def resnet_model(images,
is_training,
num_classes,
resnet_size=50,
return_intermediate_values=False,
film_generator_fn=None,
film_generator_input=None):
"""Returns resnet model, optionally returning intermediate endpoint tensors.
Args:
images: A Tensor representing a batch [N,H,W,C] of input images.
is_training: A boolean. Set to True to add operations required only when
training the classifier.
num_classes: Dimensionality of output logits emitted by final dense layer.
resnet_size: Size of resnet. One of [18, 34, 50, 101, 152, 200].
return_intermediate_values: If True, returns a dictionary of output and
intermediate activation values.
film_generator_fn: Callable that returns a List (for each block layer) of
Lists (per ResNet block) of FiLM conditioning vectors.
film_generator_input: Embedding tensor to be passed to film_generator_fn.
"""
# For bigger models, we want to use "bottleneck" layers
if resnet_size < 50:
bottleneck = False
else:
bottleneck = True
model = resnet_lib.Model(resnet_size=resnet_size,
bottleneck=bottleneck,
num_classes=num_classes,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=_get_block_sizes(resnet_size),
block_strides=[1, 2, 2, 2],
resnet_version=resnet_lib.DEFAULT_VERSION,
data_format='channels_last',
dtype=resnet_lib.DEFAULT_DTYPE)
final_dense = model(images, is_training, film_generator_fn,
film_generator_input)
if return_intermediate_values:
return resnet_endpoints(model)
else:
return final_dense
def get_resnet50_spatial(images, is_training):
"""ResNet50, but cut off last block and return before global pooling."""
num_classes = 1001 # Dummy value, unused.
model = resnet_lib.Model(resnet_size=50,
bottleneck=True,
num_classes=num_classes,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=[3, 4, 6],
block_strides=[1, 2, 2],
resnet_version=resnet_lib.DEFAULT_VERSION,
data_format='channels_last',
dtype=resnet_lib.DEFAULT_DTYPE)
# Build the graph.
_ = model(images, is_training)
# Return pre-pooled dense spatial features.
return resnet.resnet_endpoints(model)['block_layer3']
def resnet_model(images,
is_training,
num_classes,
resnet_size=50,
return_intermediate_values=False,
film_generator_fn=None,
film_generator_input=None,
pretrain_checkpoint=None):
"""Returns resnet model, optionally returning intermediate endpoint tensors.
Args:
images: A Tensor representing a batch [N,H,W,C] of input images.
is_training: A boolean. Set to True to add operations required only when
training the classifier.
num_classes: Dimensionality of output logits emitted by final dense layer.
resnet_size: Size of resnet. One of [18, 34, 50, 101, 152, 200].
return_intermediate_values: If True, returns a dictionary of output and
intermediate activation values.
film_generator_fn: Callable that returns a List (for each block layer) of
Lists (per ResNet block) of FiLM conditioning vectors.
film_generator_input: Embedding tensor to be passed to film_generator_fn.
pretrain_checkpoint: String to initialize model weights from. Does *NOT*
initialize final logits layer. ResNet checkpoints can be found here:
https://github.com/tensorflow/models/tree/master/official/r1/resnet.
"""
# For bigger models, we want to use "bottleneck" layers
if resnet_size < 50:
bottleneck = False
else:
bottleneck = True
model = resnet_lib.Model(
resnet_size=resnet_size,
bottleneck=bottleneck,
num_classes=num_classes,
num_filters=64,
kernel_size=7,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=_get_block_sizes(resnet_size),
block_strides=[1, 2, 2, 2],
resnet_version=resnet_lib.DEFAULT_VERSION,
data_format='channels_last',
dtype=resnet_lib.DEFAULT_DTYPE
)
final_dense = model(images, is_training,
film_generator_fn, film_generator_input)
if pretrain_checkpoint:
# Initialize variables in ResNet, excluding the final dense layer and any
# optimization-specific variables (e.g. Momentum, Adam Beta).
assignment_map = {}
resnet_scope = _get_resnet_scope()
for var in contrib_framework.get_variables(
scope=resnet_scope, collection=tf.GraphKeys.TRAINABLE_VARIABLES):
if 'dense' not in var.op.name:
# Remove the parent scope prefix.
name_in_ckpt = var.op.name.replace(resnet_scope, 'resnet_model/')
assignment_map[name_in_ckpt] = var
tf.train.init_from_checkpoint(pretrain_checkpoint, assignment_map)
if return_intermediate_values:
return resnet_endpoints(model)
else:
return final_dense
def resnet_model(images,
is_training,
num_classes,
resnet_size=50,
weight_decay=None,
kernel_size=7,
num_filters=64,
return_intermediate_values=False,
film_generator_fn=None,
film_generator_input=None,
pretrain_checkpoint=None):
"""Returns resnet model, optionally returning intermediate endpoint tensors.
Args:
images: A Tensor representing a batch [N,H,W,C] of input images.
is_training: A boolean. Set to True to add operations required only when
training the classifier.
num_classes: Dimensionality of output logits emitted by final dense layer.
resnet_size: Size of resnet. One of [18, 34, 50, 101, 152, 200].
weight_decay: L2 weight regularizer.
kernel_size: Size of the convolution kernels used in the resnet model.
num_filters: Number of filters used.
return_intermediate_values: If True, returns a dictionary of output and
intermediate activation values.
film_generator_fn: Callable that returns a List (for each block layer) of
Lists (per ResNet block) of FiLM conditioning vectors.
film_generator_input: Embedding tensor to be passed to film_generator_fn.
pretrain_checkpoint: String to initialize model weights from. Does *NOT*
initialize final logits layer. ResNet checkpoints can be found here:
https://github.com/tensorflow/models/tree/master/official/r1/resnet.
"""
# For bigger models, we want to use "bottleneck" layers
if resnet_size < 50:
bottleneck = False
else:
bottleneck = True
model = resnet_lib.Model(resnet_size=resnet_size,
bottleneck=bottleneck,
num_classes=num_classes,
num_filters=num_filters,
kernel_size=kernel_size,
conv_stride=2,
first_pool_size=3,
first_pool_stride=2,
block_sizes=_get_block_sizes(resnet_size),
block_strides=[1, 2, 2, 2],
resnet_version=resnet_lib.DEFAULT_VERSION,
data_format='channels_last',
weight_decay=weight_decay,
dtype=resnet_lib.DEFAULT_DTYPE)
final_dense = model(images, is_training, film_generator_fn,
film_generator_input)
if pretrain_checkpoint:
# Initialize variables in ResNet, excluding the final dense layer and any
# optimization-specific variables (e.g. Momentum, Adam Beta).
# When initializing on TPUs, use AbstractT2RModel.init_from_checkpoint_fn.
resnet_init_from_checkpoint_fn(pretrain_checkpoint)
if return_intermediate_values:
return resnet_endpoints(model)
else:
return final_dense