv.org/abs/1512.00567
Args:
images: Images returned from inputs() or distorted_inputs().
num_classes: number of classes
for_training: If set to `True`, build the inference model for training.
Kernels that operate differently for inference during training
e.g. dropout, are appropriately configured.
restore_logits: whether or not the logits layers should be restored.
Useful for fine-tuning a model with different num_classes.
scope: optional prefix string identifying the ImageNet tower.
Returns:
Logits. 2-D float Tensor.
Auxiliary Logits. 2-D float Tensor of side-head. Used for training only.
"""
# Parameters for BatchNorm.
batch_norm_params = {
# Decay for the moving averages.
'decay': BATCHNORM_MOVING_AVERAGE_DECAY,
# epsilon to prevent 0s in variance.
'epsilon': 0.001,
# calculate moving average or using exist one
'is_training': is_training
}
# Set weight_decay for weights in Conv and FC layers.
with slim.arg_scope([slim.conv2d, slim.fully_connected],
weights_regularizer=slim.l2_regularizer(FLAGS.weight_decay)):
with slim.arg_scope([slim.conv2d],
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
logits, endpoints = vgg_16(
images,
num_classes=num_classes,
dropout_keep_prob=0.8,
is_training=is_training,
scope=scope
)
# Add summaries for viewing model statistics on TensorBoard.
_activation_summaries(endpoints)
return logits
def resnet_arg_scope(weight_decay=0.0001,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
"""Defines the default ResNet arg scope.
TODO(gpapan): The batch-normalization related default values above are
appropriate for use in conjunction with the reference ResNet models
released at https://github.com/KaimingHe/deep-residual-networks. When
training ResNets from scratch, they might need to be tuned.
Args:
weight_decay: The weight decay to use for regularizing the model.
batch_norm_decay: The moving average decay when estimating layer activation
statistics in batch normalization.
batch_norm_epsilon: Small constant to prevent division by zero when
normalizing activations by their variance in batch normalization.
batch_norm_scale: If True, uses an explicit `gamma` multiplier to scale the
activations in the batch normalization layer.
Returns:
An `arg_scope` to use for the resnet models.
"""
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
# The following implies padding='SAME' for pool1, which makes feature
# alignment easier for dense prediction tasks. This is also used in
# https://github.com/facebook/fb.resnet.torch. However the accompanying
# code of 'Deep Residual Learning for Image Recognition' uses
# padding='VALID' for pool1. You can switch to that choice by setting
# slim.arg_scope([slim.max_pool2d], padding='VALID').
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
def resnet_arg_scope(weight_decay=0.0001,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
"""Defines the default ResNet arg scope.
TODO(gpapan): The batch-normalization related default values above are
appropriate for use in conjunction with the reference ResNet models
released at https://github.com/KaimingHe/deep-residual-networks. When
training ResNets from scratch, they might need to be tuned.
Args:
weight_decay: The weight decay to use for regularizing the model.
batch_norm_decay: The moving average decay when estimating layer activation
statistics in batch normalization.
batch_norm_epsilon: Small constant to prevent division by zero when
normalizing activations by their variance in batch normalization.
batch_norm_scale: If True, uses an explicit `gamma` multiplier to scale the
activations in the batch normalization layer.
Returns:
An `arg_scope` to use for the resnet models.
"""
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
# The following implies padding='SAME' for pool1, which makes feature
# alignment easier for dense prediction tasks. This is also used in
# https://github.com/facebook/fb.resnet.torch. However the accompanying
# code of 'Deep Residual Learning for Image Recognition' uses
# padding='VALID' for pool1. You can switch to that choice by setting
# slim.arg_scope([slim.max_pool2d], padding='VALID').
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
def _q_func(self, samples, collection='Q_FUNC'):
'''Returns learning signal and function.
This is the implementation for SBNs for the ELBO.
Args:
samples: dictionary of sampled latent variables
logQ: list of log q(h_i) terms
log_likelihood_func: function used to compute log probs for the latent
variables
Returns:
learning_signal: the "reward" function
function_term: part of the function that depends on the parameters
and needs to have the gradient taken through
'''
reuse=None if not self.run_q_func else True
if self.hparams.task in ['sbn', 'omni']:
with slim.arg_scope([slim.fully_connected],
weights_initializer=slim.variance_scaling_initializer(),
variables_collections=[collection, tf.GraphKeys.GLOBAL_VARIABLES, Q_COLLECTION]):
# for i in reversed(xrange(self.hparams.n_layer)):
# if i == 0:
# n_output = self.hparams.n_input
# else:
# n_output = self.hparams.n_hidden
n_output = self.hparams.n_input
i = self.hparams.n_layer - 1 # use the last layer
input = 2.0*samples[i]['activation']-1.0
h = self._create_transformation(input,
n_output,
reuse=reuse,
scope_prefix='q_func_%d' % i)
h = tf.reduce_sum(h)
self.run_q_func = True
return h, h
elif self.hparams.task == 'sp':
with slim.arg_scope([slim.fully_connected],
weights_initializer=slim.variance_scaling_initializer(),
variables_collections=[collection, tf.GraphKeys.GLOBAL_VARIABLES, Q_COLLECTION]):
n_output = int(self.hparams.n_input/2)
i = self.hparams.n_layer - 1 # use the last layer
input = 2.0*samples[i]['activation']-1.0
h = self._create_transformation(input,
n_output,
reuse=reuse,
scope_prefix='q_func_%d' % i)
self.run_q_func = True
return h, h
def resnet_arg_scope(weight_decay=0.0001,
batch_norm_decay=0.997,
batch_norm_epsilon=1e-5,
batch_norm_scale=True):
"""Defines the default ResNet arg scope.
TODO(gpapan): The batch-normalization related default values above are
appropriate for use in conjunction with the reference ResNet models
released at https://github.com/KaimingHe/deep-residual-networks. When
training ResNets from scratch, they might need to be tuned.
Args:
weight_decay: The weight decay to use for regularizing the model.
batch_norm_decay: The moving average decay when estimating layer activation
statistics in batch normalization.
batch_norm_epsilon: Small constant to prevent division by zero when
normalizing activations by their variance in batch normalization.
batch_norm_scale: If True, uses an explicit `gamma` multiplier to scale the
activations in the batch normalization layer.
Returns:
An `arg_scope` to use for the resnet models.
"""
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
'scale': batch_norm_scale,
'updates_collections': tf.GraphKeys.UPDATE_OPS,
}
with slim.arg_scope(
[slim.conv2d],
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=slim.variance_scaling_initializer(),
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
# The following implies padding='SAME' for pool1, which makes feature
# alignment easier for dense prediction tasks. This is also used in
# https://github.com/facebook/fb.resnet.torch. However the accompanying
# code of 'Deep Residual Learning for Image Recognition' uses
# padding='VALID' for pool1. You can switch to that choice by setting
# slim.arg_scope([slim.max_pool2d], padding='VALID').
with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:
return arg_sc
