def forward(self, inputs, drop_connect_rate=None):
"""
:param inputs: input tensor
:param drop_connect_rate: drop connect rate (float, between 0 and 1)
:return: output of block
"""
# Expansion and Depthwise Convolution
x = inputs
if self.expand_ratio != 1:
x = relu_fn(self._bn0(self._expand_conv(inputs)))
x = relu_fn(self._bn1(self._depthwise_conv(x)))
# Squeeze and Excitation
if self.has_se:
x_squeezed = F.adaptive_avg_pool2d(x, 1)
x_squeezed = self._se_expand(relu_fn(self._se_reduce(x_squeezed)))
x = torch.sigmoid(x_squeezed) * x
x = self._bn2(self._project_conv(x))
# Skip connection and drop connect
if self.id_skip and self.stride == 1 and self.input_filters == self.output_filters:
if drop_connect_rate:
x = drop_connect(x,
p=drop_connect_rate,
training=self.training)
x = x + inputs # skip connection
return x
def forward(self, inputs):
""" Returns output of the final convolution layer """
backbone_indices = getattr(self, 'backbone_indices', None)
if not backbone_indices:
raise ValueError('no backbone indices, something went wrong!')
# Stem
x = relu_fn(self._bn0(self._conv_stem(inputs)))
features = []
# Blocks
for idx, block in enumerate(self._blocks):
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
if idx in backbone_indices:
features.insert(0, x)
# Head
x = relu_fn(self._bn1(self._conv_head(x)))
features.insert(0, x)
return features
def forward(self, inputs):
# Stem
x = relu_fn(self._bn0(self._conv_stem(inputs)))
outputs = []
# Blocks
for idx, block in enumerate(self._blocks):
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
outputs.append(x)
# Head
x = relu_fn(self._bn1(self._conv_head(x)))
result = []
i = 0
r = list(reversed([80, 40, 24, 16])) # [592, 296, 152, 80, 35, 32]
for l in outputs:
sz = l.shape
if i < len(r) and sz[1] == r[i]:
result.append(l)
i = i + 1
return x, (*reversed(result)),
def forward(self, inputs):
x = relu_fn(self.model._bn0(self.model._conv_stem(inputs)))
global_features = []
for idx, block in enumerate(self.model._blocks):
drop_connect_rate = self.model._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self.model._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
if idx in self.idx_list:
global_features.append(x)
x = relu_fn(self.model._bn1(self.model._conv_head(x)))
global_features.append(x)
global_features.reverse()
return global_features
def extract_features(self, inputs):
""" Returns output of the final convolution layer """
# Stem
x = relu_fn(self._bn0(self._conv_stem(inputs)))
# Blocks
self.output_blocks = []
for idx, block in enumerate(self._blocks):
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
self.output_blocks.append(x)
# Head
x = relu_fn(self._bn1(self._conv_head(x)))
return self.output_blocks
def get_feature_layer(self, inputs):
""" Returns output of the final convolution layer """
x = relu_fn(self._bn0(self._conv_stem(inputs)))
for idx, block in enumerate(self._blocks):
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
return x
def extract_features(self, inputs):
ops, total_ops = 0., 0.
# conv_stem is not quantized
x, delta_ops, delta_ops_total = self._conv_stem(inputs,
is_not_quantized=True)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
# still no quantization whatsoever
delta_ops, delta_ops_total = ops_bn(x, is_not_quantized=True)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
delta_ops, delta_ops_total = ops_non_linearity(x,
is_not_quantized=True)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = relu_fn(x)
# quantization appears in these blocks:
for idx, block in enumerate(self._blocks):
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x, delta_ops, delta_ops_total = block(x, drop_connect_rate)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x, delta_ops, delta_ops_total = self._conv_head(x,
is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
delta_ops, delta_ops_total = ops_bn(x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = self._bn1(x)
delta_ops, delta_ops_total = ops_non_linearity(x,
is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = relu_fn(x)
return x, ops, total_ops
def forward(self, x):
result = []
x = relu_fn(self._bn0(self._conv_stem(x)))
result.append(x)
skip_connection_idx = 0
for idx, block in enumerate(self._blocks):
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
if idx == self._skip_connections[skip_connection_idx] - 1:
skip_connection_idx += 1
result.append(x)
return list(reversed(result))
def forward(self, inputs):
""" return a list of tensor each half in size as previous one
e.g. (5,16,128,128) -> (5,24,64,64)...
"""
x_list = [inputs if self.includeX0 else None]
x = relu_fn(self._bn0(self._conv_stem(inputs)))
for idx, block in enumerate(self._blocks):
if idx in self._layers:
x_list.append(x)
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
x_list.append(x)
return x_list
def forward(self, inputs):
outputs = []
# Stem
x = relu_fn(self._bn0(self._conv_stem(inputs)))
# Blocks
for idx, block in enumerate(self._blocks):
drop_connect_rate = self._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
if idx in self._features_idx:
outputs.append(x)
return outputs
def forward(self, inputs):
index = [31, 21, 9, 5]
out = []
x = relu_fn(self.model._bn0(self.model._conv_stem(inputs)))
# Blocks
for idx, block in enumerate(self.model._blocks):
drop_connect_rate = self.model._global_params.drop_connect_rate
if drop_connect_rate:
drop_connect_rate *= float(idx) / len(self.model._blocks)
x = block(x, drop_connect_rate=drop_connect_rate)
if idx in index:
out.append(x)
# # for debug show
# for index, t in enumerate(out):
# print(index, t.shape)
# # Head
# x = relu_fn(self.model._bn1(self.model._conv_head(x)))
return out
def forward(self, inputs, drop_connect_rate=None):
"""
Forward run of the block, see comments to EfficientNet.forward for clarification.
"""
ops, total_ops = 0., 0.
x = inputs
if self._block_args.expand_ratio != 1:
x, delta_ops, delta_ops_total = self._expand_conv(
inputs, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
delta_ops, delta_ops_total = ops_bn(x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = self._bn0(x)
delta_ops, delta_ops_total = ops_non_linearity(
x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = relu_fn(x)
x, delta_ops, delta_ops_total = self._depthwise_conv(
x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
delta_ops, delta_ops_total = ops_bn(x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = self._bn1(x)
delta_ops, delta_ops_total = ops_non_linearity(x,
is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = relu_fn(x)
if self.has_se:
delta_ops, delta_ops_total = ops_adaptive_avg_pool(
x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x_squeezed = F.adaptive_avg_pool2d(x, 1)
x_squeezed, delta_ops, delta_ops_total = self._se_reduce(
x_squeezed, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
delta_ops, delta_ops_total = ops_non_linearity(
x_squeezed, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x_squeezed = relu_fn(x_squeezed)
x_squeezed, delta_ops, delta_ops_total = self._se_expand(
x_squeezed, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
delta_ops, delta_ops_total = ops_non_linearity(
x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = torch.sigmoid(x_squeezed) * x
x, delta_ops, delta_ops_total = self._project_conv(
x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
delta_ops, delta_ops_total = ops_bn(x, is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
x = self._bn2(x)
input_filters, output_filters = self._block_args.input_filters, self._block_args.output_filters
if self.id_skip and self._block_args.stride == 1 and input_filters == output_filters:
if drop_connect_rate:
x = drop_connect(x,
p=drop_connect_rate,
training=self.training)
x = x + inputs # skip connection
delta_ops, delta_ops_total = ops_non_linearity(x,
is_not_quantized=False)
ops, total_ops = ops + delta_ops, total_ops + delta_ops_total
return x, ops, total_ops