PyTorch image models, scripts, pretrained weights -- ResNet, ResNeXT, EfficientNet, EfficientNetV2, NFNet, Vision Transformer, MixNet, MobileNet-V3/V2, RegNet, DPN, CSPNet, and more

Go to file

Ross Wightman 820b73d280 Update README.md		5 years ago
convert	Add working Gluon Xception-65 model. Some cleanup still needed.	5 years ago
notebooks	ImageNet-1k vs ImageNet-v2 comparison	5 years ago
results	Update README.md	5 years ago
timm	Add support to Dataset for class id mapping file, clean up a bit of old logic. Add results file arg for validation and update script.	5 years ago
.gitattributes	Add .gitattributes	5 years ago
.gitignore	Add gitignore	5 years ago
LICENSE	Add Apache LICENSE file	6 years ago
README.md	Update README.md	5 years ago
avg_checkpoints.py	Plural for averaging script.	5 years ago
clean_checkpoint.py	Add support to clean_checkpoint.py to remove aux_bn weights/biases from SplitBatchNorm	5 years ago
distributed_train.sh	Fix distributed train script	6 years ago
inference.py	Add checkpoint averaging script. Add headers, shebangs, exec perms to all scripts	5 years ago
requirements.txt	Update EfficientNet comments, MobileNetV3 non-TF create fns, fix factory arg checks, bump PyTorch version req to 1.2	5 years ago
setup.py	markdown in setup.py	5 years ago
sotabench.py	Update sotabench with tf_efficientnet_b8 model	5 years ago
train.py	Add warning about using sync-bn with zero initialized BN layers. Fixes #54	5 years ago
validate.py	Add support to Dataset for class id mapping file, clean up a bit of old logic. Add results file arg for validation and update script.	5 years ago

README.md

PyTorch Image Models, etc

What's New

Jan 31, 2020

Update ResNet50 weights with a new 79.038 result from further JSD / AugMix experiments. Full command line for reproduction in training section below.

Jan 11/12, 2020

Master may be a bit unstable wrt to training, these changes have been tested but not all combos
Implementations of AugMix added to existing RA and AA. Including numerous supporting pieces like JSD loss (Jensen-Shannon divergence + CE), and AugMixDataset
SplitBatchNorm adaptation layer added for implementing Auxiliary BN as per AdvProp paper
ResNet-50 AugMix trained model w/ 79% top-1 added
seresnext26tn_32x4d - 77.99 top-1, 93.75 top-5 added to tiered experiment, higher img/s than 't' and 'd'

Jan 3, 2020

Add RandAugment trained EfficientNet-B0 weight with 77.7 top-1. Trained by Michael Klachko with this code and recent hparams (see Training section)
Add avg_checkpoints.py script for post training weight averaging and update all scripts with header docstrings and shebangs.

Dec 30, 2019

Merge Dushyant Mehta's PR for SelecSLS (Selective Short and Long Range Skip Connections) networks. Good GPU memory consumption and throughput. Original: https://github.com/mehtadushy/SelecSLS-Pytorch

Dec 28, 2019

Add new model weights and training hparams (see Training Hparams section)
- efficientnet_b3 - 81.5 top-1, 95.7 top-5 at default res/crop, 81.9, 95.8 at 320x320 1.0 crop-pct
  - trained with RandAugment, ended up with an interesting but less than perfect result (see training section)
- seresnext26d_32x4d- 77.6 top-1, 93.6 top-5
  - deep stem (32, 32, 64), avgpool downsample
  - stem/dowsample from bag-of-tricks paper
- seresnext26t_32x4d- 78.0 top-1, 93.7 top-5
  - deep tiered stem (24, 48, 64), avgpool downsample (a modified 'D' variant)
  - stem sizing mods from Jeremy Howard and fastai devs discussing ResNet architecture experiments

Dec 23, 2019

Add RandAugment trained MixNet-XL weights with 80.48 top-1.
--dist-bn argument added to train.py, will distribute BN stats between nodes after each train epoch, before eval

Dec 4, 2019

Added weights from the first training from scratch of an EfficientNet (B2) with my new RandAugment implementation. Much better than my previous B2 and very close to the official AdvProp ones (80.4 top-1, 95.08 top-5).

Nov 29, 2019

Brought EfficientNet and MobileNetV3 up to date with my https://github.com/rwightman/gen-efficientnet-pytorch code. Torchscript and ONNX export compat excluded.
- AdvProp weights added
- Official TF MobileNetv3 weights added
EfficientNet and MobileNetV3 hook based 'feature extraction' classes added. Will serve as basis for using models as backbones in obj detection/segmentation tasks. Lots more to be done here...
HRNet classification models and weights added from https://github.com/HRNet/HRNet-Image-Classification
Consistency in global pooling, reset_classifer, and forward_features across models
- forward_features always returns unpooled feature maps now
Reasonable chance I broke something... let me know

Nov 22, 2019

Add ImageNet training RandAugment implementation alongside AutoAugment. PyTorch Transform compatible format, using PIL. Currently training two EfficientNet models from scratch with promising results... will update.
drop-connect cmd line arg finally added to train.py, no need to hack model fns. Works for efficientnet/mobilenetv3 based models, ignored otherwise.

Introduction

For each competition, personal, or freelance project involving images + Convolution Neural Networks, I build on top of an evolving collection of code and models. This repo contains a (somewhat) cleaned up and paired down iteration of that code. Hopefully it'll be of use to others.

The work of many others is present here. I've tried to make sure all source material is acknowledged:

Training/validation scripts evolved from early versions of the PyTorch Imagenet Examples
CUDA specific performance enhancements have been pulled from NVIDIA's APEX Examples
LR scheduler ideas from AllenNLP, FAIRseq, and SGDR: Stochastic Gradient Descent with Warm Restarts (https://arxiv.org/abs/1608.03983)
Random Erasing from Zhun Zhong (https://arxiv.org/abs/1708.04896)
Optimizers:
- RAdam by Liyuan Liu (https://arxiv.org/abs/1908.03265)
- NovoGrad by Masashi Kimura (https://arxiv.org/abs/1905.11286)
- Lookahead adapted from impl by Liam (https://arxiv.org/abs/1907.08610)

Models

I've included a few of my favourite models, but this is not an exhaustive collection. You can't do better than Cadene's collection in that regard. Most models do have pretrained weights from their respective sources or original authors.

Included models:

ResNet/ResNeXt (from torchvision with mods by myself)
- ResNet-18, ResNet-34, ResNet-50, ResNet-101, ResNet-152, ResNeXt50 (32x4d), ResNeXt101 (32x4d and 64x4d)
- 'Bag of Tricks' / Gluon C, D, E, S variations (https://arxiv.org/abs/1812.01187)
- Instagram trained / ImageNet tuned ResNeXt101-32x8d to 32x48d from from facebookresearch
- Res2Net (https://github.com/gasvn/Res2Net, https://arxiv.org/abs/1904.01169)
DLA
- Original (https://github.com/ucbdrive/dla, https://arxiv.org/abs/1707.06484)
- Res2Net (https://github.com/gasvn/Res2Net, https://arxiv.org/abs/1904.01169)
DenseNet (from torchvision)
- DenseNet-121, DenseNet-169, DenseNet-201, DenseNet-161
Squeeze-and-Excitation ResNet/ResNeXt (from Cadene with some pretrained weight additions by myself)
- SENet-154, SE-ResNet-18, SE-ResNet-34, SE-ResNet-50, SE-ResNet-101, SE-ResNet-152, SE-ResNeXt-26 (32x4d), SE-ResNeXt50 (32x4d), SE-ResNeXt101 (32x4d)
Inception-ResNet-V2 and Inception-V4 (from Cadene )
Xception
- Original variant from Cadene
- MXNet Gluon 'modified aligned' Xception-65 and 71 models from Gluon ModelZoo
PNasNet & NASNet-A (from Cadene)
DPN (from me, weights hosted by Cadene)
- DPN-68, DPN-68b, DPN-92, DPN-98, DPN-131, DPN-107
EfficientNet (from my standalone GenMobileNet) - A generic model that implements many of the efficient models that utilize similar DepthwiseSeparable and InvertedResidual blocks
- EfficientNet AdvProp (B0-B8) (https://arxiv.org/abs/1911.09665) -- TF weights ported
- EfficientNet (B0-B7) (https://arxiv.org/abs/1905.11946) -- TF weights ported, B0-B2 finetuned PyTorch
- EfficientNet-EdgeTPU (S, M, L) (https://ai.googleblog.com/2019/08/efficientnet-edgetpu-creating.html) --TF weights ported
- MixNet (https://arxiv.org/abs/1907.09595) -- TF weights ported, PyTorch finetuned (S, M, L) or trained models (XL)
- MNASNet B1, A1 (Squeeze-Excite), and Small (https://arxiv.org/abs/1807.11626) -- trained in PyTorch
- MobileNet-V2 (https://arxiv.org/abs/1801.04381)
- FBNet-C (https://arxiv.org/abs/1812.03443) -- trained in PyTorch
- Single-Path NAS (https://arxiv.org/abs/1904.02877) -- pixel1 variant
MobileNet-V3 (https://arxiv.org/abs/1905.02244) -- pretrained PyTorch model, official TF weights ported
HRNet
- code from https://github.com/HRNet/HRNet-Image-Classification, paper https://arxiv.org/abs/1908.07919
SelecSLS
- code from https://github.com/mehtadushy/SelecSLS-Pytorch, paper https://arxiv.org/abs/1907.00837

Use the --model arg to specify model for train, validation, inference scripts. Match the all lowercase creation fn for the model you'd like.

Features

Several (less common) features that I often utilize in my projects are included. Many of their additions are the reason why I maintain my own set of models, instead of using others' via PIP:

All models have a common default configuration interface and API for
- accessing/changing the classifier - get_classifier and reset_classifier
- doing a forward pass on just the features - forward_features
- these makes it easy to write consistent network wrappers that work with any of the models
All models have a consistent pretrained weight loader that adapts last linear if necessary, and from 3 to 1 channel input if desired
The train script works in several process/GPU modes:
- NVIDIA DDP w/ a single GPU per process, multiple processes with APEX present (AMP mixed-precision optional)
- PyTorch DistributedDataParallel w/ multi-gpu, single process (AMP disabled as it crashes when enabled)
- PyTorch w/ single GPU single process (AMP optional)
A dynamic global pool implementation that allows selecting from average pooling, max pooling, average + max, or concat([average, max]) at model creation. All global pooling is adaptive average by default and compatible with pretrained weights.
A 'Test Time Pool' wrapper that can wrap any of the included models and usually provide improved performance doing inference with input images larger than the training size. Idea adapted from original DPN implementation when I ported (https://github.com/cypw/DPNs)
Training schedules and techniques that provide competitive results (Cosine LR, Random Erasing, Label Smoothing, etc)
Mixup (as in https://arxiv.org/abs/1710.09412) - currently implementing/testing
An inference script that dumps output to CSV is provided as an example
AutoAugment (https://arxiv.org/abs/1805.09501) and RandAugment (https://arxiv.org/abs/1909.13719) ImageNet configurations modeled after impl for EfficientNet training (https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/autoaugment.py)
AugMix w/ JSD loss (https://arxiv.org/abs/1912.02781), JSD w/ clean + augmented mixing support works with AutoAugment and RandAugment as well
SplitBachNorm - allows splitting batch norm layers between clean and augmented (auxiliary batch norm) data

Results

A CSV file containing an ImageNet-1K validation results summary for all included models with pretrained weights and default configurations is located here

Self-trained Weights

I've leveraged the training scripts in this repository to train a few of the models with missing weights to good levels of performance. These numbers are all for 224x224 training and validation image sizing with the usual 87.5% validation crop.

Model	Prec@1 (Err)	Prec@5 (Err)	Param #	Image Scaling	Image Size
efficientnet_b3a	81.874 (18.126)	95.840 (4.160)	9.11M	bicubic	320 (1.0 crop)
efficientnet_b3	81.498 (18.502)	95.718 (4.282)	9.11M	bicubic	300
efficientnet_b2a	80.608 (19.392)	95.310 (4.690)	9.11M	bicubic	288 (1.0 crop)
mixnet_xl	80.478 (19.522)	94.932 (5.068)	11.90M	bicubic	224
efficientnet_b2	80.402 (19.598)	95.076 (4.924)	9.11M	bicubic	260
resnext50d_32x4d	79.674 (20.326)	94.868 (5.132)	25.1M	bicubic	224
resnet50	79.038 (20.962)	94.390 (5.610)	25.6M	bicubic	224
mixnet_l	78.976 (21.024	94.184 (5.816)	7.33M	bicubic	224
efficientnet_b1	78.692 (21.308)	94.086 (5.914)	7.79M	bicubic	240
resnext50_32x4d	78.512 (21.488)	94.042 (5.958)	25M	bicubic	224
seresnext26t_32x4d	77.998 (22.002)	93.708 (6.292)	16.8M	bicubic	224
seresnext26tn_32x4d	77.986 (22.014)	93.746 (6.254)	16.8M	bicubic	224
efficientnet_b0	77.698 (22.302)	93.532 (6.468)	5.29M	bicubic	224
seresnext26d_32x4d	77.602 (22.398)	93.608 (6.392)	16.8M	bicubic	224
mixnet_m	77.256 (22.744)	93.418 (6.582)	5.01M	bicubic	224
seresnext26_32x4d	77.104 (22.896)	93.316 (6.684)	16.8M	bicubic	224
resnet26d	76.68 (23.32)	93.166 (6.834)	16M	bicubic	224
mixnet_s	75.988 (24.012)	92.794 (7.206)	4.13M	bicubic	224
mobilenetv3_100	75.634 (24.366)	92.708 (7.292)	5.5M	bicubic	224
mnasnet_a1	75.448 (24.552)	92.604 (7.396)	3.89M	bicubic	224
resnet26	75.292 (24.708)	92.57 (7.43)	16M	bicubic	224
fbnetc_100	75.124 (24.876)	92.386 (7.614)	5.6M	bilinear	224
resnet34	75.110 (24.890)	92.284 (7.716)	22M	bilinear	224
seresnet34	74.808 (25.192)	92.124 (7.876)	22M	bilinear	224
mnasnet_b1	74.658 (25.342)	92.114 (7.886)	4.38M	bicubic	224
spnasnet_100	74.084 (25.916)	91.818 (8.182)	4.42M	bilinear	224
seresnet18	71.742 (28.258)	90.334 (9.666)	11.8M	bicubic	224

Ported Weights

For the models below, the model code and weight porting from Tensorflow or MXNet Gluon to Pytorch was done by myself. There are weights/models ported by others included in this repository, they are not listed below.

Model	Prec@1 (Err)	Prec@5 (Err)	Param #	Image Scaling	Image Size
tf_efficientnet_b8_ap *tfp	85.436 (14.564)	97.272 (2.728)	87.4	bicubic	672
tf_efficientnet_b8_ap	85.368 (14.632)	97.294 (2.706)	87.4	bicubic	672
tf_efficientnet_b7_ap *tfp	85.154 (14.846)	97.244 (2.756)	66.35	bicubic	600
tf_efficientnet_b7_ap	85.118 (14.882)	97.252 (2.748)	66.35	bicubic	600
tf_efficientnet_b7 *tfp	84.940 (15.060)	97.214 (2.786)	66.35	bicubic	600
tf_efficientnet_b7	84.932 (15.068)	97.208 (2.792)	66.35	bicubic	600
tf_efficientnet_b6_ap	84.786 (15.214)	97.138 (2.862)	43.04	bicubic	528
tf_efficientnet_b6_ap *tfp	84.760 (15.240)	97.124 (2.876)	43.04	bicubic	528
tf_efficientnet_b5_ap *tfp	84.276 (15.724)	96.932 (3.068)	30.39	bicubic	456
tf_efficientnet_b5_ap	84.254 (15.746)	96.976 (3.024)	30.39	bicubic	456
tf_efficientnet_b6 *tfp	84.140 (15.860)	96.852 (3.148)	43.04	bicubic	528
tf_efficientnet_b6	84.110 (15.890)	96.886 (3.114)	43.04	bicubic	528
tf_efficientnet_b5 *tfp	83.822 (16.178)	96.756 (3.244)	30.39	bicubic	456
tf_efficientnet_b5	83.812 (16.188)	96.748 (3.252)	30.39	bicubic	456
tf_efficientnet_b4_ap *tfp	83.278 (16.722)	96.376 (3.624)	19.34	bicubic	380
tf_efficientnet_b4_ap	83.248 (16.752)	96.388 (3.612)	19.34	bicubic	380
tf_efficientnet_b4	83.022 (16.978)	96.300 (3.700)	19.34	bicubic	380
tf_efficientnet_b4 *tfp	82.948 (17.052)	96.308 (3.692)	19.34	bicubic	380
tf_efficientnet_b3_ap *tfp	81.882 (18.118)	95.662 (4.338)	12.23	bicubic	300
tf_efficientnet_b3_ap	81.828 (18.172)	95.624 (4.376)	12.23	bicubic	300
tf_efficientnet_b3	81.636 (18.364)	95.718 (4.282)	12.23	bicubic	300
tf_efficientnet_b3 *tfp	81.576 (18.424)	95.662 (4.338)	12.23	bicubic	300
gluon_senet154	81.224 (18.776)	95.356 (4.644)	115.09	bicubic	224
gluon_resnet152_v1s	81.012 (18.988)	95.416 (4.584)	60.32	bicubic	224
gluon_seresnext101_32x4d	80.902 (19.098)	95.294 (4.706)	48.96	bicubic	224
gluon_seresnext101_64x4d	80.890 (19.110)	95.304 (4.696)	88.23	bicubic	224
gluon_resnext101_64x4d	80.602 (19.398)	94.994 (5.006)	83.46	bicubic	224
tf_efficientnet_el	80.534 (19.466)	95.190 (4.810)	10.59	bicubic	300
tf_efficientnet_el *tfp	80.476 (19.524)	95.200 (4.800)	10.59	bicubic	300
gluon_resnet152_v1d	80.470 (19.530)	95.206 (4.794)	60.21	bicubic	224
gluon_resnet101_v1d	80.424 (19.576)	95.020 (4.980)	44.57	bicubic	224
tf_efficientnet_b2_ap *tfp	80.420 (19.580)	95.040 (4.960)	9.11	bicubic	260
gluon_resnext101_32x4d	80.334 (19.666)	94.926 (5.074)	44.18	bicubic	224
tf_efficientnet_b2_ap	80.306 (19.694)	95.028 (4.972)	9.11	bicubic	260
gluon_resnet101_v1s	80.300 (19.700)	95.150 (4.850)	44.67	bicubic	224
tf_efficientnet_b2 *tfp	80.188 (19.812)	94.974 (5.026)	9.11	bicubic	260
tf_efficientnet_b2	80.086 (19.914)	94.908 (5.092)	9.11	bicubic	260
gluon_resnet152_v1c	79.916 (20.084)	94.842 (5.158)	60.21	bicubic	224
gluon_seresnext50_32x4d	79.912 (20.088)	94.818 (5.182)	27.56	bicubic	224
gluon_resnet152_v1b	79.692 (20.308)	94.738 (5.262)	60.19	bicubic	224
gluon_xception65	79.604 (20.396)	94.748 (5.252)	39.92	bicubic	299
gluon_resnet101_v1c	79.544 (20.456)	94.586 (5.414)	44.57	bicubic	224
tf_efficientnet_b1_ap *tfp	79.532 (20.468)	94.378 (5.622)	7.79	bicubic	240
tf_efficientnet_cc_b1_8e *tfp	79.464 (20.536)	94.492 (5.508)	39.7	bicubic	240
gluon_resnext50_32x4d	79.356 (20.644)	94.424 (5.576)	25.03	bicubic	224
gluon_resnet101_v1b	79.304 (20.696)	94.524 (5.476)	44.55	bicubic	224
tf_efficientnet_cc_b1_8e	79.298 (20.702)	94.364 (5.636)	39.7	bicubic	240
tf_efficientnet_b1_ap	79.278 (20.722)	94.308 (5.692)	7.79	bicubic	240
tf_efficientnet_b1 *tfp	79.172 (20.828)	94.450 (5.550)	7.79	bicubic	240
gluon_resnet50_v1d	79.074 (20.926)	94.476 (5.524)	25.58	bicubic	224
tf_efficientnet_em *tfp	78.958 (21.042)	94.458 (5.542)	6.90	bicubic	240
tf_mixnet_l *tfp	78.846 (21.154)	94.212 (5.788)	7.33	bilinear	224
tf_efficientnet_b1	78.826 (21.174)	94.198 (5.802)	7.79	bicubic	240
gluon_inception_v3	78.804 (21.196)	94.380 (5.620)	27.16M	bicubic	299
tf_mixnet_l	78.770 (21.230)	94.004 (5.996)	7.33	bicubic	224
tf_efficientnet_em	78.742 (21.258)	94.332 (5.668)	6.90	bicubic	240
gluon_resnet50_v1s	78.712 (21.288)	94.242 (5.758)	25.68	bicubic	224
tf_efficientnet_cc_b0_8e *tfp	78.314 (21.686)	93.790 (6.210)	24.0	bicubic	224
gluon_resnet50_v1c	78.010 (21.990)	93.988 (6.012)	25.58	bicubic	224
tf_efficientnet_cc_b0_8e	77.908 (22.092)	93.656 (6.344)	24.0	bicubic	224
tf_inception_v3	77.856 (22.144)	93.644 (6.356)	27.16M	bicubic	299
tf_efficientnet_cc_b0_4e *tfp	77.746 (22.254)	93.552 (6.448)	13.3	bicubic	224
tf_efficientnet_es *tfp	77.616 (22.384)	93.750 (6.250)	5.44	bicubic	224
gluon_resnet50_v1b	77.578 (22.422)	93.718 (6.282)	25.56	bicubic	224
adv_inception_v3	77.576 (22.424)	93.724 (6.276)	27.16M	bicubic	299
tf_efficientnet_b0_ap *tfp	77.514 (22.486)	93.576 (6.424)	5.29	bicubic	224
tf_efficientnet_cc_b0_4e	77.304 (22.696)	93.332 (6.668)	13.3	bicubic	224
tf_efficientnet_es	77.264 (22.736)	93.600 (6.400)	5.44	bicubic	224
tf_efficientnet_b0 *tfp	77.258 (22.742)	93.478 (6.522)	5.29	bicubic	224
tf_efficientnet_b0_ap	77.084 (22.916)	93.254 (6.746)	5.29	bicubic	224
tf_mixnet_m *tfp	77.072 (22.928)	93.368 (6.632)	5.01	bilinear	224
tf_mixnet_m	76.950 (23.050)	93.156 (6.844)	5.01	bicubic	224
tf_efficientnet_b0	76.848 (23.152)	93.228 (6.772)	5.29	bicubic	224
tf_mixnet_s *tfp	75.800 (24.200)	92.788 (7.212)	4.13	bilinear	224
tf_mobilenetv3_large_100 *tfp	75.768 (24.232)	92.710 (7.290)	5.48	bilinear	224
tf_mixnet_s	75.648 (24.352)	92.636 (7.364)	4.13	bicubic	224
tf_mobilenetv3_large_100	75.516 (24.484)	92.600 (7.400)	5.48	bilinear	224
gluon_resnet34_v1b	74.580 (25.420)	91.988 (8.012)	21.80	bicubic	224
tf_mobilenetv3_large_075 *tfp	73.730 (26.270)	91.616 (8.384)	3.99	bilinear	224
tf_mobilenetv3_large_075	73.442 (26.558)	91.352 (8.648)	3.99	bilinear	224
tf_mobilenetv3_large_minimal_100 *tfp	72.678 (27.322)	90.860 (9.140)	3.92	bilinear	224
tf_mobilenetv3_large_minimal_100	72.244 (27.756)	90.636 (9.364)	3.92	bilinear	224
tf_mobilenetv3_small_100 *tfp	67.918 (32.082)	87.958 (12.042	2.54	bilinear	224
tf_mobilenetv3_small_100	67.918 (32.082)	87.662 (12.338)	2.54	bilinear	224
tf_mobilenetv3_small_075 *tfp	66.142 (33.858)	86.498 (13.502)	2.04	bilinear	224
tf_mobilenetv3_small_075	65.718 (34.282)	86.136 (13.864)	2.04	bilinear	224
tf_mobilenetv3_small_minimal_100 *tfp	63.378 (36.622)	84.802 (15.198)	2.04	bilinear	224
tf_mobilenetv3_small_minimal_100	62.898 (37.102)	84.230 (15.770)	2.04	bilinear	224

Models with *tfp next to them were scored with --tf-preprocessing flag.

The tf_efficientnet, tf_mixnet models require an equivalent for 'SAME' padding as their arch results in asymmetric padding. I've added this in the model creation wrapper, but it does come with a performance penalty.

Sources for original weights:

tf_efficientnet*: Tensorflow TPU
tf_efficientnet_e*: Tensorflow TPU
tf_mixnet*: Tensorflow TPU
tf_inception*: Tensorflow Slim
gluon_*: MxNet Gluon

Training Hyperparameters

EfficientNet-B2 with RandAugment - 80.4 top-1, 95.1 top-5

These params are for dual Titan RTX cards with NVIDIA Apex installed:

./distributed_train.sh 2 /imagenet/ --model efficientnet_b2 -b 128 --sched step --epochs 450 --decay-epochs 2.4 --decay-rate .97 --opt rmsproptf --opt-eps .001 -j 8 --warmup-lr 1e-6 --weight-decay 1e-5 --drop 0.3 --drop-connect 0.2 --model-ema --model-ema-decay 0.9999 --aa rand-m9-mstd0.5 --remode pixel --reprob 0.2 --amp --lr .016

MixNet-XL with RandAugment - 80.5 top-1, 94.9 top-5

This params are for dual Titan RTX cards with NVIDIA Apex installed:

./distributed_train.sh 2 /imagenet/ --model mixnet_xl -b 128 --sched step --epochs 450 --decay-epochs 2.4 --decay-rate .969 --opt rmsproptf --opt-eps .001 -j 8 --warmup-lr 1e-6 --weight-decay 1e-5 --drop 0.3 --drop-connect 0.2 --model-ema --model-ema-decay 0.9999 --aa rand-m9-mstd0.5 --remode pixel --reprob 0.3 --amp --lr .016 --dist-bn reduce

SE-ResNeXt-26-D and SE-ResNeXt-26-T

These hparams (or similar) work well for a wide range of ResNet architecture, generally a good idea to increase the epoch # as the model size increases... ie approx 180-200 for ResNe(X)t50, and 220+ for larger. Increase batch size and LR proportionally for better GPUs or with AMP enabled. These params were for 2 1080Ti cards:

./distributed_train.sh 2 /imagenet/ --model seresnext26t_32x4d --lr 0.1 --warmup-epochs 5 --epochs 160 --weight-decay 1e-4 --sched cosine --reprob 0.4 --remode pixel -b 112

EfficientNet-B3 with RandAugment - 81.5 top-1, 95.7 top-5

The training of this model started with the same command line as EfficientNet-B2 w/ RA above. After almost three weeks of training the process crashed. The results weren't looking amazing so I resumed the training several times with tweaks to a few params (increase RE prob, decrease rand-aug, increase ema-decay). Nothing looked great. I ended up averaging the best checkpoints from all restarts. The result is mediocre at default res/crop but oddly performs much better with a full image test crop of 1.0.

EfficientNet-B0 with RandAugment - 77.7 top-1, 95.3 top-5

Michael Klachko achieved these results with the command line for B2 adapted for larger batch size, with the recommended B0 dropout rate of 0.2.

./distributed_train.sh 2 /imagenet/ --model efficientnet_b0 -b 384 --sched step --epochs 450 --decay-epochs 2.4 --decay-rate .97 --opt rmsproptf --opt-eps .001 -j 8 --warmup-lr 1e-6 --weight-decay 1e-5 --drop 0.2 --drop-connect 0.2 --model-ema --model-ema-decay 0.9999 --aa rand-m9-mstd0.5 --remode pixel --reprob 0.2 --amp --lr .048

ResNet50 with JSD loss and RandAugment (clean + 2x RA augs) - 79.04 top-1, 94.39 top-5

Trained on two older 1080Ti cards, this took a while. Only slightly, non statistically better ImageNet validation result than my first good AugMix training of 78.99. However, these weights are more robust on tests with ImageNetV2, ImageNet-Sketch, etc. Unlike my first AugMix runs, I've enabled SplitBatchNorm, disabled random erasing on the clean split, and cranked up random erasing prob on the 2 augmented paths.

./distributed_train.sh 2 /imagenet -b 64 --model resnet50 --sched cosine --epochs 200 --lr 0.05 --amp --remode pixel --reprob 0.6 --aug-splits 3 --aa rand-m9-mstd0.5-inc1 --resplit --split-bn --jsd --dist-bn reduce

TODO dig up some more

Usage

Environment

All development and testing has been done in Conda Python 3 environments on Linux x86-64 systems, specifically Python 3.6.x and 3.7.x. Little to no care has been taken to be Python 2.x friendly and I don't plan to support it. If you run into any challenges running on Windows, or other OS, I'm definitely open to looking into those issues so long as it's in a reproducible (read Conda) environment.

PyTorch versions 1.2 and 1.3.1 have been tested with this code.

I've tried to keep the dependencies minimal, the setup is as per the PyTorch default install instructions for Conda:

conda create -n torch-env
conda activate torch-env
conda install -c pytorch pytorch torchvision cudatoolkit=10
conda install pyyaml

Pip

This package can be installed via pip. Currently, the model factory (timm.create_model) is the most useful component to use via a pip install.

Install (after conda env/install):

pip install timm

Use:

>>> import timm
>>> m = timm.create_model('mobilenetv3_100', pretrained=True)
>>> m.eval()

Scripts

A train, validation, inference, and checkpoint cleaning script included in the github root folder. Scripts are not currently packaged in the pip release.

Training

The variety of training args is large and not all combinations of options (or even options) have been fully tested. For the training dataset folder, specify the folder to the base that contains a train and validation folder.

To train an SE-ResNet34 on ImageNet, locally distributed, 4 GPUs, one process per GPU w/ cosine schedule, random-erasing prob of 50% and per-pixel random value:

./distributed_train.sh 4 /data/imagenet --model seresnet34 --sched cosine --epochs 150 --warmup-epochs 5 --lr 0.4 --reprob 0.5 --remode pixel --batch-size 256 -j 4

NOTE: NVIDIA APEX should be installed to run in per-process distributed via DDP or to enable AMP mixed precision with the --amp flag

Validation / Inference

Validation and inference scripts are similar in usage. One outputs metrics on a validation set and the other outputs topk class ids in a csv. Specify the folder containing validation images, not the base as in training script.

To validate with the model's pretrained weights (if they exist):

python validate.py /imagenet/validation/ --model seresnext26_32x4d --pretrained

To run inference from a checkpoint:

python inference.py /imagenet/validation/ --model mobilenetv3_100 --checkpoint ./output/model_best.pth.tar

TODO

A number of additions planned in the future for various projects, incl

Do a model performance (speed + accuracy) benchmarking across all models (make runable as script)
Complete feature map extraction across all model types and build obj detection/segmentation models and scripts (or integrate backbones with mmdetection, detectron2)