def __init__(self, blocks_args=None, global_params=None):
super().__init__()
assert isinstance(blocks_args, list), 'blocks_args should be a list'
assert len(blocks_args) > 0, 'block args must be greater than 0'
self._global_params = global_params
self._blocks_args = blocks_args
# Build blocks
self._blocks = nn.ModuleList([])
self._array = []
for block_args in self._blocks_args:
self.temp = []
# Update block input and output filters based on depth multiplier.
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters,
self._global_params),
output_filters=round_filters(block_args.output_filters,
self._global_params),
num_repeat=round_repeats(block_args.num_repeat,
self._global_params))
# The first block needs to take care of stride and filter size increase.
self._blocks.append(MBConvBlock(block_args, self._global_params))
self.temp.append(MBConvBlock(block_args, self._global_params))
if block_args.num_repeat > 1:
block_args = block_args._replace(
input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(
MBConvBlock(block_args, self._global_params))
self.temp.append(MBConvBlock(block_args, self._global_params))
self._array.append(nn.Sequential(*self.temp))
def __init__(self, blocks_args=None, global_params=None):
super().__init__()
assert isinstance(blocks_args, list), 'blocks_args should be a list'
assert len(blocks_args) > 0, 'block args must be greater than 0'
self._global_params = global_params
self._blocks_args = blocks_args
Conv2d = get_same_padding_conv2d(image_size=global_params.image_size)
bn_mom = 1 - self._global_params.batch_norm_momentum
bn_eps = self._global_params.batch_norm_epsilon
in_channels = 3 # rgb
out_channels = round_filters(
32, self._global_params) # number of output channels
self._conv_stem = Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=2,
bias=False)
self._bn0 = nn.BatchNorm2d(num_features=out_channels,
momentum=bn_mom,
eps=bn_eps)
self._blocks = nn.ModuleList([])
for block_args in self._blocks_args:
# Update block input and output filters based on depth multiplier.
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters,
self._global_params),
output_filters=round_filters(block_args.output_filters,
self._global_params),
num_repeat=round_repeats(block_args.num_repeat,
self._global_params))
self._blocks.append(MBConvBlock(block_args, self._global_params))
if block_args.num_repeat > 1:
block_args = block_args._replace(
input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(
MBConvBlock(block_args, self._global_params))
in_channels = block_args.output_filters # output of final block
out_channels = round_filters(1280, self._global_params)
self._conv_head = Conv2d(in_channels,
out_channels,
kernel_size=1,
bias=False)
self._bn1 = nn.BatchNorm2d(num_features=out_channels,
momentum=bn_mom,
eps=bn_eps)
self._dropout = self._global_params.dropout_rate
self._fc = nn.Linear(out_channels, self._global_params.num_classes)
def __init__(self, blocks_args=None, global_params=None):
super().__init__()
assert isinstance(blocks_args, list), "blocks_args should be a list"
assert len(blocks_args) > 0, "block args must be greater than 0"
self._global_params = global_params
self._blocks_args = blocks_args
# Get static or dynamic convolution depending on image size
Conv2d = get_same_padding_conv2d(image_size=global_params.image_size)
# Batch norm parameters
bn_mom = 1 - self._global_params.batch_norm_momentum
bn_eps = self._global_params.batch_norm_epsilon
# Build blocks
self._blocks = nn.ModuleList([])
for i, block_args in enumerate(self._blocks_args):
# Update block input and output filters based on depth multiplier.
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters,
self._global_params),
output_filters=round_filters(block_args.output_filters,
self._global_params),
num_repeat=round_repeats(block_args.num_repeat,
self._global_params),
)
print(i, block_args)
# The first block needs to take care of stride and filter size increase.
self._blocks.append(MBConvBlock(block_args, self._global_params))
if block_args.num_repeat > 1:
block_args = block_args._replace(
input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(
MBConvBlock(block_args, self._global_params))
# Head
in_channels = block_args.output_filters # output of final block
out_channels = round_filters(1280, self._global_params)
self._conv_head = Conv2d(in_channels,
out_channels,
kernel_size=1,
bias=False)
self._bn1 = nn.BatchNorm2d(num_features=out_channels,
momentum=bn_mom,
eps=bn_eps)
# Define the activation function
self._swish = MemoryEfficientSwish()
def __init__(self, blocks_args=None, global_params=None):
super().__init__()
assert isinstance(blocks_args, list), 'blocks_args should be a list'
assert len(blocks_args) > 0, 'block args must be greater than 0'
self._global_params = global_params
self._blocks_args = blocks_args
# Get static or dynamic convolution depending on image size
Conv2d = get_same_padding_conv2d(image_size=global_params.image_size)
# Batch norm parameters
bn_mom = 1 - self._global_params.batch_norm_momentum
bn_eps = self._global_params.batch_norm_epsilon
# Stem
in_channels = 3 # rgb
out_channels = round_filters(32, self._global_params) # number of output channels
self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)
self._bn0 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
# Build blocks
self._blocks = nn.ModuleList([])
for block_args in self._blocks_args:
# Update block input and output filters based on depth multiplier.
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters, self._global_params),
output_filters=round_filters(block_args.output_filters, self._global_params),
num_repeat=round_repeats(block_args.num_repeat, self._global_params)
)
# The first block needs to take care of stride and filter size increase.
self._blocks.append(MBConvBlock(block_args, self._global_params))
if block_args.num_repeat > 1:
block_args = block_args._replace(input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(MBConvBlock(block_args, self._global_params))
# Head
in_channels = block_args.output_filters # output of final block
out_channels = round_filters(1280, self._global_params)
self._conv_head = Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
self._bn1 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
# Final linear layer
self._avg_pooling = nn.AdaptiveAvgPool2d(1)
self._dropout = nn.Dropout(self._global_params.dropout_rate)
self._fc = nn.Linear(out_channels, self._global_params.num_classes)
self._swish = MemoryEfficientSwish()
def __init__(
self,
efficientnet: nn.Module,
input_size: int,
extract_blocks: Iterable[int] = [5],
freeze: bool = False,
) -> NoReturn:
super().__init__()
self.freeze = freeze
self.input_size = input_size
self._global_params = efficientnet._global_params
# Forget about the fc layers, but copy out everything else.
self._conv_stem = efficientnet._conv_stem
self._bn0 = efficientnet._bn0
self._swish = efficientnet._swish
# EfficientNet saves all layers (including repeats) in a flat _blocks
# parameter. To keep the model comparable we keep the same approach
# but map to extract_blocks when needed.
self.block_to_layer = [-1] # block index to layer index (w/o first).
self.layer_to_block = {} # Probably a bit more efficient forward pass.
layer = -1
for i, block_args in enumerate(efficientnet._blocks_args):
block_args = block_args._replace(num_repeat=round_repeats(
block_args.num_repeat, self._global_params))
layer += block_args.num_repeat
self.block_to_layer.append(layer)
self.layer_to_block[layer] = i + 1
assert layer == len(efficientnet._blocks) - 1 # Last layer.
# Attach blocks only as needed.
assert len(extract_blocks) == len(set(
extract_blocks)), "extract_blocks must only contain unique indices"
assert (sorted(extract_blocks) == extract_blocks
), "attach_block must be sorted in increasing order"
# Last two blocks are conv & fc layer
assert set(extract_blocks) <= set(
range(len(self.block_to_layer) +
1)), "only block indices from 0 to {} are valid".format(
len(self.block_to_layer))
self.extract_blocks = extract_blocks
max_block = max(extract_blocks)
if max_block < len(self.block_to_layer) - 1:
# Last layer inclusive.
self._blocks = efficientnet._blocks[:self.
block_to_layer[max_block] + 1]
else:
self._blocks = efficientnet._blocks
self.orig_len_blocks = len(efficientnet._blocks) # For drop connect.
if max_block >= len(self.block_to_layer):
self._conv_head = efficientnet._conv_head
self._bn1 = efficientnet._bn1
if max_block == len(self.block_to_layer) + 1:
self._avg_pooling = efficientnet._avg_pooling
self._dropout = efficientnet._dropout
self.fc = efficientnet._fc
# freeze learnable model parameters
if self.freeze:
for param in self.parameters():
param.requires_grad = False
import torch