def build_resnet_backbone(cfg):
arch = cfg.MODEL.BACKBONE.ARCH
in_planes = cfg.MODEL.BACKBONE.IN_PLANES
base_planes = cfg.MODEL.BACKBONE.BASE_PLANES
layer_planes = cfg.MODEL.BACKBONE.LAYER_PLANES
down_samples = cfg.MODEL.BACKBONE.DOWNSAMPLES
conv_layer = get_conv(cfg)
norm_layer = get_norm(cfg)
act_layer = get_act(cfg)
zero_init_residual = cfg.MODEL.RECOGNIZER.ZERO_INIT_RESIDUAL
# for attention
with_attentions = cfg.MODEL.ATTENTION.WITH_ATTENTIONS
reduction = cfg.MODEL.ATTENTION.REDUCTION
attention_type = cfg.MODEL.ATTENTION.ATTENTION_TYPE
block_layer, layer_blocks, groups, width_per_group = arch_settings[arch]
return ResNetBackbone(
in_channels=in_planes,
base_channels=base_planes,
layer_channels=layer_planes,
layer_blocks=layer_blocks,
down_samples=down_samples,
groups=groups,
width_per_group=width_per_group,
with_attentions=with_attentions,
reduction=reduction,
attention_type=attention_type,
block_layer=block_layer,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer,
zero_init_residual=zero_init_residual,
)
def build_mnasnet(cfg):
arch = cfg.MODEL.BACKBONE.ARCH
in_planes = cfg.MODEL.BACKBONE.IN_PLANES
norm_layer = get_norm(cfg)
# compression
width_multiplier = cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER
round_nearest = cfg.MODEL.COMPRESSION.ROUND_NEAREST
# attention
with_attention = cfg.MODEL.ATTENTION.WITH_ATTENTION
reduction = cfg.MODEL.ATTENTION.REDUCTION
attention_type = cfg.MODEL.ATTENTION.ATTENTION_TYPE
# conv
conv_layer = get_conv(cfg)
# act
act_layer = get_act(cfg)
out_planes = 1280
stage_setting = arch_settings[arch]
return MNASNetBackbone(in_channels=in_planes,
out_channels=out_planes,
stage_setting=stage_setting,
width_multiplier=width_multiplier,
round_nearest=round_nearest,
with_attention=with_attention,
reduction=reduction,
attention_type=attention_type,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer)
def build_sfv2_backbone(cfg):
arch = cfg.MODEL.BACKBONE.ARCH
in_channels = cfg.MODEL.BACKBONE.IN_PLANES
base_channels = cfg.MODEL.BACKBONE.BASE_PLANES
round_nearest = cfg.MODEL.COMPRESSION.ROUND_NEAREST
block_layer, stage_channels, stage_blocks, out_channels = arch_settings[
arch]
base_channels = make_divisible(base_channels, round_nearest)
for i in range(len(stage_channels)):
stage_channels[i] = make_divisible(stage_channels[i], round_nearest)
out_channels = make_divisible(out_channels, round_nearest)
down_samples = cfg.MODEL.BACKBONE.DOWNSAMPLES
conv_layer = get_conv(cfg)
norm_layer = get_norm(cfg)
act_layer = get_act(cfg)
return ShuffleNetV2Backbone(in_channels=in_channels,
base_channels=base_channels,
out_channels=out_channels,
stage_channels=stage_channels,
stage_blocks=stage_blocks,
downsamples=down_samples,
block_layer=block_layer,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer)
def build_sfv1_backbone(cfg):
arch = cfg.MODEL.BACKBONE.ARCH
round_nearest = cfg.MODEL.COMPRESSION.ROUND_NEAREST
in_channels = cfg.MODEL.BACKBONE.IN_PLANES
base_channels = cfg.MODEL.BACKBONE.BASE_PLANES
downsamples = cfg.MODEL.BACKBONE.DOWNSAMPLES
with_groups = cfg.MODEL.BACKBONE.WITH_GROUPS
conv_layer = get_conv(cfg)
norm_layer = get_norm(cfg)
act_layer = get_act(cfg)
block_layer, groups, stage_channels, layer_blocks, width_multiplier = arch_settings[
arch]
base_channels = make_divisible(base_channels * width_multiplier,
round_nearest)
for i in range(len(stage_channels)):
stage_channels[i] = make_divisible(
stage_channels[i] * width_multiplier, round_nearest)
return ShuffleNetV1Backbone(
in_channels=in_channels,
base_channels=base_channels,
groups=groups,
stage_channels=stage_channels,
stage_blocks=layer_blocks,
downsamples=downsamples,
with_groups=with_groups,
block_layer=block_layer,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer,
)
def build_mbv3_backbone(cfg):
arch = cfg.MODEL.BACKBONE.ARCH
in_channels = cfg.MODEL.BACKBONE.IN_PLANES
norm_layer = get_norm(cfg)
# compression
width_multiplier = cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER
round_nearest = cfg.MODEL.COMPRESSION.ROUND_NEAREST
# attention
with_attention = cfg.MODEL.ATTENTION.WITH_ATTENTION
reduction = cfg.MODEL.ATTENTION.REDUCTION
attention_type = cfg.MODEL.ATTENTION.ATTENTION_TYPE
# conv
conv_layer = get_conv(cfg)
# act
act_layer = get_act(cfg)
sigmoid_type = cfg.MODEL.ACT.SIGMOID_TYPE
base_channels, feature_dims, inner_dims, layer_setting = arch_settings[
arch]
return MobileNetV3Backbone(in_channels=in_channels,
base_channels=base_channels,
out_channels=feature_dims,
width_multiplier=width_multiplier,
round_nearest=round_nearest,
with_attention=with_attention,
reduction=reduction,
attention_type=attention_type,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer,
sigmoid_type=sigmoid_type)
def build_resnet3d_backbone(cfg):
arch = cfg.MODEL.BACKBONE.ARCH
torchvision_pretrained = cfg.MODEL.RECOGNIZER.TORCHVISION_PRETRAINED
conv_layer = get_conv(cfg)
norm_layer = get_norm(cfg)
act_layer = get_act(cfg)
zero_init_residual = cfg.MODEL.RECOGNIZER.ZERO_INIT_RESIDUAL
# for backbone
in_planes = cfg.MODEL.BACKBONE.IN_PLANES
base_planes = cfg.MODEL.BACKBONE.BASE_PLANES
conv1_kernel = cfg.MODEL.BACKBONE.CONV1_KERNEL
conv1_stride = cfg.MODEL.BACKBONE.CONV1_STRIDE
conv1_padding = cfg.MODEL.BACKBONE.CONV1_PADDING
pool1_kernel = cfg.MODEL.BACKBONE.POOL1_KERNEL
pool1_stride = cfg.MODEL.BACKBONE.POOL1_STRIDE
pool1_padding = cfg.MODEL.BACKBONE.POOL1_PADDING
with_pool2 = cfg.MODEL.BACKBONE.WITH_POOL2
layer_planes = cfg.MODEL.BACKBONE.LAYER_PLANES
down_samples = cfg.MODEL.BACKBONE.DOWNSAMPLES
temporal_strides = cfg.MODEL.BACKBONE.TEMPORAL_STRIDES
inflate_list = cfg.MODEL.BACKBONE.INFLATE_LIST
inflate_style = cfg.MODEL.BACKBONE.INFLATE_STYLE
block_layer, layer_blocks, groups, width_per_group = arch_settings[arch]
state_dict_2d = load_state_dict_from_url(model_urls[arch], progress=True) \
if torchvision_pretrained else None
return ResNet3DBackbone(in_planes=in_planes,
base_planes=base_planes,
conv1_kernel=conv1_kernel,
conv1_stride=conv1_stride,
conv1_padding=conv1_padding,
pool1_kernel=pool1_kernel,
pool1_stride=pool1_stride,
pool1_padding=pool1_padding,
with_pool2=with_pool2,
layer_planes=layer_planes,
layer_blocks=layer_blocks,
down_samples=down_samples,
temporal_strides=temporal_strides,
inflate_list=inflate_list,
inflate_style=inflate_style,
groups=groups,
width_per_group=width_per_group,
block_layer=block_layer,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer,
zero_init_residual=zero_init_residual,
state_dict_2d=state_dict_2d)
def test_resnet_gn():
cfg.MODEL.NORM.TYPE = 'GroupNorm'
norm_layer = get_norm(cfg)
print(norm_layer)
# for custom
model = ResNet(arch="resnet50", num_classes=1000, norm_layer=norm_layer)
print(model)
test_data(model, (1, 3, 224, 224), (1, 1000))
# resnetxt_32x4d
model = ResNet(arch="resnext50_32x4d",
num_classes=1000,
norm_layer=norm_layer)
print(model)
test_data(model, (3, 3, 224, 224), (3, 1000))
def build_torchvision_mbv2(cfg):
torchvision_pretrained = cfg.MODEL.RECOGNIZER.TORCHVISION_PRETRAINED
pretrained_num_classes = cfg.MODEL.RECOGNIZER.PRETRAINED_NUM_CLASSES
fix_bn = cfg.MODEL.NORM.FIX_BN
partial_bn = cfg.MODEL.NORM.PARTIAL_BN
num_classes = cfg.MODEL.HEAD.NUM_CLASSES
norm_layer = get_norm(cfg)
width_multiplier = cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER
return TorchvisionMobileNetV2(
num_classes=num_classes,
torchvision_pretrained=torchvision_pretrained,
pretrained_num_classes=pretrained_num_classes,
width_multiplier=width_multiplier,
fix_bn=fix_bn,
partial_bn=partial_bn,
norm_layer=norm_layer)
def build_mbv2_backbone(cfg):
in_channels = cfg.MODEL.BACKBONE.IN_PLANES
base_channels = cfg.MODEL.BACKBONE.BASE_PLANES
out_channels = cfg.MODEL.HEAD.FEATURE_DIMS
round_nearest = cfg.MODEL.COMPRESSION.ROUND_NEAREST
width_multiplier = cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER
conv_layer = get_conv(cfg)
norm_layer = get_norm(cfg)
act_layer = get_act(cfg)
inverted_residual_setting = [
# t, c, n, s
[1, 16, 1, 1],
[6, 24, 2, 2],
[6, 32, 3, 2],
[6, 64, 4, 2],
[6, 96, 3, 1],
[6, 160, 3, 2],
[6, 320, 1, 1],
]
base_channels = make_divisible(base_channels * width_multiplier,
round_nearest)
for i in range(len(inverted_residual_setting)):
channel = inverted_residual_setting[i][1]
inverted_residual_setting[i][1] = make_divisible(
channel * width_multiplier, round_nearest)
out_channels = make_divisible(out_channels * width_multiplier,
round_nearest)
return MobileNetV2Backbone(
in_channels=in_channels,
out_channels=out_channels,
base_channels=base_channels,
inverted_residual_setting=inverted_residual_setting,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer)
def build_mbv1_backbone(cfg):
in_planes = cfg.MODEL.BACKBONE.IN_PLANES
base_planes = cfg.MODEL.BACKBONE.BASE_PLANES
layer_planes = cfg.MODEL.BACKBONE.LAYER_PLANES
strides = cfg.MODEL.BACKBONE.STRIDES
conv_layer = get_conv(cfg)
norm_layer = get_norm(cfg)
act_layer = get_act(cfg)
width_multiplier = cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER
round_nearest = cfg.MODEL.COMPRESSION.ROUND_NEAREST
base_planes = make_divisible(base_planes * width_multiplier, round_nearest)
layer_planes = [
make_divisible(layer_plane * width_multiplier, round_nearest)
for layer_plane in layer_planes
]
return MobileNetV1Backbone(in_channels=in_planes,
base_channels=base_planes,
layer_channels=layer_planes,
strides=strides,
conv_layer=conv_layer,
norm_layer=norm_layer,
act_layer=act_layer)
