Migrations
/
pytorch-image-models
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Make 'regnetz' model experiments closer to actual RegNetZ, bottleneck expansion, expand from in_chs, no shortcut on stride 2, tweak model sizes

Browse Source
pull/880/head
Ross Wightman 3 years ago
parent b81e79aae9
commit 0ca687f224
2 changed files with 171 additions and 138 deletions
Show Stats Download Patch File Download Diff File

52
timm/models/byoanet.py
Unescape View File

@ -57,8 +57,14 @@ default_cfgs = {
input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)), input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)),
'lambda_resnet26t': _cfg( 'lambda_resnet26t': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/lambda_resnet26t_256-b040fce6.pth', url='',
min_input_size=(3, 128, 128), input_size=(3, 256, 256), pool_size=(8, 8)),
'lambda_resnet50ts': _cfg(
url='',
min_input_size=(3, 128, 128), input_size=(3, 256, 256), pool_size=(8, 8)), min_input_size=(3, 128, 128), input_size=(3, 256, 256), pool_size=(8, 8)),
'lambda_resnet26rt_256': _cfg(
url='',
fixed_input_size=True, input_size=(3, 256, 256), pool_size=(8, 8)),
} }
@ -198,6 +204,33 @@ model_cfgs = dict(
self_attn_layer='lambda', self_attn_layer='lambda',
self_attn_kwargs=dict(r=9) self_attn_kwargs=dict(r=9)
), ),
lambda_resnet50ts=ByoModelCfg(
blocks=(
ByoBlockCfg(type='bottle', d=3, c=256, s=1, gs=0, br=0.25),
interleave_blocks(types=('bottle', 'self_attn'), every=4, d=4, c=512, s=2, gs=0, br=0.25),
interleave_blocks(types=('bottle', 'self_attn'), d=6, c=1024, s=2, gs=0, br=0.25),
interleave_blocks(types=('bottle', 'self_attn'), d=3, c=2048, s=2, gs=0, br=0.25),
),
stem_chs=64,
stem_type='tiered',
stem_pool='maxpool',
act_layer='silu',
self_attn_layer='lambda',
self_attn_kwargs=dict(r=9)
),
lambda_resnet26rt_256=ByoModelCfg(
blocks=(
ByoBlockCfg(type='bottle', d=2, c=256, s=1, gs=0, br=0.25),
ByoBlockCfg(type='bottle', d=2, c=512, s=2, gs=0, br=0.25),
interleave_blocks(types=('bottle', 'self_attn'), d=2, c=1024, s=2, gs=0, br=0.25),
ByoBlockCfg(type='self_attn', d=2, c=2048, s=2, gs=0, br=0.25),
),
stem_chs=64,
stem_type='tiered',
stem_pool='maxpool',
self_attn_layer='lambda',
self_attn_kwargs=dict(r=None)
),
) )
@ -275,6 +308,21 @@ def eca_halonext26ts(pretrained=False, **kwargs):
@register_model @register_model
def lambda_resnet26t(pretrained=False, **kwargs): def lambda_resnet26t(pretrained=False, **kwargs):
""" Lambda-ResNet-26T. Lambda layers in last two stages. """ Lambda-ResNet-26-T. Lambda layers w/ conv pos in last two stages.
""" """
return _create_byoanet('lambda_resnet26t', pretrained=pretrained, **kwargs) return _create_byoanet('lambda_resnet26t', pretrained=pretrained, **kwargs)
@register_model
def lambda_resnet50ts(pretrained=False, **kwargs):
""" Lambda-ResNet-50-TS. SiLU act. Lambda layers w/ conv pos in last two stages.
"""
return _create_byoanet('lambda_resnet50ts', pretrained=pretrained, **kwargs)
@register_model
def lambda_resnet26rt_256(pretrained=False, **kwargs):
""" Lambda-ResNet-26-R-T. Lambda layers w/ rel pos embed in last two stages.
"""
kwargs.setdefault('img_size', 256)
return _create_byoanet('lambda_resnet26rt_256', pretrained=pretrained, **kwargs)

257
timm/models/byobnet.py
Unescape View File

@ -51,6 +51,16 @@ def _cfg(url='', **kwargs):
} }
def _cfgr(url='', **kwargs):
return {
'url': url, 'num_classes': 1000, 'input_size': (3, 256, 256), 'pool_size': (8, 8),
'crop_pct': 0.9, 'interpolation': 'bicubic',
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
'first_conv': 'stem.conv1.conv', 'classifier': 'head.fc',
**kwargs
}
default_cfgs = { default_cfgs = {
# GPU-Efficient (ResNet) weights # GPU-Efficient (ResNet) weights
'gernet_s': _cfg( 'gernet_s': _cfg(
@ -92,65 +102,50 @@ default_cfgs = {
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnet51q_ra2-d47dcc76.pth', url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnet51q_ra2-d47dcc76.pth',
first_conv='stem.conv1', input_size=(3, 256, 256), pool_size=(8, 8), first_conv='stem.conv1', input_size=(3, 256, 256), pool_size=(8, 8),
test_input_size=(3, 288, 288), crop_pct=1.0), test_input_size=(3, 288, 288), crop_pct=1.0),
'resnet61q': _cfg( 'resnet61q': _cfgr(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnet61q_ra2-6afc536c.pth', url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnet61q_ra2-6afc536c.pth',
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), test_input_size=(3, 288, 288), crop_pct=1.0),
test_input_size=(3, 288, 288), crop_pct=1.0, interpolation='bicubic'),
'resnext26ts': _cfgr(
'resnext26ts': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/resnext26ts_256_ra2-8bbd9106.pth'),
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/resnext26ts_256_ra2-8bbd9106.pth', 'gcresnext26ts': _cfgr(
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnext26ts_256-e414378b.pth'),
'gcresnext26ts': _cfg( 'seresnext26ts': _cfgr(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnext26ts_256-e414378b.pth', url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/seresnext26ts_256-6f0d74a3.pth'),
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), 'eca_resnext26ts': _cfgr(
'seresnext26ts': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/eca_resnext26ts_256-5a1d030f.pth'),
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/seresnext26ts_256-6f0d74a3.pth', 'bat_resnext26ts': _cfgr(
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'),
'eca_resnext26ts': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/eca_resnext26ts_256-5a1d030f.pth',
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'),
'bat_resnext26ts': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/bat_resnext26ts_256-fa6fd595.pth', url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/bat_resnext26ts_256-fa6fd595.pth',
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic',
min_input_size=(3, 256, 256)), min_input_size=(3, 256, 256)),
'resnet32ts': _cfg( 'resnet32ts': _cfgr(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/resnet32ts_256-aacf5250.pth', url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/resnet32ts_256-aacf5250.pth'),
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), 'resnet33ts': _cfgr(
'resnet33ts': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/resnet33ts_256-e91b09a4.pth'),
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/resnet33ts_256-e91b09a4.pth', 'gcresnet33ts': _cfgr(
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnet33ts_256-0e0cd345.pth'),
'gcresnet33ts': _cfg( 'seresnet33ts': _cfgr(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnet33ts_256-0e0cd345.pth', url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/seresnet33ts_256-f8ad44d9.pth'),
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), 'eca_resnet33ts': _cfgr(
'seresnet33ts': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/eca_resnet33ts_256-8f98face.pth'),
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/seresnet33ts_256-f8ad44d9.pth',
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), 'gcresnet50t': _cfgr(
'eca_resnet33ts': _cfg( url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnet50t_256-96374d1c.pth'),
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/eca_resnet33ts_256-8f98face.pth',
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), 'gcresnext50ts': _cfgr(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnext50ts_256-3e0f515e.pth'),
'gcresnet50t': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnet50t_256-96374d1c.pth', # experimental models, likely to change ot be removed
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), 'regnetz_b': _cfgr(
'gcresnext50ts': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/gcresnext50ts_256-3e0f515e.pth',
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'),
# experimental models
'regnetz_b': _cfg(
url='', url='',
mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5),
input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), input_size=(3, 224, 224), pool_size=(7, 7), first_conv='stem.conv'),
'regnetz_c': _cfg( 'regnetz_c': _cfgr(
url='', url='',
mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), imean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), first_conv='stem.conv'),
input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'), 'regnetz_d': _cfgr(
'regnetz_d': _cfg(
url='', url='',
mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
first_conv='stem.conv1.conv', input_size=(3, 256, 256), pool_size=(8, 8), interpolation='bicubic'),
} }
@ -507,46 +502,52 @@ model_cfgs = dict(
# experimental models, closer to a RegNetZ than a ResNet. Similar to EfficientNets but w/ groups instead of DW # experimental models, closer to a RegNetZ than a ResNet. Similar to EfficientNets but w/ groups instead of DW
regnetz_b=ByoModelCfg( regnetz_b=ByoModelCfg(
blocks=( blocks=(
ByoBlockCfg(type='bottle', d=2, c=192, s=2, gs=24, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=2, c=48, s=2, gs=16, br=3),
ByoBlockCfg(type='bottle', d=6, c=384, s=2, gs=24, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=6, c=96, s=2, gs=16, br=3),
ByoBlockCfg(type='bottle', d=12, c=768, s=2, gs=24, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=12, c=192, s=2, gs=16, br=3),
ByoBlockCfg(type='bottle', d=2, c=1536, s=2, gs=24, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=2, c=288, s=2, gs=16, br=3),
), ),
stem_chs=32, stem_chs=32,
stem_pool='', stem_pool='',
num_features=1792, downsample='',
num_features=1536,
act_layer='silu', act_layer='silu',
attn_layer='se', attn_layer='se',
attn_kwargs=dict(rd_ratio=0.25), attn_kwargs=dict(rd_ratio=0.25),
block_kwargs=dict(bottle_in=True, linear_out=True),
), ),
regnetz_c=ByoModelCfg( regnetz_c=ByoModelCfg(
blocks=( blocks=(
ByoBlockCfg(type='bottle', d=2, c=128, s=2, gs=16, br=0.5, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=2, c=48, s=2, gs=16, br=4),
ByoBlockCfg(type='bottle', d=6, c=512, s=2, gs=32, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=6, c=96, s=2, gs=16, br=4),
ByoBlockCfg(type='bottle', d=12, c=768, s=2, gs=32, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=12, c=192, s=2, gs=16, br=4),
ByoBlockCfg(type='bottle', d=2, c=1536, s=2, gs=64, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=2, c=288, s=2, gs=16, br=4),
), ),
stem_chs=32, stem_chs=32,
stem_pool='', stem_pool='',
num_features=1792, downsample='',
num_features=1536,
act_layer='silu', act_layer='silu',
attn_layer='se', attn_layer='se',
attn_kwargs=dict(rd_ratio=0.25), attn_kwargs=dict(rd_ratio=0.25),
block_kwargs=dict(bottle_in=True, linear_out=True),
), ),
regnetz_d=ByoModelCfg( regnetz_d=ByoModelCfg(
blocks=( blocks=(
ByoBlockCfg(type='bottle', d=3, c=256, s=1, gs=64, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=3, c=64, s=1, gs=32, br=4),
ByoBlockCfg(type='bottle', d=6, c=512, s=2, gs=64, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=6, c=128, s=2, gs=32, br=4),
ByoBlockCfg(type='bottle', d=12, c=768, s=2, gs=64, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=12, c=256, s=2, gs=32, br=4),
ByoBlockCfg(type='bottle', d=3, c=1536, s=2, gs=64, br=0.25, block_kwargs=dict(linear_out=True)), ByoBlockCfg(type='bottle', d=3, c=384, s=2, gs=32, br=4),
), ),
stem_chs=128, stem_chs=64,
stem_type='quad', stem_type='tiered',
stem_pool='', stem_pool='',
downsample='',
num_features=1792, num_features=1792,
act_layer='silu', act_layer='silu',
attn_layer='se', attn_layer='se',
attn_kwargs=dict(rd_ratio=0.25), attn_kwargs=dict(rd_ratio=0.25),
block_kwargs=dict(bottle_in=True, linear_out=True),
), ),
) )
@ -802,11 +803,17 @@ class DownsampleAvg(nn.Module):
return self.conv(self.pool(x)) return self.conv(self.pool(x))
def create_downsample(downsample_type, layers: LayerFn, **kwargs): def create_shortcut(downsample_type, layers: LayerFn, in_chs, out_chs, stride, dilation, **kwargs):
if downsample_type == 'avg': assert downsample_type in ('avg', 'conv1x1', '')
return DownsampleAvg(**kwargs) if in_chs != out_chs or stride != 1 or dilation[0] != dilation[1]:
if not downsample_type:
return None # no shortcut
elif downsample_type == 'avg':
return DownsampleAvg(in_chs, out_chs, stride=stride, dilation=dilation[0], **kwargs)
else:
return layers.conv_norm_act(in_chs, out_chs, kernel_size=1, stride=stride, dilation=dilation[0], **kwargs)
else: else:
return layers.conv_norm_act(kwargs.pop('in_chs'), kwargs.pop('out_chs'), kernel_size=1, **kwargs) return nn.Identity() # identity shortcut
class BasicBlock(nn.Module): class BasicBlock(nn.Module):
@ -822,12 +829,9 @@ class BasicBlock(nn.Module):
mid_chs = make_divisible(out_chs * bottle_ratio) mid_chs = make_divisible(out_chs * bottle_ratio)
groups = num_groups(group_size, mid_chs) groups = num_groups(group_size, mid_chs)
if in_chs != out_chs or stride != 1 or dilation[0] != dilation[1]: self.shortcut = create_shortcut(
self.shortcut = create_downsample( downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation,
downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation[0], apply_act=False, layers=layers)
apply_act=False, layers=layers)
else:
self.shortcut = nn.Identity()
self.conv1_kxk = layers.conv_norm_act(in_chs, mid_chs, kernel_size, stride=stride, dilation=dilation[0]) self.conv1_kxk = layers.conv_norm_act(in_chs, mid_chs, kernel_size, stride=stride, dilation=dilation[0])
self.attn = nn.Identity() if attn_last or layers.attn is None else layers.attn(mid_chs) self.attn = nn.Identity() if attn_last or layers.attn is None else layers.attn(mid_chs)
@ -838,23 +842,21 @@ class BasicBlock(nn.Module):
self.act = nn.Identity() if linear_out else layers.act(inplace=True) self.act = nn.Identity() if linear_out else layers.act(inplace=True)
def init_weights(self, zero_init_last: bool = False): def init_weights(self, zero_init_last: bool = False):
if zero_init_last: if zero_init_last and self.shortcut is not None:
nn.init.zeros_(self.conv2_kxk.bn.weight) nn.init.zeros_(self.conv2_kxk.bn.weight)
for attn in (self.attn, self.attn_last): for attn in (self.attn, self.attn_last):
if hasattr(attn, 'reset_parameters'): if hasattr(attn, 'reset_parameters'):
attn.reset_parameters() attn.reset_parameters()
def forward(self, x): def forward(self, x):
shortcut = self.shortcut(x) shortcut = x
# residual path
x = self.conv1_kxk(x) x = self.conv1_kxk(x)
x = self.conv2_kxk(x) x = self.conv2_kxk(x)
x = self.attn(x) x = self.attn(x)
x = self.drop_path(x) x = self.drop_path(x)
if self.shortcut is not None:
x = self.act(x + shortcut) x = x + self.shortcut(shortcut)
return x return self.act(x)
class BottleneckBlock(nn.Module): class BottleneckBlock(nn.Module):
@ -862,24 +864,18 @@ class BottleneckBlock(nn.Module):
""" """
def __init__(self, in_chs, out_chs, kernel_size=3, stride=1, dilation=(1, 1), bottle_ratio=1., group_size=None, def __init__(self, in_chs, out_chs, kernel_size=3, stride=1, dilation=(1, 1), bottle_ratio=1., group_size=None,
downsample='avg', attn_last=False, linear_out=False, extra_conv=False, layers: LayerFn = None, downsample='avg', attn_last=False, linear_out=False, extra_conv=False, bottle_in=False,
drop_block=None, drop_path_rate=0.): layers: LayerFn = None, drop_block=None, drop_path_rate=0.):
super(BottleneckBlock, self).__init__() super(BottleneckBlock, self).__init__()
layers = layers or LayerFn() layers = layers or LayerFn()
mid_chs = make_divisible(out_chs * bottle_ratio) mid_chs = make_divisible((in_chs if bottle_in else out_chs) * bottle_ratio)
groups = num_groups(group_size, mid_chs) groups = num_groups(group_size, mid_chs)
if in_chs != out_chs or stride != 1 or dilation[0] != dilation[1]: self.shortcut = create_shortcut(
self.shortcut = create_downsample( downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation,
downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation[0], apply_act=False, layers=layers)
apply_act=False, layers=layers)
else:
self.shortcut = nn.Identity()
self.conv1_1x1 = layers.conv_norm_act(in_chs, mid_chs, 1) self.conv1_1x1 = layers.conv_norm_act(in_chs, mid_chs, 1)
self.conv2_kxk = layers.conv_norm_act(
mid_chs, mid_chs, kernel_size, stride=stride, dilation=dilation[0],
groups=groups, drop_block=drop_block)
self.conv2_kxk = layers.conv_norm_act( self.conv2_kxk = layers.conv_norm_act(
mid_chs, mid_chs, kernel_size, stride=stride, dilation=dilation[0], mid_chs, mid_chs, kernel_size, stride=stride, dilation=dilation[0],
groups=groups, drop_block=drop_block) groups=groups, drop_block=drop_block)
@ -895,15 +891,14 @@ class BottleneckBlock(nn.Module):
self.act = nn.Identity() if linear_out else layers.act(inplace=True) self.act = nn.Identity() if linear_out else layers.act(inplace=True)
def init_weights(self, zero_init_last: bool = False): def init_weights(self, zero_init_last: bool = False):
if zero_init_last: if zero_init_last and self.shortcut is not None:
nn.init.zeros_(self.conv3_1x1.bn.weight) nn.init.zeros_(self.conv3_1x1.bn.weight)
for attn in (self.attn, self.attn_last): for attn in (self.attn, self.attn_last):
if hasattr(attn, 'reset_parameters'): if hasattr(attn, 'reset_parameters'):
attn.reset_parameters() attn.reset_parameters()
def forward(self, x): def forward(self, x):
shortcut = self.shortcut(x) shortcut = x
x = self.conv1_1x1(x) x = self.conv1_1x1(x)
x = self.conv2_kxk(x) x = self.conv2_kxk(x)
x = self.conv2b_kxk(x) x = self.conv2b_kxk(x)
@ -911,9 +906,9 @@ class BottleneckBlock(nn.Module):
x = self.conv3_1x1(x) x = self.conv3_1x1(x)
x = self.attn_last(x) x = self.attn_last(x)
x = self.drop_path(x) x = self.drop_path(x)
if self.shortcut is not None:
x = self.act(x + shortcut) x = x + self.shortcut(shortcut)
return x return self.act(x)
class DarkBlock(nn.Module): class DarkBlock(nn.Module):
@ -935,12 +930,9 @@ class DarkBlock(nn.Module):
mid_chs = make_divisible(out_chs * bottle_ratio) mid_chs = make_divisible(out_chs * bottle_ratio)
groups = num_groups(group_size, mid_chs) groups = num_groups(group_size, mid_chs)
if in_chs != out_chs or stride != 1 or dilation[0] != dilation[1]: self.shortcut = create_shortcut(
self.shortcut = create_downsample( downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation,
downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation[0], apply_act=False, layers=layers)
apply_act=False, layers=layers)
else:
self.shortcut = nn.Identity()
self.conv1_1x1 = layers.conv_norm_act(in_chs, mid_chs, 1) self.conv1_1x1 = layers.conv_norm_act(in_chs, mid_chs, 1)
self.attn = nn.Identity() if attn_last or layers.attn is None else layers.attn(mid_chs) self.attn = nn.Identity() if attn_last or layers.attn is None else layers.attn(mid_chs)
@ -952,22 +944,22 @@ class DarkBlock(nn.Module):
self.act = nn.Identity() if linear_out else layers.act(inplace=True) self.act = nn.Identity() if linear_out else layers.act(inplace=True)
def init_weights(self, zero_init_last: bool = False): def init_weights(self, zero_init_last: bool = False):
if zero_init_last: if zero_init_last and self.shortcut is not None:
nn.init.zeros_(self.conv2_kxk.bn.weight) nn.init.zeros_(self.conv2_kxk.bn.weight)
for attn in (self.attn, self.attn_last): for attn in (self.attn, self.attn_last):
if hasattr(attn, 'reset_parameters'): if hasattr(attn, 'reset_parameters'):
attn.reset_parameters() attn.reset_parameters()
def forward(self, x): def forward(self, x):
shortcut = self.shortcut(x) shortcut = x
x = self.conv1_1x1(x) x = self.conv1_1x1(x)
x = self.attn(x) x = self.attn(x)
x = self.conv2_kxk(x) x = self.conv2_kxk(x)
x = self.attn_last(x) x = self.attn_last(x)
x = self.drop_path(x) x = self.drop_path(x)
x = self.act(x + shortcut) if self.shortcut is not None:
return x x = x + self.shortcut(shortcut)
return self.act(x)
class EdgeBlock(nn.Module): class EdgeBlock(nn.Module):
@ -988,12 +980,9 @@ class EdgeBlock(nn.Module):
mid_chs = make_divisible(out_chs * bottle_ratio) mid_chs = make_divisible(out_chs * bottle_ratio)
groups = num_groups(group_size, mid_chs) groups = num_groups(group_size, mid_chs)
if in_chs != out_chs or stride != 1 or dilation[0] != dilation[1]: self.shortcut = create_shortcut(
self.shortcut = create_downsample( downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation,
downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation[0], apply_act=False, layers=layers)
apply_act=False, layers=layers)
else:
self.shortcut = nn.Identity()
self.conv1_kxk = layers.conv_norm_act( self.conv1_kxk = layers.conv_norm_act(
in_chs, mid_chs, kernel_size, stride=stride, dilation=dilation[0], in_chs, mid_chs, kernel_size, stride=stride, dilation=dilation[0],
@ -1005,22 +994,22 @@ class EdgeBlock(nn.Module):
self.act = nn.Identity() if linear_out else layers.act(inplace=True) self.act = nn.Identity() if linear_out else layers.act(inplace=True)
def init_weights(self, zero_init_last: bool = False): def init_weights(self, zero_init_last: bool = False):
if zero_init_last: if zero_init_last and self.shortcut is not None:
nn.init.zeros_(self.conv2_1x1.bn.weight) nn.init.zeros_(self.conv2_1x1.bn.weight)
for attn in (self.attn, self.attn_last): for attn in (self.attn, self.attn_last):
if hasattr(attn, 'reset_parameters'): if hasattr(attn, 'reset_parameters'):
attn.reset_parameters() attn.reset_parameters()
def forward(self, x): def forward(self, x):
shortcut = self.shortcut(x) shortcut = x
x = self.conv1_kxk(x) x = self.conv1_kxk(x)
x = self.attn(x) x = self.attn(x)
x = self.conv2_1x1(x) x = self.conv2_1x1(x)
x = self.attn_last(x) x = self.attn_last(x)
x = self.drop_path(x) x = self.drop_path(x)
x = self.act(x + shortcut) if self.shortcut is not None:
return x x = x + self.shortcut(shortcut)
return self.act(x)
class RepVggBlock(nn.Module): class RepVggBlock(nn.Module):
@ -1065,8 +1054,7 @@ class RepVggBlock(nn.Module):
x = self.drop_path(x) # not in the paper / official impl, experimental x = self.drop_path(x) # not in the paper / official impl, experimental
x = x + identity x = x + identity
x = self.attn(x) # no attn in the paper / official impl, experimental x = self.attn(x) # no attn in the paper / official impl, experimental
x = self.act(x) return self.act(x)
return x
class SelfAttnBlock(nn.Module): class SelfAttnBlock(nn.Module):
@ -1074,19 +1062,16 @@ class SelfAttnBlock(nn.Module):
""" """
def __init__(self, in_chs, out_chs, kernel_size=3, stride=1, dilation=(1, 1), bottle_ratio=1., group_size=None, def __init__(self, in_chs, out_chs, kernel_size=3, stride=1, dilation=(1, 1), bottle_ratio=1., group_size=None,
downsample='avg', extra_conv=False, linear_out=False, post_attn_na=True, feat_size=None, downsample='avg', extra_conv=False, linear_out=False, bottle_in=False, post_attn_na=True,
layers: LayerFn = None, drop_block=None, drop_path_rate=0.): feat_size=None, layers: LayerFn = None, drop_block=None, drop_path_rate=0.):
super(SelfAttnBlock, self).__init__() super(SelfAttnBlock, self).__init__()
assert layers is not None assert layers is not None
mid_chs = make_divisible(out_chs * bottle_ratio) mid_chs = make_divisible((in_chs if bottle_in else out_chs) * bottle_ratio)
groups = num_groups(group_size, mid_chs) groups = num_groups(group_size, mid_chs)
if in_chs != out_chs or stride != 1 or dilation[0] != dilation[1]: self.shortcut = create_shortcut(
self.shortcut = create_downsample( downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation,
downsample, in_chs=in_chs, out_chs=out_chs, stride=stride, dilation=dilation[0], apply_act=False, layers=layers)
apply_act=False, layers=layers)
else:
self.shortcut = nn.Identity()
self.conv1_1x1 = layers.conv_norm_act(in_chs, mid_chs, 1) self.conv1_1x1 = layers.conv_norm_act(in_chs, mid_chs, 1)
if extra_conv: if extra_conv:
@ -1105,7 +1090,7 @@ class SelfAttnBlock(nn.Module):
self.act = nn.Identity() if linear_out else layers.act(inplace=True) self.act = nn.Identity() if linear_out else layers.act(inplace=True)
def init_weights(self, zero_init_last: bool = False): def init_weights(self, zero_init_last: bool = False):
if zero_init_last: if zero_init_last and self.shortcut is not None:
nn.init.zeros_(self.conv3_1x1.bn.weight) nn.init.zeros_(self.conv3_1x1.bn.weight)
if hasattr(self.self_attn, 'reset_parameters'): if hasattr(self.self_attn, 'reset_parameters'):
self.self_attn.reset_parameters() self.self_attn.reset_parameters()

Write Preview
Loading…
Cancel
Save