Skip to content

Commit daeaa11

Browse files
rwightmanrwightman
committed
Add initial sotabench attempt. Split create_transform out of create_loader. Update requirements.txt
1 parent aeaaad7 commit daeaa11

File tree

4 files changed

+254
-16
lines changed

4 files changed

+254
-16
lines changed

‎requirements.txt‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,2 +1,2 @@
1-
torch~=1.0
2-
torchvision
1+
torch>=1.1.0
2+
torchvision>=0.3.0

‎sotabench.py‎

Lines changed: 214 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,214 @@
1+
from torchbench.image_classification import ImageNet
2+
from timm import create_model, list_models
3+
from timm.data import resolve_data_config, create_transform
4+
5+
NUM_GPU = 1
6+
BATCH_SIZE = 256 * NUM_GPU
7+
8+
9+
def _attrib(paper_model_name='', paper_arxiv_id='', batch_size=BATCH_SIZE):
10+
return dict(
11+
paper_model_name=paper_model_name,
12+
paper_arxiv_id=paper_arxiv_id,
13+
batch_size=batch_size)
14+
15+
model_map = dict(
16+
#adv_inception_v3=_attrib(paper_model_name='Adversarial Inception V3', paper_arxiv_id=),
17+
#densenet121=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
18+
#densenet161=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
19+
#densenet169=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
20+
#densenet201=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
21+
dpn68=_attrib(
22+
paper_model_name='DPN-68', paper_arxiv_id='1707.01629'),
23+
dpn68b=_attrib(
24+
paper_model_name='DPN-68b', paper_arxiv_id='1707.01629'),
25+
dpn92=_attrib(
26+
paper_model_name='DPN-92', paper_arxiv_id='1707.01629'),
27+
dpn98=_attrib(
28+
paper_model_name='DPN-98', paper_arxiv_id='1707.01629'),
29+
dpn107=_attrib(
30+
paper_model_name='DPN-107', paper_arxiv_id='1707.01629'),
31+
dpn131=_attrib(
32+
paper_model_name='DPN-131', paper_arxiv_id='1707.01629'),
33+
efficientnet_b0=_attrib(
34+
paper_model_name='EfficientNet-B0', paper_arxiv_id='1905.11946'),
35+
efficientnet_b1=_attrib(
36+
paper_model_name='EfficientNet-B1', paper_arxiv_id='1905.11946'),
37+
efficientnet_b2=_attrib(
38+
paper_model_name='EfficientNet-B2', paper_arxiv_id='1905.11946'),
39+
#ens_adv_inception_resnet_v2=_attrib(paper_model_name=, paper_arxiv_id=),
40+
fbnetc_100=_attrib(
41+
paper_model_name='FBNet-C', paper_arxiv_id='1812.03443'),
42+
gluon_inception_v3=_attrib(
43+
paper_model_name='Inception V3', paper_arxiv_id='1512.00567'),
44+
gluon_resnet18_v1b=_attrib(
45+
paper_model_name='ResNet-18', paper_arxiv_id='1812.01187'),
46+
gluon_resnet34_v1b=_attrib(
47+
paper_model_name='ResNet-34', paper_arxiv_id='1812.01187'),
48+
gluon_resnet50_v1b=_attrib(
49+
paper_model_name='ResNet-50', paper_arxiv_id='1812.01187'),
50+
gluon_resnet50_v1c=_attrib(
51+
paper_model_name='ResNet-50-C', paper_arxiv_id='1812.01187'),
52+
gluon_resnet50_v1d=_attrib(
53+
paper_model_name='ResNet-50-D', paper_arxiv_id='1812.01187'),
54+
gluon_resnet50_v1s=_attrib(
55+
paper_model_name='ResNet-50-S', paper_arxiv_id='1812.01187'),
56+
gluon_resnet101_v1b=_attrib(
57+
paper_model_name='ResNet-101', paper_arxiv_id='1812.01187'),
58+
gluon_resnet101_v1c=_attrib(
59+
paper_model_name='ResNet-101-C', paper_arxiv_id='1812.01187'),
60+
gluon_resnet101_v1d=_attrib(
61+
paper_model_name='ResNet-101-D', paper_arxiv_id='1812.01187'),
62+
gluon_resnet101_v1s=_attrib(
63+
paper_model_name='ResNet-101-S', paper_arxiv_id='1812.01187'),
64+
gluon_resnet152_v1b=_attrib(
65+
paper_model_name='ResNet-152', paper_arxiv_id='1812.01187'),
66+
gluon_resnet152_v1c=_attrib(
67+
paper_model_name='ResNet-152-C', paper_arxiv_id='1812.01187'),
68+
gluon_resnet152_v1d=_attrib(
69+
paper_model_name='ResNet-152-D', paper_arxiv_id='1812.01187'),
70+
gluon_resnet152_v1s=_attrib(
71+
paper_model_name='ResNet-152-S', paper_arxiv_id='1812.01187'),
72+
gluon_resnext50_32x4d=_attrib(
73+
paper_model_name='ResNeXt-50 32x4d', paper_arxiv_id='1812.01187'),
74+
gluon_resnext101_32x4d=_attrib(
75+
paper_model_name='ResNeXt-101 32x4d', paper_arxiv_id='1812.01187'),
76+
gluon_resnext101_64x4d=_attrib(
77+
paper_model_name='ResNeXt-101 64x4d', paper_arxiv_id='1812.01187'),
78+
gluon_senet154=_attrib(
79+
paper_model_name='SENet-154', paper_arxiv_id='1812.01187'),
80+
gluon_seresnext50_32x4d=_attrib(
81+
paper_model_name='SE-ResNeXt-50 32x4d', paper_arxiv_id='1812.01187'),
82+
gluon_seresnext101_32x4d=_attrib(
83+
paper_model_name='SE-ResNeXt-101 32x4d', paper_arxiv_id='1812.01187'),
84+
gluon_seresnext101_64x4d=_attrib(
85+
paper_model_name='SE-ResNeXt-101 64x4d', paper_arxiv_id='1812.01187'),
86+
gluon_xception65=_attrib(
87+
paper_model_name='Modified Aligned Xception', paper_arxiv_id='1802.02611', batch_size=BATCH_SIZE//2),
88+
ig_resnext101_32x8d=_attrib(
89+
paper_model_name='ResNeXt-101 32×8d', paper_arxiv_id='1805.00932'),
90+
ig_resnext101_32x16d=_attrib(
91+
paper_model_name='ResNeXt-101 32×16d', paper_arxiv_id='1805.00932'),
92+
ig_resnext101_32x32d=_attrib(
93+
paper_model_name='ResNeXt-101 32×32d', paper_arxiv_id='1805.00932', batch_size=BATCH_SIZE//2),
94+
ig_resnext101_32x48d=_attrib(
95+
paper_model_name='ResNeXt-101 32×48d', paper_arxiv_id='1805.00932', batch_size=BATCH_SIZE//4),
96+
inception_resnet_v2=_attrib(
97+
paper_model_name='Inception ResNet V2', paper_arxiv_id='1602.07261'),
98+
#inception_v3=dict(paper_model_name='Inception V3', paper_arxiv_id=), # same weights as torchvision
99+
inception_v4=_attrib(
100+
paper_model_name='Inception V4', paper_arxiv_id='1602.07261'),
101+
mixnet_l=_attrib(
102+
paper_model_name='MixNet-L', paper_arxiv_id='1907.09595'),
103+
mixnet_m=_attrib(
104+
paper_model_name='MixNet-M', paper_arxiv_id='1907.09595'),
105+
mixnet_s=_attrib(
106+
paper_model_name='MixNet-S', paper_arxiv_id='1907.09595'),
107+
mnasnet_100=_attrib(
108+
paper_model_name='MnasNet-B1', paper_arxiv_id='1807.11626'),
109+
mobilenetv3_100=_attrib(
110+
paper_model_name='MobileNet V3(1.0)', paper_arxiv_id='1905.02244'),
111+
nasnetalarge=_attrib(
112+
paper_model_name='NASNet-A Large', paper_arxiv_id='1707.07012', batch_size=BATCH_SIZE//4),
113+
pnasnet5large=_attrib(
114+
paper_model_name='PNASNet-5', paper_arxiv_id='1712.00559', batch_size=BATCH_SIZE//4),
115+
resnet18=_attrib(
116+
paper_model_name='ResNet-18', paper_arxiv_id='1812.01187'),
117+
resnet26=_attrib(
118+
paper_model_name='ResNet-26', paper_arxiv_id='1812.01187'),
119+
resnet26d=_attrib(
120+
paper_model_name='ResNet-26-D', paper_arxiv_id='1812.01187'),
121+
resnet34=_attrib(
122+
paper_model_name='ResNet-34', paper_arxiv_id='1812.01187'),
123+
resnet50=_attrib(
124+
paper_model_name='ResNet-50', paper_arxiv_id='1812.01187'),
125+
#resnet101=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
126+
#resnet152=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
127+
resnext50_32x4d=_attrib(
128+
paper_model_name='ResNeXt-50 32x4d', paper_arxiv_id='1812.01187'),
129+
resnext50d_32x4d=_attrib(
130+
paper_model_name='ResNeXt-50-D 32x4d', paper_arxiv_id='1812.01187'),
131+
#resnext101_32x8d=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
132+
semnasnet_100=_attrib(
133+
paper_model_name='MnasNet-A1', paper_arxiv_id='1807.11626'),
134+
senet154=_attrib(
135+
paper_model_name='SENet-154', paper_arxiv_id='1709.01507'),
136+
seresnet18=_attrib(
137+
paper_model_name='SE-ResNet-18', paper_arxiv_id='1709.01507'),
138+
seresnet34=_attrib(
139+
paper_model_name='SE-ResNet-34', paper_arxiv_id='1709.01507'),
140+
seresnet50=_attrib(
141+
paper_model_name='SE-ResNet-50', paper_arxiv_id='1709.01507'),
142+
seresnet101=_attrib(
143+
paper_model_name='SE-ResNet-101', paper_arxiv_id='1709.01507'),
144+
seresnet152=_attrib(
145+
paper_model_name='SE-ResNet-152', paper_arxiv_id='1709.01507'),
146+
seresnext26_32x4d=_attrib(
147+
paper_model_name='SE-ResNeXt-26 32x4d', paper_arxiv_id='1709.01507'),
148+
seresnext50_32x4d=_attrib(
149+
paper_model_name='SE-ResNeXt-50 32x4d', paper_arxiv_id='1709.01507'),
150+
seresnext101_32x4d=_attrib(
151+
paper_model_name='SE-ResNeXt-101 32x4d', paper_arxiv_id='1709.01507'),
152+
spnasnet_100=_attrib(
153+
paper_model_name='Single-Path NAS', paper_arxiv_id='1904.02877'),
154+
tf_efficientnet_b0=_attrib(
155+
paper_model_name='EfficientNet-B0', paper_arxiv_id='1905.11946'),
156+
tf_efficientnet_b1=_attrib(
157+
paper_model_name='EfficientNet-B1', paper_arxiv_id='1905.11946'),
158+
tf_efficientnet_b2=_attrib(
159+
paper_model_name='EfficientNet-B2', paper_arxiv_id='1905.11946'),
160+
tf_efficientnet_b3=_attrib(
161+
paper_model_name='EfficientNet-B3', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//2),
162+
tf_efficientnet_b4=_attrib(
163+
paper_model_name='EfficientNet-B4', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//2),
164+
tf_efficientnet_b5=_attrib(
165+
paper_model_name='EfficientNet-B5', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//4),
166+
tf_efficientnet_b6=_attrib(
167+
paper_model_name='EfficientNet-B6', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//8),
168+
tf_efficientnet_b7=_attrib(
169+
paper_model_name='EfficientNet-B7', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//8),
170+
tf_inception_v3=_attrib(
171+
paper_model_name='Inception V3', paper_arxiv_id='1512.00567'),
172+
tf_mixnet_l=_attrib(
173+
paper_model_name='MixNet-L', paper_arxiv_id='1907.09595'),
174+
tf_mixnet_m=_attrib(
175+
paper_model_name='MixNet-M', paper_arxiv_id='1907.09595'),
176+
tf_mixnet_s=_attrib(
177+
paper_model_name='MixNet-S', paper_arxiv_id='1907.09595'),
178+
#tv_resnet34=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
179+
#tv_resnet50=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
180+
#tv_resnext50_32x4d=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
181+
#wide_resnet50_2=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
182+
#wide_resnet101_2=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
183+
xception=_attrib(
184+
paper_model_name='Xception', paper_arxiv_id='1610.02357'),
185+
)
186+
187+
model_names = list_models(pretrained=True)
188+
189+
for model_name in model_names:
190+
if model_name not in model_map:
191+
print('Skipping %s' % model_name)
192+
continue
193+
194+
# create model from name
195+
model = create_model(model_name, pretrained=True)
196+
param_count = sum([m.numel() for m in model.parameters()])
197+
print('Model %s created, param count: %d' % (model_name, param_count))
198+
199+
# get appropriate transform for model's default pretrained config
200+
data_config = resolve_data_config(dict(), model=model, verbose=True)
201+
input_transform = create_transform(**data_config)
202+
203+
# Run the benchmark
204+
ImageNet.benchmark(
205+
model=model,
206+
paper_model_name=model_map[model_name]['paper_model_name'],
207+
paper_arxiv_id=model_map[model_name]['paper_arxiv_id'],
208+
input_transform=input_transform,
209+
batch_size=model_map[model_name]['batch_size'],
210+
num_gpu=NUM_GPU,
211+
#data_root=DATA_ROOT
212+
)
213+
214+

‎timm/data/__init__.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2,5 +2,5 @@
22
from .config import resolve_data_config
33
from .dataset import Dataset, DatasetTar
44
from .transforms import *
5-
from .loader import create_loader
5+
from .loader import create_loader, create_transform
66
from .mixup import mixup_target, FastCollateMixup

‎timm/data/loader.py‎

Lines changed: 37 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -86,25 +86,17 @@ def mixup_enabled(self, x):
8686
self.loader.collate_fn.mixup_enabled = x
8787

8888

89-
def create_loader(
90-
dataset,
89+
def create_transform(
9190
input_size,
92-
batch_size,
9391
is_training=False,
94-
use_prefetcher=True,
95-
rand_erase_prob=0.,
96-
rand_erase_mode='const',
92+
use_prefetcher=False,
9793
color_jitter=0.4,
9894
interpolation='bilinear',
9995
mean=IMAGENET_DEFAULT_MEAN,
10096
std=IMAGENET_DEFAULT_STD,
101-
num_workers=1,
102-
distributed=False,
10397
crop_pct=None,
104-
collate_fn=None,
105-
fp16=False,
106-
tf_preprocessing=False,
107-
):
98+
tf_preprocessing=False):
99+
108100
if isinstance(input_size, tuple):
109101
img_size = input_size[-2:]
110102
else:
@@ -132,7 +124,39 @@ def create_loader(
132124
std=std,
133125
crop_pct=crop_pct)
134126

135-
dataset.transform = transform
127+
return transform
128+
129+
130+
def create_loader(
131+
dataset,
132+
input_size,
133+
batch_size,
134+
is_training=False,
135+
use_prefetcher=True,
136+
rand_erase_prob=0.,
137+
rand_erase_mode='const',
138+
color_jitter=0.4,
139+
interpolation='bilinear',
140+
mean=IMAGENET_DEFAULT_MEAN,
141+
std=IMAGENET_DEFAULT_STD,
142+
num_workers=1,
143+
distributed=False,
144+
crop_pct=None,
145+
collate_fn=None,
146+
fp16=False,
147+
tf_preprocessing=False,
148+
):
149+
dataset.transform = create_transform(
150+
input_size,
151+
is_training=is_training,
152+
use_prefetcher=use_prefetcher,
153+
color_jitter=color_jitter,
154+
interpolation=interpolation,
155+
mean=mean,
156+
std=std,
157+
crop_pct=crop_pct,
158+
tf_preprocessing=tf_preprocessing,
159+
)
136160

137161
sampler = None
138162
if distributed:

0 commit comments

Comments
 (0)