rwightman / pytorch-image-models

TOP 1 ACCURACY TOP 5 ACCURACY
SPEED
MODEL CODE PAPER
ε-REPR
CODE PAPER
ε-REPR
PAPER
GLOBAL RANK
Adversarial Inception V3
77.6% -- 93.7% -- 269.1 #178
DLA-102
78.0% -- 94.0% -- 384.0 #166
DLA-169
78.7% -- 94.3% -- 368.5 #132
DLA-34
74.6% -- 92.1% -- 404.7 #244
DLA-46-C
64.9% -- 86.3% -- 407.5 #352
DLA-60
77.0% -- 93.3% -- 391.0 #203
DLA-X-102
78.5% -- 94.2% -- 381.6 #135
DLA-X-102 64
79.5% -- 94.6% -- 348.3 #95
DLA-X-46-C
66.0% -- 87.0% -- 405.5 #334
DLA-X-60
78.2% -- 94.0% -- 387.5 #156
DLA-X-60-C
67.9% -- 88.4% -- 404.8 #319
DPN-107
(224x224)
80.2% -- 94.9% -- 215.4 #80
DPN-107
(320x320, Mean-Max Pooling)
81.8% -- 95.9% -- 106.6 #44
DPN-131
(224x224)
79.8% 80.1% 94.7% 94.9% 240.0 #94
DPN-131
(320x320, Mean-Max Pooling)
81.4% 81.5% 95.8% 95.8% 118.6 #47
DPN-68
(224x224)
76.3% 76.4% 93.0% 93.1% 357.0 #215
DPN-68
(320x320, Mean-Max Pooling)
78.5% 78.5% 94.4% 94.5% 292.3 #142
DPN-68b
(224x224)
77.5% -- 93.8% -- 363.5 #170
DPN-68b
(320x320, Mean-Max Pooling)
79.4% -- 95.0% -- 295.3 #97
DPN-92
(224x224)
80.0% 79.3%
94.8% 94.6% 338.0 #85
DPN-92
(320x320, Mean-Max Pooling)
81.3% 81.0%
95.7% 95.5% 244.4 #50
DPN-98
(224x224)
79.6% 80.0% 94.6% 94.8% 317.9 #100
DPN-98
(320x320, Mean-Max Pooling)
81.2% 81.3% 95.7% 95.6% 164.0 #52
EfficientNet-B0
76.9% 76.3%
93.2% 93.2% 363.4 #207
EfficientNet-B0
(AutoAugment)
76.8% -- 93.2% -- 358.6 #204
EfficientNet-B1
78.7% 78.8% 94.1% 94.4% 352.0 #135
EfficientNet-B1
(AutoAugment)
78.8% -- 94.2% -- 339.9 #139
EfficientNet-B2
79.8% 79.8% 94.7% 94.9% 330.4 #87
EfficientNet-B2
(AutoAugment)
80.1% -- 94.9% -- 325.8 #81
EfficientNet-B3
(AutoAugment)
81.6% -- 95.7% -- 300.5 #50
EfficientNet-B4
(AutoAugment)
83.0% -- 96.3% -- 234.5 #29
EfficientNet-B5
(RandAugment)
83.8% -- 96.8% -- 126.4 #21
EfficientNet-B6
(AutoAugment)
84.1% -- 96.9% -- 73.6 #19
EfficientNet-B7
(RandAugment)
84.9% -- 97.2% -- 43.1 #12
EfficientNet-EdgeTPU-L
80.4% -- 95.2% -- 282.0 #65
EfficientNet-EdgeTPU-M
78.7% -- 94.3% -- 336.5 #134
EfficientNet-EdgeTPU-S
77.3% -- 93.6% -- 358.9 #188
Ensemble Adversarial Inception V3
80.0% -- 94.9% -- 289.2 #80
FBNet-C
75.1% 74.9% 92.4% -- 422.1 #233
Inception ResNet V2
80.5% 80.1%
95.3% 95.1% 287.1 #61
Inception V3
78.8% 78.8% 94.4% 94.4% 303.2 #127
Inception V4
80.2% -- 95.0% -- 288.0 #73
MixNet-L
79.0% 78.9% 94.2% 94.2% 356.0 #121
MixNet-M
77.3% 77.0% 93.2% 93.3% 350.5 #189
MixNet-S
76.0% 75.8% 92.6% 92.8% 360.8 #225
MixNet-XL
80.1% -- 95.0% -- 347.4 #70
MnasNet-A1
75.5% 75.2% 92.6% 92.5% 356.2 #226
MnasNet-B1
74.7% -- 92.1% -- 356.9 #244
MobileNet V3-Large 1.0
75.6% 75.2%
92.7% -- 366.8 #225
Modified Aligned Xception
79.6% 79.8% 94.7% 94.8% 260.0 #92
NASNet-A Large
82.6% -- 96.0% -- 116.7 #35
PNASNet-5
82.7% 82.9% 96.0% 96.2% 118.1 #31
Res2Net-50 14x8s
78.2% -- 93.8% -- 386.0 #162
Res2Net-50 26x4s
77.9% -- 93.9% -- 394.0 #168
Res2Net-50 26x6s
78.6% -- 94.1% -- 382.9 #147
Res2Net-50 26x8s
79.2% -- 94.4% -- 377.1 #104
Res2Net-50 48x2s
77.5% -- 93.5% -- 386.9 #181
Res2Net-DLA-60
78.5% 79.5% 94.2% -- 391.6 #146
Res2NeXt-101 26x4s
79.2% -- 94.4% -- 377.8 #112
Res2NeXt-50
78.2% -- 93.9% -- 377.0 #157
Res2NeXt-DLA-60
78.4% -- 94.1% -- 377.1 #144
ResNet-101
79.3% -- 94.5% -- 346.2 #102
ResNet-101-C
79.5% -- 94.6% -- 349.5 #101
ResNet-101-D
80.4% -- 95.0% -- 346.8 #71
ResNet-101-S
80.3% -- 95.2% -- 341.9 #67
ResNet-152
79.7% -- 94.7% -- 339.9 #89
ResNet-152-C
79.9% -- 94.8% -- 333.1 #81
ResNet-152-D
80.5% -- 95.2% -- 337.3 #65
ResNet-152-S
81.0% -- 95.4% -- 335.2 #57
ResNet-18
70.8% -- 89.1% -- 368.7 #308
ResNet-18
72.6% -- 91.4% -- 423.6 #277
ResNet-26
75.3% -- 92.6% -- 354.9 #227
ResNet-26-D
76.7% -- 93.2% -- 352.7 #209
ResNet-34
74.6% -- 92.3% -- 372.3 #235
ResNet-50
78.5% -- 93.7% -- 355.0 #147
ResNet-50
79.2% -- 96.0% -- 397.4 #103
ResNet-50
(288x288 Mean-Max Pooling)
80.3% -- 95.6% -- 346.0 #67
ResNet-50-C
78.0% -- 94.0% -- 362.6 #169
ResNet-50-D
79.1% 77.2%
94.5% 93.5%
357.3 #108
ResNet-50-S
78.7% -- 94.2% -- 352.4 #133
ResNeXt-101 32x16d
83.3% -- 96.1% -- 150.9 #38
ResNeXt-101 32x16d
84.2% -- 97.2% -- 150.9 #14
ResNeXt-101 32x16d
(288x288 Mean-Max Pooling)
84.3% -- 97.3% -- 81.6 #13
ResNeXt-101 32x16d
(288x288 Mean-Max Pooling)
85.0% -- 97.6% -- 84.1 #9
ResNeXt-101 32x32d
85.1% 85.1% 97.4% 97.5% 60.5 #7
ResNeXt-101 32x32d
(288x288 Mean-Max Pooling)
85.9% -- 97.8% -- 37.6 #3
ResNeXt-101 32x48d
85.4% 85.4% 97.6% 97.6% 31.6 #6
ResNeXt-101 32x48d
(288x288 Mean-Max Pooling)
86.1% -- 97.9% -- 19.1 #2
ResNeXt-101 32x4d
80.9% -- 95.7% -- 372.3 #48
ResNeXt-101 32x4d
80.3% -- 94.9% -- 342.0 #68
ResNeXt-101 32x4d
(288x288 Mean-Max Pooling)
84.0% -- 97.2% -- 283.3 #11
ResNeXt-101 32x8d
81.6% -- 97.2% -- 286.4 #46
ResNeXt-101 32x8d
82.7% 82.2%
96.6% 96.4% 286.2 #29
ResNeXt-101 32x8d
(288x288 Mean-Max Pooling)
83.5% -- 97.1% -- 174.6 #16
ResNeXt-101 32x8d
(288x288 Mean-Max Pooling)
85.1% -- 96.5% -- 175.3 #32
ResNeXt-101 64x4d
80.6% -- 95.0% -- 277.5 #62
ResNeXt-50 32x4d
82.2% -- 96.2% -- 393.7 #37
ResNeXt-50 32x4d
78.5% -- 94.4% -- 345.2 #143
ResNeXt-50 32x4d
(288x288 Mean-Max Pooling)
81.3% -- 96.8% -- 337.8 #52
ResNeXt-50-D 32x4d
79.7% -- 94.9% -- 348.5 #90
SENet-154
81.3% 82.7% 95.5% 96.2% 215.6 #55
SENet-154
81.2% -- 95.4% -- 214.8 #60
SE-ResNet-101
78.4% -- 94.3% -- 391.5 #151
SE-ResNet-152
78.7% -- 94.4% -- 371.8 #138
SE-ResNet-18
71.8% -- 90.3% -- 354.7 #288
SE-ResNet-34
74.8% -- 92.1% -- 410.7 #243
SE-ResNet-50
77.6% -- 93.8% -- 399.1 #173
SE-ResNeXt-101 32x4d
80.2% -- 95.0% -- 375.5 #69
SE-ResNeXt-101 32x4d
80.9% -- 95.3% -- 335.2 #60
SE-ResNeXt-101 64x4d
80.9% -- 95.3% -- 266.0 #61
SE-ResNeXt-26 32x4d
77.1% -- 93.3% -- 353.9 #193
SE-ResNeXt-50 32x4d
79.1% -- 94.4% -- 393.5 #111
SE-ResNeXt-50 32x4d
79.9% -- 94.8% -- 347.1 #82
Single-Path NAS
74.1% 75.0% 91.8% 92.2% 417.0 #256
Xception
79.0% 79.0% 94.4% 94.5% 295.6 #119
See Full Build Details +get badge code
[![SotaBench](https://img.shields.io/endpoint.svg?url=https://sotabench.com/api/v0/badge/gh/rwightman/pytorch-image-models)](https://sotabench.com/user/rwightman/repos/rwightman/pytorch-image-models)

How the Repository is Evaluated

The full sotabench.py file - source
import torch
from torchbench.image_classification import ImageNet
from timm import create_model
from timm.data import resolve_data_config, create_transform
from timm.models import TestTimePoolHead
import os

NUM_GPU = 1
BATCH_SIZE = 256 * NUM_GPU


def _entry(model_name, paper_model_name, paper_arxiv_id, batch_size=BATCH_SIZE,
           ttp=False, args=dict(), model_desc=None):
    return dict(
        model=model_name,
        model_description=model_desc,
        paper_model_name=paper_model_name,
        paper_arxiv_id=paper_arxiv_id,
        batch_size=batch_size,
        ttp=ttp,
        args=args)

# NOTE For any original PyTorch models, I'll remove from this list when you add to sotabench to
# avoid overlap and confusion. Please contact me.
model_list = [
    ## Weights ported by myself from other frameworks or trained myself in PyTorch
    _entry('adv_inception_v3', 'Adversarial Inception V3', '1611.01236',
           model_desc='Ported from official Tensorflow weights'),
    _entry('ens_adv_inception_resnet_v2', 'Ensemble Adversarial Inception V3', '1705.07204',
           model_desc='Ported from official Tensorflow weights'),
    _entry('dpn68', 'DPN-68 (224x224)', '1707.01629'),
    _entry('dpn68b', 'DPN-68b (224x224)', '1707.01629'),
    _entry('dpn92', 'DPN-92 (224x224)', '1707.01629'),
    _entry('dpn98', 'DPN-98 (224x224)', '1707.01629'),
    _entry('dpn107', 'DPN-107 (224x224)', '1707.01629'),
    _entry('dpn131', 'DPN-131 (224x224)', '1707.01629'),
    _entry('dpn68', 'DPN-68 (320x320, Mean-Max Pooling)', '1707.01629', ttp=True, args=dict(img_size=320)),
    _entry('dpn68b', 'DPN-68b (320x320, Mean-Max Pooling)', '1707.01629', ttp=True, args=dict(img_size=320)),
    _entry('dpn92', 'DPN-92 (320x320, Mean-Max Pooling)', '1707.01629',
           ttp=True, args=dict(img_size=320), batch_size=BATCH_SIZE//2),
    _entry('dpn98', 'DPN-98 (320x320, Mean-Max Pooling)', '1707.01629',
           ttp=True, args=dict(img_size=320), batch_size=BATCH_SIZE//2),
    _entry('dpn107', 'DPN-107 (320x320, Mean-Max Pooling)', '1707.01629',
           ttp=True, args=dict(img_size=320), batch_size=BATCH_SIZE//4),
    _entry('dpn131', 'DPN-131 (320x320, Mean-Max Pooling)', '1707.01629',
           ttp=True, args=dict(img_size=320), batch_size=BATCH_SIZE//4),
    _entry('efficientnet_b0', 'EfficientNet-B0', '1905.11946'),
    _entry('efficientnet_b1', 'EfficientNet-B1', '1905.11946'),
    _entry('efficientnet_b2', 'EfficientNet-B2', '1905.11946'),
    _entry('fbnetc_100', 'FBNet-C', '1812.03443',
           model_desc='Trained in PyTorch with RMSProp, exponential LR decay'),
    _entry('gluon_inception_v3', 'Inception V3', '1512.00567', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet18_v1b', 'ResNet-18', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet34_v1b', 'ResNet-34', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet50_v1b', 'ResNet-50', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet50_v1c', 'ResNet-50-C', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet50_v1d', 'ResNet-50-D', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet50_v1s', 'ResNet-50-S', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet101_v1b', 'ResNet-101', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet101_v1c', 'ResNet-101-C', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet101_v1d', 'ResNet-101-D', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet101_v1s', 'ResNet-101-S', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet152_v1b', 'ResNet-152', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet152_v1c', 'ResNet-152-C', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet152_v1d', 'ResNet-152-D', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnet152_v1s', 'ResNet-152-S', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnext50_32x4d', 'ResNeXt-50 32x4d', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnext101_32x4d', 'ResNeXt-101 32x4d', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_resnext101_64x4d', 'ResNeXt-101 64x4d', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_senet154', 'SENet-154', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_seresnext50_32x4d', 'SE-ResNeXt-50 32x4d', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_seresnext101_32x4d', 'SE-ResNeXt-101 32x4d', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_seresnext101_64x4d', 'SE-ResNeXt-101 64x4d', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
    _entry('gluon_xception65', 'Modified Aligned Xception', '1802.02611', batch_size=BATCH_SIZE//2,
           model_desc='Ported from GluonCV Model Zoo'),
    _entry('mixnet_xl', 'MixNet-XL', '1907.09595', model_desc="My own scaling beyond paper's MixNet Large"),
    _entry('mixnet_l', 'MixNet-L', '1907.09595'),
    _entry('mixnet_m', 'MixNet-M', '1907.09595'),
    _entry('mixnet_s', 'MixNet-S', '1907.09595'),
    _entry('mnasnet_100', 'MnasNet-B1', '1807.11626'),
    _entry('mobilenetv3_100', 'MobileNet V3-Large 1.0', '1905.02244',
           model_desc='Trained in PyTorch with RMSProp, exponential LR decay, and hyper-params matching '
                      'paper as closely as possible.'),
    _entry('resnet18', 'ResNet-18', '1812.01187'),
    _entry('resnet26', 'ResNet-26', '1812.01187', model_desc='Block cfg of ResNet-34 w/ Bottleneck'),
    _entry('resnet26d', 'ResNet-26-D', '1812.01187',
           model_desc='Block cfg of ResNet-34 w/ Bottleneck, deep stem, and avg-pool in downsample layers.'),
    _entry('resnet34', 'ResNet-34', '1812.01187'),
    _entry('resnet50', 'ResNet-50', '1812.01187'),
    _entry('resnext50_32x4d', 'ResNeXt-50 32x4d', '1812.01187'),
    _entry('resnext50d_32x4d', 'ResNeXt-50-D 32x4d', '1812.01187',
           model_desc="'D' variant (3x3 deep stem w/ avg-pool downscale). Trained with "
                      "SGD w/ cosine LR decay, random-erasing (gaussian per-pixel noise) and label-smoothing"),
    _entry('semnasnet_100', 'MnasNet-A1', '1807.11626'),
    _entry('seresnet18', 'SE-ResNet-18', '1709.01507'),
    _entry('seresnet34', 'SE-ResNet-34', '1709.01507'),
    _entry('seresnext26_32x4d', 'SE-ResNeXt-26 32x4d', '1709.01507',
           model_desc='Block cfg of SE-ResNeXt-34 w/ Bottleneck, deep stem, and avg-pool in downsample layers.'),
    _entry('spnasnet_100', 'Single-Path NAS', '1904.02877',
           model_desc='Trained in PyTorch with SGD, cosine LR decay'),
    _entry('tf_efficientnet_b0', 'EfficientNet-B0 (AutoAugment)', '1905.11946',
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_b1', 'EfficientNet-B1 (AutoAugment)', '1905.11946',
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_b2', 'EfficientNet-B2 (AutoAugment)', '1905.11946',
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_b3', 'EfficientNet-B3 (AutoAugment)', '1905.11946', batch_size=BATCH_SIZE//2,
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_b4', 'EfficientNet-B4 (AutoAugment)', '1905.11946', batch_size=BATCH_SIZE//2,
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_b5', 'EfficientNet-B5 (RandAugment)', '1905.11946', batch_size=BATCH_SIZE//4,
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_b6', 'EfficientNet-B6 (AutoAugment)', '1905.11946', batch_size=BATCH_SIZE//8,
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_b7', 'EfficientNet-B7 (RandAugment)', '1905.11946', batch_size=BATCH_SIZE//8,
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_es', 'EfficientNet-EdgeTPU-S', '1905.11946',
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_em', 'EfficientNet-EdgeTPU-M', '1905.11946',
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_efficientnet_el', 'EfficientNet-EdgeTPU-L', '1905.11946', batch_size=BATCH_SIZE//2,
           model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_inception_v3', 'Inception V3', '1512.00567', model_desc='Ported from official Tensorflow weights'),
    _entry('tf_mixnet_l', 'MixNet-L', '1907.09595', model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_mixnet_m', 'MixNet-M', '1907.09595', model_desc='Ported from official Google AI Tensorflow weights'),
    _entry('tf_mixnet_s', 'MixNet-S', '1907.09595', model_desc='Ported from official Google AI Tensorflow weights'),

    ## Cadene ported weights (to remove if Cadene adds sotabench)
    _entry('inception_resnet_v2', 'Inception ResNet V2', '1602.07261'),
    _entry('inception_v4', 'Inception V4', '1602.07261'),
    _entry('nasnetalarge', 'NASNet-A Large', '1707.07012', batch_size=BATCH_SIZE // 4),
    _entry('pnasnet5large', 'PNASNet-5', '1712.00559', batch_size=BATCH_SIZE // 4),
    _entry('seresnet50', 'SE-ResNet-50', '1709.01507'),
    _entry('seresnet101', 'SE-ResNet-101', '1709.01507'),
    _entry('seresnet152', 'SE-ResNet-152', '1709.01507'),
    _entry('seresnext50_32x4d', 'SE-ResNeXt-50 32x4d', '1709.01507'),
    _entry('seresnext101_32x4d', 'SE-ResNeXt-101 32x4d', '1709.01507'),
    _entry('senet154', 'SENet-154', '1709.01507'),
    _entry('xception', 'Xception', '1610.02357',  batch_size=BATCH_SIZE//2),

    ## Torchvision weights
    # _entry('densenet121'),
    # _entry('densenet161'),
    # _entry('densenet169'),
    # _entry('densenet201'),
    # _entry('inception_v3', paper_model_name='Inception V3', ),
    # _entry('tv_resnet34', , ),
    # _entry('tv_resnet50', , ),
    # _entry('resnet101', , ),
    # _entry('resnet152', , ),
    # _entry('tv_resnext50_32x4d', , ),
    # _entry('resnext101_32x8d', ),
    # _entry('wide_resnet50_2' , ),
    # _entry('wide_resnet101_2', , ),

    ## Facebook WSL weights
    _entry('ig_resnext101_32x8d', 'ResNeXt-101 32x8d', '1805.00932',
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),
    _entry('ig_resnext101_32x16d', 'ResNeXt-101 32x16d', '1805.00932',
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),
    _entry('ig_resnext101_32x32d', 'ResNeXt-101 32x32d', '1805.00932', batch_size=BATCH_SIZE // 2,
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),
    _entry('ig_resnext101_32x48d', 'ResNeXt-101 32x48d', '1805.00932', batch_size=BATCH_SIZE // 4,
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),

    _entry('ig_resnext101_32x8d', 'ResNeXt-101 32x8d (288x288 Mean-Max Pooling)', '1805.00932',
           ttp=True, args=dict(img_size=288),
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),
    _entry('ig_resnext101_32x16d', 'ResNeXt-101 32x16d (288x288 Mean-Max Pooling)', '1805.00932',
           ttp=True, args=dict(img_size=288), batch_size=BATCH_SIZE // 2,
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),
    _entry('ig_resnext101_32x32d', 'ResNeXt-101 32x32d (288x288 Mean-Max Pooling)', '1805.00932',
           ttp=True, args=dict(img_size=288), batch_size=BATCH_SIZE // 4,
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),
    _entry('ig_resnext101_32x48d', 'ResNeXt-101 32x48d (288x288 Mean-Max Pooling)', '1805.00932',
           ttp=True, args=dict(img_size=288), batch_size=BATCH_SIZE // 8,
           model_desc='Weakly-Supervised pre-training on 1B Instagram hashtag dataset by Facebook Research'),

    ## Facebook SSL weights
    _entry('ssl_resnet18', 'ResNet-18', '1905.00546',
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnet50', 'ResNet-50', '1905.00546',
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext50_32x4d', 'ResNeXt-50 32x4d', '1905.00546',
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext101_32x4d', 'ResNeXt-101 32x4d', '1905.00546',
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext101_32x8d', 'ResNeXt-101 32x8d', '1905.00546',
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext101_32x16d', 'ResNeXt-101 32x16d', '1905.00546',
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),

    _entry('ssl_resnet50', 'ResNet-50 (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext50_32x4d', 'ResNeXt-50 32x4d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext101_32x4d', 'ResNeXt-101 32x4d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext101_32x8d', 'ResNeXt-101 32x8d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),
    _entry('ssl_resnext101_32x16d', 'ResNeXt-101 32x16d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288), batch_size=BATCH_SIZE // 2,
           model_desc='Semi-Supervised pre-training on YFCC100M dataset by Facebook Research'),

    ## Facebook SWSL weights
    _entry('swsl_resnet18', 'ResNet-18', '1905.00546',
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnet50', 'ResNet-50', '1905.00546',
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext50_32x4d', 'ResNeXt-50 32x4d', '1905.00546',
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext101_32x4d', 'ResNeXt-101 32x4d', '1905.00546',
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext101_32x8d', 'ResNeXt-101 32x8d', '1905.00546',
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext101_32x16d', 'ResNeXt-101 32x16d', '1905.00546',
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),

    _entry('swsl_resnet50', 'ResNet-50 (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext50_32x4d', 'ResNeXt-50 32x4d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext101_32x4d', 'ResNeXt-101 32x4d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext101_32x8d', 'ResNeXt-101 32x8d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288),
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),
    _entry('swsl_resnext101_32x16d', 'ResNeXt-101 32x16d (288x288 Mean-Max Pooling)', '1905.00546',
           ttp=True, args=dict(img_size=288), batch_size=BATCH_SIZE // 2,
           model_desc='Semi-Weakly-Supervised pre-training on 1 billion unlabelled dataset by Facebook Research'),

    ## DLA official impl weights (to remove if sotabench added to source)
    _entry('dla34', 'DLA-34', '1707.06484'),
    _entry('dla46_c', 'DLA-46-C', '1707.06484'),
    _entry('dla46x_c', 'DLA-X-46-C', '1707.06484'),
    _entry('dla60x_c', 'DLA-X-60-C', '1707.06484'),
    _entry('dla60', 'DLA-60', '1707.06484'),
    _entry('dla60x', 'DLA-X-60', '1707.06484'),
    _entry('dla102', 'DLA-102', '1707.06484'),
    _entry('dla102x', 'DLA-X-102', '1707.06484'),
    _entry('dla102x2', 'DLA-X-102 64', '1707.06484'),
    _entry('dla169', 'DLA-169', '1707.06484'),

    ## Res2Net official impl weights (to remove if sotabench added to source)
    _entry('res2net50_26w_4s', 'Res2Net-50 26x4s', '1904.01169'),
    _entry('res2net50_14w_8s', 'Res2Net-50 14x8s', '1904.01169'),
    _entry('res2net50_26w_6s', 'Res2Net-50 26x6s', '1904.01169'),
    _entry('res2net50_26w_8s', 'Res2Net-50 26x8s', '1904.01169'),
    _entry('res2net50_48w_2s', 'Res2Net-50 48x2s', '1904.01169'),
    _entry('res2net101_26w_4s', 'Res2NeXt-101 26x4s', '1904.01169'),
    _entry('res2next50', 'Res2NeXt-50', '1904.01169'),
    _entry('dla60_res2net', 'Res2Net-DLA-60', '1904.01169'),
    _entry('dla60_res2next', 'Res2NeXt-DLA-60', '1904.01169'),
]

for m in model_list:
    model_name = m['model']
    # create model from name
    model = create_model(model_name, pretrained=True)
    param_count = sum([m.numel() for m in model.parameters()])
    print('Model %s, %s created. Param count: %d' % (model_name, m['paper_model_name'], param_count))

    # get appropriate transform for model's default pretrained config
    data_config = resolve_data_config(m['args'], model=model, verbose=True)
    if m['ttp']:
        model = TestTimePoolHead(model, model.default_cfg['pool_size'])
        data_config['crop_pct'] = 1.0
    input_transform = create_transform(**data_config)

    # Run the benchmark
    ImageNet.benchmark(
        model=model,
        model_description=m.get('model_description', None),
        paper_model_name=m['paper_model_name'],
        paper_arxiv_id=m['paper_arxiv_id'],
        input_transform=input_transform,
        batch_size=m['batch_size'],
        num_gpu=NUM_GPU,
        data_root=os.environ.get('IMAGENET_DIR', './imagenet')
    )

    torch.cuda.empty_cache()


STATUS
BUILD
COMMIT MESSAGE
RUN TIME
Fix non-prefetch variant of Mixup. Fixes #50
rwightman   4748c6d  ·  Nov 02 2019
0h:32m:07s
Add TF RandAug weights for B5/B7 EfficientNet models.
rwightman   0d58c50  ·  Oct 30 2019
0h:57m:20s
Better differentiate sotabench WSL, SSL, and SWSL models via mod…
rwightman   62105ed  ·  Oct 20 2019
0h:31m:09s
Map pretrained checkpoint to cpu to avoid issue with some pretra…
rwightman   c099374  ·  Oct 19 2019
0h:29m:28s
Add Facebook Research Semi-Supervised and Semi-Weakly Supervised…
rwightman a9eb484    b93fcf0  ·  Oct 19 2019
7h:41m:03s
6h:58m:34s
Add support for loading args from yaml file (and saving them wit…
rwightman   187ecba  ·  Sep 09 2019
0h:27m:02s
Fix Mobilenet V3 model name for sotabench. Minor res2net cleanup.
rwightman   d3ba34e  ·  Sep 05 2019
2h:26m:03s
Silly typos.
rwightman   b5a8bb5  ·  Sep 04 2019
3h:23m:37s
sotabench fail
rwightman   7dc5d7a  ·  Sep 04 2019
1h:11m:43s
Merge pull request #35 from rwightman/res2net_dla Add Res2net a…
rwightman   96364fc  (+2 commits )  ·  Sep 04 2019
0h:58m:37s
1h:36m:10s
0h:58m:34s