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

Add initial sotabench attempt. Split create_transform out of create_loader. Update requirements.txt

Browse Source
pull/30/head
Ross Wightman 5 years ago
parent aeaaad7304
commit daeaa113e2
4 changed files with 254 additions and 16 deletions
Show Stats Download Patch File Download Diff File

4
requirements.txt
Unescape View File

@ -1,2 +1,2 @@
torch~=1.0 torch>=1.1.0
torchvision torchvision>=0.3.0

214
sotabench.py
Unescape View File

@ -0,0 +1,214 @@
from torchbench.image_classification import ImageNet
from timm import create_model, list_models
from timm.data import resolve_data_config, create_transform
NUM_GPU = 1
BATCH_SIZE = 256 * NUM_GPU
def _attrib(paper_model_name='', paper_arxiv_id='', batch_size=BATCH_SIZE):
return dict(
paper_model_name=paper_model_name,
paper_arxiv_id=paper_arxiv_id,
batch_size=batch_size)
model_map = dict(
#adv_inception_v3=_attrib(paper_model_name='Adversarial Inception V3', paper_arxiv_id=),
#densenet121=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#densenet161=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#densenet169=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#densenet201=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
dpn68=_attrib(
paper_model_name='DPN-68', paper_arxiv_id='1707.01629'),
dpn68b=_attrib(
paper_model_name='DPN-68b', paper_arxiv_id='1707.01629'),
dpn92=_attrib(
paper_model_name='DPN-92', paper_arxiv_id='1707.01629'),
dpn98=_attrib(
paper_model_name='DPN-98', paper_arxiv_id='1707.01629'),
dpn107=_attrib(
paper_model_name='DPN-107', paper_arxiv_id='1707.01629'),
dpn131=_attrib(
paper_model_name='DPN-131', paper_arxiv_id='1707.01629'),
efficientnet_b0=_attrib(
paper_model_name='EfficientNet-B0', paper_arxiv_id='1905.11946'),
efficientnet_b1=_attrib(
paper_model_name='EfficientNet-B1', paper_arxiv_id='1905.11946'),
efficientnet_b2=_attrib(
paper_model_name='EfficientNet-B2', paper_arxiv_id='1905.11946'),
#ens_adv_inception_resnet_v2=_attrib(paper_model_name=, paper_arxiv_id=),
fbnetc_100=_attrib(
paper_model_name='FBNet-C', paper_arxiv_id='1812.03443'),
gluon_inception_v3=_attrib(
paper_model_name='Inception V3', paper_arxiv_id='1512.00567'),
gluon_resnet18_v1b=_attrib(
paper_model_name='ResNet-18', paper_arxiv_id='1812.01187'),
gluon_resnet34_v1b=_attrib(
paper_model_name='ResNet-34', paper_arxiv_id='1812.01187'),
gluon_resnet50_v1b=_attrib(
paper_model_name='ResNet-50', paper_arxiv_id='1812.01187'),
gluon_resnet50_v1c=_attrib(
paper_model_name='ResNet-50-C', paper_arxiv_id='1812.01187'),
gluon_resnet50_v1d=_attrib(
paper_model_name='ResNet-50-D', paper_arxiv_id='1812.01187'),
gluon_resnet50_v1s=_attrib(
paper_model_name='ResNet-50-S', paper_arxiv_id='1812.01187'),
gluon_resnet101_v1b=_attrib(
paper_model_name='ResNet-101', paper_arxiv_id='1812.01187'),
gluon_resnet101_v1c=_attrib(
paper_model_name='ResNet-101-C', paper_arxiv_id='1812.01187'),
gluon_resnet101_v1d=_attrib(
paper_model_name='ResNet-101-D', paper_arxiv_id='1812.01187'),
gluon_resnet101_v1s=_attrib(
paper_model_name='ResNet-101-S', paper_arxiv_id='1812.01187'),
gluon_resnet152_v1b=_attrib(
paper_model_name='ResNet-152', paper_arxiv_id='1812.01187'),
gluon_resnet152_v1c=_attrib(
paper_model_name='ResNet-152-C', paper_arxiv_id='1812.01187'),
gluon_resnet152_v1d=_attrib(
paper_model_name='ResNet-152-D', paper_arxiv_id='1812.01187'),
gluon_resnet152_v1s=_attrib(
paper_model_name='ResNet-152-S', paper_arxiv_id='1812.01187'),
gluon_resnext50_32x4d=_attrib(
paper_model_name='ResNeXt-50 32x4d', paper_arxiv_id='1812.01187'),
gluon_resnext101_32x4d=_attrib(
paper_model_name='ResNeXt-101 32x4d', paper_arxiv_id='1812.01187'),
gluon_resnext101_64x4d=_attrib(
paper_model_name='ResNeXt-101 64x4d', paper_arxiv_id='1812.01187'),
gluon_senet154=_attrib(
paper_model_name='SENet-154', paper_arxiv_id='1812.01187'),
gluon_seresnext50_32x4d=_attrib(
paper_model_name='SE-ResNeXt-50 32x4d', paper_arxiv_id='1812.01187'),
gluon_seresnext101_32x4d=_attrib(
paper_model_name='SE-ResNeXt-101 32x4d', paper_arxiv_id='1812.01187'),
gluon_seresnext101_64x4d=_attrib(
paper_model_name='SE-ResNeXt-101 64x4d', paper_arxiv_id='1812.01187'),
gluon_xception65=_attrib(
paper_model_name='Modified Aligned Xception', paper_arxiv_id='1802.02611', batch_size=BATCH_SIZE//2),
ig_resnext101_32x8d=_attrib(
paper_model_name='ResNeXt-101 32×8d', paper_arxiv_id='1805.00932'),
ig_resnext101_32x16d=_attrib(
paper_model_name='ResNeXt-101 32×16d', paper_arxiv_id='1805.00932'),
ig_resnext101_32x32d=_attrib(
paper_model_name='ResNeXt-101 32×32d', paper_arxiv_id='1805.00932', batch_size=BATCH_SIZE//2),
ig_resnext101_32x48d=_attrib(
paper_model_name='ResNeXt-101 32×48d', paper_arxiv_id='1805.00932', batch_size=BATCH_SIZE//4),
inception_resnet_v2=_attrib(
paper_model_name='Inception ResNet V2', paper_arxiv_id='1602.07261'),
#inception_v3=dict(paper_model_name='Inception V3', paper_arxiv_id=), # same weights as torchvision
inception_v4=_attrib(
paper_model_name='Inception V4', paper_arxiv_id='1602.07261'),
mixnet_l=_attrib(
paper_model_name='MixNet-L', paper_arxiv_id='1907.09595'),
mixnet_m=_attrib(
paper_model_name='MixNet-M', paper_arxiv_id='1907.09595'),
mixnet_s=_attrib(
paper_model_name='MixNet-S', paper_arxiv_id='1907.09595'),
mnasnet_100=_attrib(
paper_model_name='MnasNet-B1', paper_arxiv_id='1807.11626'),
mobilenetv3_100=_attrib(
paper_model_name='MobileNet V3(1.0)', paper_arxiv_id='1905.02244'),
nasnetalarge=_attrib(
paper_model_name='NASNet-A Large', paper_arxiv_id='1707.07012', batch_size=BATCH_SIZE//4),
pnasnet5large=_attrib(
paper_model_name='PNASNet-5', paper_arxiv_id='1712.00559', batch_size=BATCH_SIZE//4),
resnet18=_attrib(
paper_model_name='ResNet-18', paper_arxiv_id='1812.01187'),
resnet26=_attrib(
paper_model_name='ResNet-26', paper_arxiv_id='1812.01187'),
resnet26d=_attrib(
paper_model_name='ResNet-26-D', paper_arxiv_id='1812.01187'),
resnet34=_attrib(
paper_model_name='ResNet-34', paper_arxiv_id='1812.01187'),
resnet50=_attrib(
paper_model_name='ResNet-50', paper_arxiv_id='1812.01187'),
#resnet101=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#resnet152=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
resnext50_32x4d=_attrib(
paper_model_name='ResNeXt-50 32x4d', paper_arxiv_id='1812.01187'),
resnext50d_32x4d=_attrib(
paper_model_name='ResNeXt-50-D 32x4d', paper_arxiv_id='1812.01187'),
#resnext101_32x8d=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
semnasnet_100=_attrib(
paper_model_name='MnasNet-A1', paper_arxiv_id='1807.11626'),
senet154=_attrib(
paper_model_name='SENet-154', paper_arxiv_id='1709.01507'),
seresnet18=_attrib(
paper_model_name='SE-ResNet-18', paper_arxiv_id='1709.01507'),
seresnet34=_attrib(
paper_model_name='SE-ResNet-34', paper_arxiv_id='1709.01507'),
seresnet50=_attrib(
paper_model_name='SE-ResNet-50', paper_arxiv_id='1709.01507'),
seresnet101=_attrib(
paper_model_name='SE-ResNet-101', paper_arxiv_id='1709.01507'),
seresnet152=_attrib(
paper_model_name='SE-ResNet-152', paper_arxiv_id='1709.01507'),
seresnext26_32x4d=_attrib(
paper_model_name='SE-ResNeXt-26 32x4d', paper_arxiv_id='1709.01507'),
seresnext50_32x4d=_attrib(
paper_model_name='SE-ResNeXt-50 32x4d', paper_arxiv_id='1709.01507'),
seresnext101_32x4d=_attrib(
paper_model_name='SE-ResNeXt-101 32x4d', paper_arxiv_id='1709.01507'),
spnasnet_100=_attrib(
paper_model_name='Single-Path NAS', paper_arxiv_id='1904.02877'),
tf_efficientnet_b0=_attrib(
paper_model_name='EfficientNet-B0', paper_arxiv_id='1905.11946'),
tf_efficientnet_b1=_attrib(
paper_model_name='EfficientNet-B1', paper_arxiv_id='1905.11946'),
tf_efficientnet_b2=_attrib(
paper_model_name='EfficientNet-B2', paper_arxiv_id='1905.11946'),
tf_efficientnet_b3=_attrib(
paper_model_name='EfficientNet-B3', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//2),
tf_efficientnet_b4=_attrib(
paper_model_name='EfficientNet-B4', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//2),
tf_efficientnet_b5=_attrib(
paper_model_name='EfficientNet-B5', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//4),
tf_efficientnet_b6=_attrib(
paper_model_name='EfficientNet-B6', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//8),
tf_efficientnet_b7=_attrib(
paper_model_name='EfficientNet-B7', paper_arxiv_id='1905.11946', batch_size=BATCH_SIZE//8),
tf_inception_v3=_attrib(
paper_model_name='Inception V3', paper_arxiv_id='1512.00567'),
tf_mixnet_l=_attrib(
paper_model_name='MixNet-L', paper_arxiv_id='1907.09595'),
tf_mixnet_m=_attrib(
paper_model_name='MixNet-M', paper_arxiv_id='1907.09595'),
tf_mixnet_s=_attrib(
paper_model_name='MixNet-S', paper_arxiv_id='1907.09595'),
#tv_resnet34=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#tv_resnet50=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#tv_resnext50_32x4d=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#wide_resnet50_2=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
#wide_resnet101_2=_attrib(paper_model_name=, paper_arxiv_id=), # same weights as torchvision
xception=_attrib(
paper_model_name='Xception', paper_arxiv_id='1610.02357'),
)
model_names = list_models(pretrained=True)
for model_name in model_names:
if model_name not in model_map:
print('Skipping %s' % model_name)
continue
# create model from name
model = create_model(model_name, pretrained=True)
param_count = sum([m.numel() for m in model.parameters()])
print('Model %s created, param count: %d' % (model_name, param_count))
# get appropriate transform for model's default pretrained config
data_config = resolve_data_config(dict(), model=model, verbose=True)
input_transform = create_transform(**data_config)
# Run the benchmark
ImageNet.benchmark(
model=model,
paper_model_name=model_map[model_name]['paper_model_name'],
paper_arxiv_id=model_map[model_name]['paper_arxiv_id'],
input_transform=input_transform,
batch_size=model_map[model_name]['batch_size'],
num_gpu=NUM_GPU,
#data_root=DATA_ROOT
)

2
timm/data/__init__.py
Unescape View File

@ -2,5 +2,5 @@ from .constants import *
from .config import resolve_data_config from .config import resolve_data_config
from .dataset import Dataset, DatasetTar from .dataset import Dataset, DatasetTar
from .transforms import * from .transforms import *
from .loader import create_loader from .loader import create_loader, create_transform
from .mixup import mixup_target, FastCollateMixup from .mixup import mixup_target, FastCollateMixup

50
timm/data/loader.py
Unescape View File

@ -86,25 +86,17 @@ class PrefetchLoader:
self.loader.collate_fn.mixup_enabled = x self.loader.collate_fn.mixup_enabled = x
def create_loader( def create_transform(
dataset,
input_size, input_size,
batch_size,
is_training=False, is_training=False,
use_prefetcher=True, use_prefetcher=False,
rand_erase_prob=0.,
rand_erase_mode='const',
color_jitter=0.4, color_jitter=0.4,
interpolation='bilinear', interpolation='bilinear',
mean=IMAGENET_DEFAULT_MEAN, mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD, std=IMAGENET_DEFAULT_STD,
num_workers=1,
distributed=False,
crop_pct=None, crop_pct=None,
collate_fn=None, tf_preprocessing=False):
fp16=False,
tf_preprocessing=False,
):
if isinstance(input_size, tuple): if isinstance(input_size, tuple):
img_size = input_size[-2:] img_size = input_size[-2:]
else: else:
@ -132,7 +124,39 @@ def create_loader(
std=std, std=std,
crop_pct=crop_pct) crop_pct=crop_pct)
dataset.transform = transform return transform
def create_loader(
dataset,
input_size,
batch_size,
is_training=False,
use_prefetcher=True,
rand_erase_prob=0.,
rand_erase_mode='const',
color_jitter=0.4,
interpolation='bilinear',
mean=IMAGENET_DEFAULT_MEAN,
std=IMAGENET_DEFAULT_STD,
num_workers=1,
distributed=False,
crop_pct=None,
collate_fn=None,
fp16=False,
tf_preprocessing=False,
):
dataset.transform = create_transform(
input_size,
is_training=is_training,
use_prefetcher=use_prefetcher,
color_jitter=color_jitter,
interpolation=interpolation,
mean=mean,
std=std,
crop_pct=crop_pct,
tf_preprocessing=tf_preprocessing,
)
sampler = None sampler = None
if distributed: if distributed:

Write Preview
Loading…
Cancel
Save