def _maybe_squeeze_and_excite(
params,
input_filters_or_mask,
inner_activation,
gating_activation,
enabled
):
"""Generate a squeeze-and-excite layer or identity function."""
def make_squeeze_and_excite():
return _squeeze_and_excite(
params=params,
input_filters_or_mask=input_filters_or_mask,
inner_activation=inner_activation,
gating_activation=gating_activation)
# We use explicit bool comparisons to make sure a user doesn't pass in a
# bad configuration like enabled=OneOf([False, True, 42]).
if isinstance(enabled, bool):
return make_squeeze_and_excite() if enabled else layers.Identity()
elif isinstance(enabled, schema.OneOf):
options = []
for choice in enabled.choices:
options.append(make_squeeze_and_excite() if choice else layers.Identity())
return layers.maybe_switch_v2(enabled.mask, options)
else:
raise ValueError('Unsupported value for "enabled": {}'.format(enabled))
def _conv(params,
input_filters_or_mask,
output_filters_or_mask,
kernel_size,
strides = (1, 1),
activation = None,
use_batch_norm = True):
"""Conv2D + batch norm + activation, optionally searching over kernel size."""
if isinstance(kernel_size, schema.OneOf) and len(kernel_size.choices) > 1:
choices = []
for kernel_size_value in kernel_size.choices:
choices.append(
_conv_with_fixed_kernel(
params=params,
input_filters_or_mask=input_filters_or_mask,
output_filters_or_mask=output_filters_or_mask,
kernel_size=kernel_size_value,
strides=strides,
activation=activation,
use_batch_norm=use_batch_norm))
return layers.maybe_switch_v2(kernel_size.mask, choices)
else:
if isinstance(kernel_size, schema.OneOf):
kernel_size_value = kernel_size.choices[0]
else:
kernel_size_value = kernel_size
return _conv_with_fixed_kernel(
params=params,
input_filters_or_mask=input_filters_or_mask,
output_filters_or_mask=output_filters_or_mask,
kernel_size=kernel_size_value,
strides=strides,
activation=activation,
use_batch_norm=use_batch_norm)
def _depthwise_conv(
params,
filters_or_mask,
kernel_size,
strides = (1, 1),
activation = None):
"""Depthwise conv + BN + activation, optionally searching over kernel size."""
if isinstance(kernel_size, schema.OneOf) and len(kernel_size.choices) > 1:
choices = []
for kernel_size_value in kernel_size.choices:
choices.append(
_depthwise_conv_with_fixed_kernel(
params=params,
filters_or_mask=filters_or_mask,
kernel_size=kernel_size_value,
strides=strides,
activation=activation))
return layers.maybe_switch_v2(kernel_size.mask, choices)
else:
if isinstance(kernel_size, schema.OneOf):
kernel_size_value = kernel_size.choices[0]
else:
kernel_size_value = kernel_size
return _depthwise_conv_with_fixed_kernel(
params=params,
filters_or_mask=filters_or_mask,
kernel_size=kernel_size_value,
strides=strides,
activation=activation)
def _build_oneof(params, layer_spec, input_filters, output_filters,
filters_base):
"""Select one of N possible choices."""
if len(layer_spec.choices
) > 1 and not params['force_stateless_batch_norm']:
raise ValueError(
'force_stateless_batch_norm must be true for models containing '
'Switch layers (e.g., when performing architecture searches).')
choices = [
_build_layer(params, choice, input_filters, output_filters,
filters_base) for choice in layer_spec.choices
]
return layers.maybe_switch_v2(layer_spec.mask, choices)
