Skip to content

Commit 6c7932f

Browse files
rwightmanrwightman
committed
Update sotabench.py, tweak VovNet cfg
1 parent d3ee3de commit 6c7932f

File tree

2 files changed

+83
-46
lines changed

2 files changed

+83
-46
lines changed

‎sotabench.py‎

Lines changed: 34 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -135,6 +135,12 @@ def _entry(model_name, paper_model_name, paper_arxiv_id, batch_size=BATCH_SIZE,
135135

136136
_entry('resnetblur50', 'ResNet-Blur-50', '1904.11486'),
137137

138+
_entry('densenet121', 'DenseNet-121', '1608.06993'),
139+
_entry('densenetblur121d', 'DenseNet-Blur-121D', '1904.11486',
140+
model_desc='DenseNet with blur pooling and deep stem'),
141+
142+
_entry('ese_vovnet39b', 'VoVNet-39-V2', '1911.06667'),
143+
138144
_entry('tf_efficientnet_b0', 'EfficientNet-B0 (AutoAugment)', '1905.11946',
139145
model_desc='Ported from official Google AI Tensorflow weights'),
140146
_entry('tf_efficientnet_b1', 'EfficientNet-B1 (AutoAugment)', '1905.11946',
@@ -389,6 +395,34 @@ def _entry(model_name, paper_model_name, paper_arxiv_id, batch_size=BATCH_SIZE,
389395
model_desc='Originally from https://github.com/mehtadushy/SelecSLS-Pytorch'),
390396
_entry('selecsls60b', 'SelecSLS-60_B', '1907.00837',
391397
model_desc='Originally from https://github.com/mehtadushy/SelecSLS-Pytorch'),
398+
399+
## RegNet official impl weighs
400+
_entry('regnetx_002', 'RegNetX-200MF', '2003.13678'),
401+
_entry('regnetx_004', 'RegNetX-400MF', '2003.13678'),
402+
_entry('regnetx_006', 'RegNetX-600MF', '2003.13678'),
403+
_entry('regnetx_008', 'RegNetX-800MF', '2003.13678'),
404+
_entry('regnetx_016', 'RegNetX-1.6GF', '2003.13678'),
405+
_entry('regnetx_032', 'RegNetX-3.2GF', '2003.13678'),
406+
_entry('regnetx_040', 'RegNetX-4.0GF', '2003.13678'),
407+
_entry('regnetx_064', 'RegNetX-6.4GF', '2003.13678'),
408+
_entry('regnetx_080', 'RegNetX-8.0GF', '2003.13678'),
409+
_entry('regnetx_120', 'RegNetX-12GF', '2003.13678'),
410+
_entry('regnetx_160', 'RegNetX-16GF', '2003.13678'),
411+
_entry('regnetx_320', 'RegNetX-32GF', '2003.13678', batch_size=BATCH_SIZE // 2),
412+
413+
_entry('regnety_002', 'RegNetY-200MF', '2003.13678'),
414+
_entry('regnety_004', 'RegNetY-400MF', '2003.13678'),
415+
_entry('regnety_006', 'RegNetY-600MF', '2003.13678'),
416+
_entry('regnety_008', 'RegNetY-800MF', '2003.13678'),
417+
_entry('regnety_016', 'RegNetY-1.6GF', '2003.13678'),
418+
_entry('regnety_032', 'RegNetY-3.2GF', '2003.13678'),
419+
_entry('regnety_040', 'RegNetY-4.0GF', '2003.13678'),
420+
_entry('regnety_064', 'RegNetY-6.4GF', '2003.13678'),
421+
_entry('regnety_080', 'RegNetY-8.0GF', '2003.13678'),
422+
_entry('regnety_120', 'RegNetY-12GF', '2003.13678'),
423+
_entry('regnety_160', 'RegNetY-16GF', '2003.13678'),
424+
_entry('regnety_320', 'RegNetY-32GF', '2003.13678', batch_size=BATCH_SIZE // 2),
425+
392426
]
393427

394428
for m in model_list:

‎timm/models/vovnet.py‎

Lines changed: 49 additions & 46 deletions
Original file line numberDiff line numberDiff line change
@@ -28,19 +28,19 @@
2828
# https://github.com/stigma0617/VoVNet.pytorch/blob/master/models_vovnet/vovnet.py
2929
model_cfgs = dict(
3030
vovnet39a=dict(
31-
stem_ch=[64, 64, 128],
32-
stage_conv_ch=[128, 160, 192, 224],
33-
stage_out_ch=[256, 512, 768, 1024],
31+
stem_chs=[64, 64, 128],
32+
stage_conv_chs=[128, 160, 192, 224],
33+
stage_out_chs=[256, 512, 768, 1024],
3434
layer_per_block=5,
3535
block_per_stage=[1, 1, 2, 2],
3636
residual=False,
3737
depthwise=False,
3838
attn='',
3939
),
4040
vovnet57a=dict(
41-
stem_ch=[64, 64, 128],
42-
stage_conv_ch=[128, 160, 192, 224],
43-
stage_out_ch=[256, 512, 768, 1024],
41+
stem_chs=[64, 64, 128],
42+
stage_conv_chs=[128, 160, 192, 224],
43+
stage_out_chs=[256, 512, 768, 1024],
4444
layer_per_block=5,
4545
block_per_stage=[1, 1, 4, 3],
4646
residual=False,
@@ -49,9 +49,9 @@
4949

5050
),
5151
ese_vovnet19b_slim_dw=dict(
52-
stem_ch=[64, 64, 64],
53-
stage_conv_ch=[64, 80, 96, 112],
54-
stage_out_ch=[112, 256, 384, 512],
52+
stem_chs=[64, 64, 64],
53+
stage_conv_chs=[64, 80, 96, 112],
54+
stage_out_chs=[112, 256, 384, 512],
5555
layer_per_block=3,
5656
block_per_stage=[1, 1, 1, 1],
5757
residual=True,
@@ -60,29 +60,29 @@
6060

6161
),
6262
ese_vovnet19b_dw=dict(
63-
stem_ch=[64, 64, 64],
64-
stage_conv_ch=[128, 160, 192, 224],
65-
stage_out_ch=[256, 512, 768, 1024],
63+
stem_chs=[64, 64, 64],
64+
stage_conv_chs=[128, 160, 192, 224],
65+
stage_out_chs=[256, 512, 768, 1024],
6666
layer_per_block=3,
6767
block_per_stage=[1, 1, 1, 1],
6868
residual=True,
6969
depthwise=True,
7070
attn='ese',
7171
),
7272
ese_vovnet19b_slim=dict(
73-
stem_ch=[64, 64, 128],
74-
stage_conv_ch=[64, 80, 96, 112],
75-
stage_out_ch=[112, 256, 384, 512],
73+
stem_chs=[64, 64, 128],
74+
stage_conv_chs=[64, 80, 96, 112],
75+
stage_out_chs=[112, 256, 384, 512],
7676
layer_per_block=3,
7777
block_per_stage=[1, 1, 1, 1],
7878
residual=True,
7979
depthwise=False,
8080
attn='ese',
8181
),
8282
ese_vovnet19b=dict(
83-
stem_ch=[64, 64, 128],
84-
stage_conv_ch=[128, 160, 192, 224],
85-
stage_out_ch=[256, 512, 768, 1024],
83+
stem_chs=[64, 64, 128],
84+
stage_conv_chs=[128, 160, 192, 224],
85+
stage_out_chs=[256, 512, 768, 1024],
8686
layer_per_block=3,
8787
block_per_stage=[1, 1, 1, 1],
8888
residual=True,
@@ -91,19 +91,19 @@
9191

9292
),
9393
ese_vovnet39b=dict(
94-
stem_ch=[64, 64, 128],
95-
stage_conv_ch=[128, 160, 192, 224],
96-
stage_out_ch=[256, 512, 768, 1024],
94+
stem_chs=[64, 64, 128],
95+
stage_conv_chs=[128, 160, 192, 224],
96+
stage_out_chs=[256, 512, 768, 1024],
9797
layer_per_block=5,
9898
block_per_stage=[1, 1, 2, 2],
9999
residual=True,
100100
depthwise=False,
101101
attn='ese',
102102
),
103103
ese_vovnet57b=dict(
104-
stem_ch=[64, 64, 128],
105-
stage_conv_ch=[128, 160, 192, 224],
106-
stage_out_ch=[256, 512, 768, 1024],
104+
stem_chs=[64, 64, 128],
105+
stage_conv_chs=[128, 160, 192, 224],
106+
stage_out_chs=[256, 512, 768, 1024],
107107
layer_per_block=5,
108108
block_per_stage=[1, 1, 4, 3],
109109
residual=True,
@@ -112,26 +112,28 @@
112112

113113
),
114114
ese_vovnet99b=dict(
115-
stem_ch=[64, 64, 128],
116-
stage_conv_ch=[128, 160, 192, 224],
117-
stage_out_ch=[256, 512, 768, 1024],
115+
stem_chs=[64, 64, 128],
116+
stage_conv_chs=[128, 160, 192, 224],
117+
stage_out_chs=[256, 512, 768, 1024],
118118
layer_per_block=5,
119119
block_per_stage=[1, 3, 9, 3],
120120
residual=True,
121121
depthwise=False,
122122
attn='ese',
123123
),
124124
eca_vovnet39b=dict(
125-
stem_ch=[64, 64, 128],
126-
stage_conv_ch=[128, 160, 192, 224],
127-
stage_out_ch=[256, 512, 768, 1024],
125+
stem_chs=[64, 64, 128],
126+
stage_conv_chs=[128, 160, 192, 224],
127+
stage_out_chs=[256, 512, 768, 1024],
128128
layer_per_block=5,
129129
block_per_stage=[1, 1, 2, 2],
130130
residual=True,
131131
depthwise=False,
132132
attn='eca',
133133
),
134134
)
135+
model_cfgs['ese_vovnet39b_evos'] = model_cfgs['ese_vovnet39b']
136+
model_cfgs['ese_vovnet99b_iabn'] = model_cfgs['ese_vovnet99b']
135137

136138

137139
def _cfg(url=''):
@@ -154,6 +156,8 @@ def _cfg(url=''):
154156
ese_vovnet57b=_cfg(url=''),
155157
ese_vovnet99b=_cfg(url=''),
156158
eca_vovnet39b=_cfg(url=''),
159+
ese_vovnet39b_evos=_cfg(url=''),
160+
eee_vovnet99b_iabn=_cfg(url=''),
157161
)
158162

159163

@@ -277,33 +281,33 @@ def __init__(self, cfg, in_chans=3, num_classes=1000, global_pool='avg', drop_ra
277281
self.drop_rate = drop_rate
278282
assert stem_stride in (4, 2)
279283

280-
stem_ch = cfg["stem_ch"]
281-
stage_conv_ch = cfg["stage_conv_ch"]
282-
stage_out_ch = cfg["stage_out_ch"]
284+
stem_chs = cfg["stem_chs"]
285+
stage_conv_chs = cfg["stage_conv_chs"]
286+
stage_out_chs = cfg["stage_out_chs"]
283287
block_per_stage = cfg["block_per_stage"]
284288
layer_per_block = cfg["layer_per_block"]
285289

286290
# Stem module
287291
last_stem_stride = stem_stride // 2
288292
conv_type = SeparableConvBnAct if cfg["depthwise"] else ConvBnAct
289293
self.stem = nn.Sequential(*[
290-
ConvBnAct(in_chans, stem_ch[0], 3, stride=2, norm_layer=norm_layer),
291-
conv_type(stem_ch[0], stem_ch[1], 3, stride=1, norm_layer=norm_layer),
292-
conv_type(stem_ch[1], stem_ch[2], 3, stride=last_stem_stride, norm_layer=norm_layer),
294+
ConvBnAct(in_chans, stem_chs[0], 3, stride=2, norm_layer=norm_layer),
295+
conv_type(stem_chs[0], stem_chs[1], 3, stride=1, norm_layer=norm_layer),
296+
conv_type(stem_chs[1], stem_chs[2], 3, stride=last_stem_stride, norm_layer=norm_layer),
293297
])
294298

295299
# OSA stages
296-
in_ch_list = stem_ch[-1:] + stage_out_ch[:-1]
300+
in_ch_list = stem_chs[-1:] + stage_out_chs[:-1]
297301
stage_args = dict(
298302
residual=cfg["residual"], depthwise=cfg["depthwise"], attn=cfg["attn"], norm_layer=norm_layer)
299303
stages = []
300304
for i in range(4): # num_stages
301305
downsample = stem_stride == 2 or i > 0 # first stage has no stride/downsample if stem_stride is 4
302306
stages += [OsaStage(
303-
in_ch_list[i], stage_conv_ch[i], stage_out_ch[i], block_per_stage[i], layer_per_block,
307+
in_ch_list[i], stage_conv_chs[i], stage_out_chs[i], block_per_stage[i], layer_per_block,
304308
downsample=downsample, **stage_args)
305309
]
306-
self.num_features = stage_out_ch[i]
310+
self.num_features = stage_out_chs[i]
307311
self.stages = nn.Sequential(*stages)
308312

309313
self.head = ClassifierHead(self.num_features, num_classes, pool_type=global_pool, drop_rate=drop_rate)
@@ -398,14 +402,13 @@ def eca_vovnet39b(pretrained=False, **kwargs):
398402

399403
# Experimental Models
400404

401-
@register_model
402-
def ese_vovnet39b_iabn(pretrained=False, **kwargs):
403-
norm_layer = get_norm_act_layer('iabn')
404-
return _vovnet('ese_vovnet39b', pretrained=pretrained, norm_layer=norm_layer, **kwargs)
405-
406-
407405
@register_model
408406
def ese_vovnet39b_evos(pretrained=False, **kwargs):
409407
def norm_act_fn(num_features, **kwargs):
410408
return create_norm_act('EvoNormSample', num_features, jit=False, **kwargs)
411-
return _vovnet('ese_vovnet39b', pretrained=pretrained, norm_layer=norm_act_fn, **kwargs)
409+
return _vovnet('ese_vovnet39b_evos', pretrained=pretrained, norm_layer=norm_act_fn, **kwargs)
410+
411+
@register_model
412+
def ese_vovnet99b_iabn(pretrained=False, **kwargs):
413+
norm_layer = get_norm_act_layer('iabn')
414+
return _vovnet('ese_vovnet99b_iabn', pretrained=pretrained, norm_layer=norm_layer, **kwargs)

0 commit comments

Comments
 (0)