Skip to content

[WIP] Add ResNet-RS models #554

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 22 commits into from
May 4, 2021
Merged

[WIP] Add ResNet-RS models #554

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
22 commits
Select commit Hold shift + click to select a range
8449932
Add ResNet-RS models
amaarora Apr 10, 2021
b117e16
Only include resnet-rs changes
amaarora Apr 12, 2021
206407a
remove whitespace diff
amaarora Apr 12, 2021
3f12ad3
remove whitespace diff
amaarora Apr 12, 2021
ca2877f
EOF newline
amaarora Apr 12, 2021
3df71a1
Update time
amaarora Apr 12, 2021
823162c
increase time
amaarora Apr 12, 2021
530f8dd
Add first conv
amaarora Apr 12, 2021
973b880
Try running only resnetv2_101x1_bitm on Linux runner
amaarora Apr 12, 2021
bf63f0c
Add to exclude filter
amaarora Apr 12, 2021
094f04d
Run test_model_forward_features for all
amaarora Apr 12, 2021
f89bf2c
Add to exclude ftrs
amaarora Apr 12, 2021
a17ce02
back to defaults
amaarora Apr 12, 2021
84fd045
only run test_forward_features
amaarora Apr 12, 2021
6076728
run all tests
amaarora Apr 12, 2021
59ca945
Run all tests
amaarora Apr 12, 2021
3a73401
Add bigger resnetrs to model filters to fix Github CLI
amaarora Apr 12, 2021
ea9f935
Remove resnetv2_101x1_bitm from exclude feat features
amaarora Apr 12, 2021
e7df94a
Merge branch 'master' into resnet-rs
amaarora May 2, 2021
69f8c71
Remove hardcoded values
amaarora May 2, 2021
c7b4041
Make sure reduction ratio in resnetrs is 0.25
amaarora May 3, 2021
1f45c6e
There is no bias in replaced maxpool so remove it
amaarora May 3, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 3 additions & 2 deletions tests/test_models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,8 +22,9 @@
if 'GITHUB_ACTIONS' in os.environ: # and 'Linux' in platform.system():
# GitHub Linux runner is slower and hits memory limits sooner than MacOS, exclude bigger models
EXCLUDE_FILTERS = [
'*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm', '*101x3_bitm',
'*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*'] + NON_STD_FILTERS
'*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm',
'*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*',
'*resnetrs200*', '*resnetrs270*', '*resnetrs350*', '*resnetrs420*'] + NON_STD_FILTERS
else:
EXCLUDE_FILTERS = NON_STD_FILTERS

Expand Down
99 changes: 90 additions & 9 deletions timm/models/resnet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -236,7 +236,23 @@ def _cfg(url='', **kwargs):
interpolation='bicubic'),
'resnetblur50': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnetblur50-84f4748f.pth',
interpolation='bicubic')
interpolation='bicubic'),

# ResNet-RS models
'resnetrs50': _cfg(
interpolation='bicubic', first_conv='conv1.0'),
'resnetrs101': _cfg(
interpolation='bicubic', first_conv='conv1.0'),
'resnetrs152': _cfg(
interpolation='bicubic', first_conv='conv1.0'),
'resnetrs200': _cfg(
interpolation='bicubic', first_conv='conv1.0'),
'resnetrs270': _cfg(
interpolation='bicubic', first_conv='conv1.0'),
'resnetrs350': _cfg(
interpolation='bicubic', first_conv='conv1.0'),
'resnetrs420': _cfg(
interpolation='bicubic', first_conv='conv1.0'),
}


Expand Down Expand Up @@ -318,7 +334,7 @@ class Bottleneck(nn.Module):

def __init__(self, inplanes, planes, stride=1, downsample=None, cardinality=1, base_width=64,
reduce_first=1, dilation=1, first_dilation=None, act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d,
attn_layer=None, aa_layer=None, drop_block=None, drop_path=None):
attn_layer=None, aa_layer=None, drop_block=None, drop_path=None, **kwargs):
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@rwightman FYI this is change that you might want to review specifically. I have added **kwargs to Bottleneck block that get passed to attention layers.

This is to pass in reduction_ratio=0.25 for se layers as mentioned in the paper.

super(Bottleneck, self).__init__()

width = int(math.floor(planes * (base_width / 64)) * cardinality)
Expand All @@ -341,7 +357,7 @@ def __init__(self, inplanes, planes, stride=1, downsample=None, cardinality=1, b
self.conv3 = nn.Conv2d(width, outplanes, kernel_size=1, bias=False)
self.bn3 = norm_layer(outplanes)

self.se = create_attn(attn_layer, outplanes)
self.se = create_attn(attn_layer, outplanes, **kwargs)
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These **kwargs get passed to Bottleneck using block_args in model_config as in L1112 model definition for resnetrs152


self.act3 = act_layer(inplace=True)
self.downsample = downsample
Expand Down Expand Up @@ -545,11 +561,12 @@ def __init__(self, block, layers, num_classes=1000, in_chans=3,
cardinality=1, base_width=64, stem_width=64, stem_type='',
output_stride=32, block_reduce_first=1, down_kernel_size=1, avg_down=False,
act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d, aa_layer=None, drop_rate=0.0, drop_path_rate=0.,
drop_block_rate=0., global_pool='avg', zero_init_last_bn=True, block_args=None):
drop_block_rate=0., global_pool='avg', zero_init_last_bn=True, block_args=None, replace_stem_max_pool=False):
block_args = block_args or dict()
assert output_stride in (8, 16, 32)
self.num_classes = num_classes
self.drop_rate = drop_rate
self.replace_stem_max_pool = replace_stem_max_pool
super(ResNet, self).__init__()

# Stem
Expand All @@ -574,12 +591,19 @@ def __init__(self, block, layers, num_classes=1000, in_chans=3,
self.feature_info = [dict(num_chs=inplanes, reduction=2, module='act1')]

# Stem Pooling
if aa_layer is not None:
self.maxpool = nn.Sequential(*[
nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
aa_layer(channels=inplanes, stride=2)])
if not self.replace_stem_max_pool:
if aa_layer is not None:
self.maxpool = nn.Sequential(*[
nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
aa_layer(channels=inplanes, stride=2)])
else:
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
else:
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.maxpool = nn.Sequential(*[
nn.Conv2d(inplanes, inplanes, 3, stride=2, padding=1, bias=False),
norm_layer(inplanes),
act_layer(inplace=True)
])

# Feature Blocks
channels = [64, 128, 256, 512]
Expand Down Expand Up @@ -1065,6 +1089,63 @@ def ecaresnet50d(pretrained=False, **kwargs):
return _create_resnet('ecaresnet50d', pretrained, **model_args)


@register_model
def resnetrs50(pretrained=False, **kwargs):
model_args = dict(
block=Bottleneck, layers=[3, 4, 6, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
avg_down=True, block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
return _create_resnet('resnetrs50', pretrained, **model_args)


@register_model
def resnetrs101(pretrained=False, **kwargs):
model_args = dict(
block=Bottleneck, layers=[3, 4, 23, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
avg_down=True, block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
return _create_resnet('resnetrs101', pretrained, **model_args)


@register_model
def resnetrs152(pretrained=False, **kwargs):
model_args = dict(
block=Bottleneck, layers=[3, 8, 36, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
avg_down=True, block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
return _create_resnet('resnetrs152', pretrained, **model_args)


@register_model
def resnetrs200(pretrained=False, **kwargs):
model_args = dict(
block=Bottleneck, layers=[3, 24, 36, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
avg_down=True, block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
return _create_resnet('resnetrs200', pretrained, **model_args)


@register_model
def resnetrs270(pretrained=False, **kwargs):
model_args = dict(
block=Bottleneck, layers=[4, 29, 53, 4], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
avg_down=True, block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
return _create_resnet('resnetrs270', pretrained, **model_args)



@register_model
def resnetrs350(pretrained=False, **kwargs):
model_args = dict(
block=Bottleneck, layers=[4, 36, 72, 4], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
avg_down=True, block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
return _create_resnet('resnetrs350', pretrained, **model_args)


@register_model
def resnetrs420(pretrained=False, **kwargs):
model_args = dict(
block=Bottleneck, layers=[4, 44, 87, 4], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
avg_down=True, block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
return _create_resnet('resnetrs420', pretrained, **model_args)


@register_model
def ecaresnet50d_pruned(pretrained=False, **kwargs):
"""Constructs a ResNet-50-D model pruned with eca.
Expand Down