def _build_head_layers(self, model_args):
self._head_layers = build_conv_bn_act(self._global_params,
model_args.feature_mix_layer,
add_bn=True,
act_fn=self.act_fn)
self._head_layers.append(
GlobalAvgPool(self._global_params, keep_dims=True))
def _build_proj(self, output_filters):
kwargs = dict(
cls=Conv2D.__name__,
output_filters=output_filters,
kernel_size=[1, 1],
strides=[1, 1])
self._proj_layers = build_conv_bn_act(self._global_params, kwargs, add_bn=True, act_fn=None)
def _build_dwise(self, output_filters):
kernel_size = self._block_args.kernel_size
kwargs = dict(
cls=DepthwiseConv2D.__name__,
kernel_size=[kernel_size, kernel_size],
strides=self._block_args.strides)
self._dwise_layers = build_conv_bn_act(self._global_params, kwargs, add_bn=True, act_fn=self.act_fn)
def _build_expand(self, output_filters):
if self._block_args.expand_ratio != 1:
kwargs = dict(
cls=Conv2D.__name__,
output_filters=output_filters,
kernel_size=[1, 1],
strides=[1, 1])
else:
kwargs = None
self._expand_layers = build_conv_bn_act(self._global_params, kwargs, add_bn=True, act_fn=self.act_fn)
def _build_dwise(self, output_filters):
kernel_sizes = self._block_args.kernel_size
input_filters = self._block_args.input_filters
expand_ratios = self._block_args.expand_ratio
filter_splits = [int(input_filters * er) for er in expand_ratios]
assert self.expand_filters == sum(filter_splits)
kwargs = dict(
cls=MixConv.__name__,
kernel_sizes=kernel_sizes,
filter_splits=filter_splits,
strides=self._block_args.strides)
self._dwise_layers = build_conv_bn_act(self._global_params, kwargs, add_bn=True, act_fn=self.act_fn)
def _build_head_layers(self, model_args):
final_expand_layer = model_args.get('final_expand_layer')
if final_expand_layer:
self._head_layers = build_conv_bn_act(
self._global_params,
model_args.final_expand_layer,
add_bn=True,
act_fn=self.act_fn)
gap = GlobalAvgPool(self._global_params, keep_dims=True)
feature_mix_layer = graph.construct_with_args(
self._global_params, **model_args.feature_mix_layer)
final_act_fn = Activation(self._global_params, self.act_fn)
self._head_layers.extend([gap, feature_mix_layer, final_act_fn])
else:
super(MBV3CompatJsonNet, self)._build_head_layers(model_args)
def _build_expand(self, output_filters):
kernel_size = self._block_args.kernel_size
if self.kernel_front:
expand_ksize = [kernel_size, kernel_size]
else:
expand_ksize = [self.default_ksize, self.default_ksize]
if self.stride_front:
expand_strides = self._block_args.strides
else:
expand_strides = [1, 1]
kwargs = dict(cls=Conv2D.__name__,
output_filters=output_filters,
kernel_size=expand_ksize,
strides=expand_strides)
self._expand_layers = build_conv_bn_act(self._global_params,
kwargs,
add_bn=True,
act_fn=self.act_fn)
def _build_proj(self, output_filters):
kernel_size = self._block_args.kernel_size
if self.kernel_front:
proj_ksize = [self.default_ksize, self.default_ksize]
else:
proj_ksize = [kernel_size, kernel_size]
strides = self._block_args.strides
if self.stride_front:
proj_strides = [1, 1]
else:
proj_strides = strides
kwargs = dict(cls=Conv2D.__name__,
output_filters=output_filters,
kernel_size=proj_ksize,
strides=proj_strides)
self._proj_layers = build_conv_bn_act(self._global_params,
kwargs,
add_bn=True,
act_fn=None)
def _build_dwise(self, output_filters):
kwargs = None
self._dwise_layers = build_conv_bn_act(self._global_params, kwargs)
def _build_stem_layers(self, model_args):
self._stem_layers = build_conv_bn_act(self._global_params,
model_args.first_conv,
add_bn=True,
act_fn=self.act_fn)