- Sponsors
- What's New
- Introduction
- Models
- Features
- Results
- Getting Started (Documentation)
- Train, Validation, Inference Scripts
- Awesome PyTorch Resources
- Licenses
- Citing
Thanks to the following for hardware support:
- TPU Research Cloud (TRC) (https://sites.research.google/trc/about/)
- Nvidia (https://www.nvidia.com/en-us/)
And a big thanks to all GitHub sponsors who helped with some of my costs before I joined Hugging Face.
For a few months now, timm
has been part of the Hugging Face ecosystem. Yearly, we survey users of our tools to see what we could do better, what we need to continue doing, or what we need to stop doing.
If you have a couple of minutes and want to participate in shaping the future of the ecosystem, please share your thoughts: hf.co/oss-survey 🙏
- ConvNeXt-V2 models and weights added to existing
convnext.py
- Paper: ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders
- Reference impl: https://github.com/facebookresearch/ConvNeXt-V2 (NOTE: weights currently CC-BY-NC)
- Add FlexiViT models and weights from https://github.com/google-research/big_vision (check out paper at https://arxiv.org/abs/2212.08013)
- NOTE currently resizing is static on model creation, on-the-fly dynamic / train patch size sampling is a WIP
- Many more models updated to multi-weight and downloadable via HF hub now (convnext, efficientnet, mobilenet, vision_transformer*, beit)
- More model pretrained tag and adjustments, some model names changed (working on deprecation translations, consider main branch DEV branch right now, use 0.6.x for stable use)
- More ImageNet-12k (subset of 22k) pretrain models popping up:
efficientnet_b5.in12k_ft_in1k
- 85.9 @ 448x448vit_medium_patch16_gap_384.in12k_ft_in1k
- 85.5 @ 384x384vit_medium_patch16_gap_256.in12k_ft_in1k
- 84.5 @ 256x256convnext_nano.in12k_ft_in1k
- 82.9 @ 288x288
- Add 'EVA l' to
vision_transformer.py
, MAE style ViT-L/14 MIM pretrain w/ EVA-CLIP targets, FT on ImageNet-1k (w/ ImageNet-22k intermediate for some)- original source: https://github.com/baaivision/EVA
model | top1 | param_count | gmac | macts | hub |
---|---|---|---|---|---|
eva_large_patch14_336.in22k_ft_in22k_in1k | 89.2 | 304.5 | 191.1 | 270.2 | link |
eva_large_patch14_336.in22k_ft_in1k | 88.7 | 304.5 | 191.1 | 270.2 | link |
eva_large_patch14_196.in22k_ft_in22k_in1k | 88.6 | 304.1 | 61.6 | 63.5 | link |
eva_large_patch14_196.in22k_ft_in1k | 87.9 | 304.1 | 61.6 | 63.5 | link |
- Add 'EVA g', BEiT style ViT-g/14 model weights w/ both MIM pretrain and CLIP pretrain to
beit.py
.- original source: https://github.com/baaivision/EVA
- paper: https://arxiv.org/abs/2211.07636
model | top1 | param_count | gmac | macts | hub |
---|---|---|---|---|---|
eva_giant_patch14_560.m30m_ft_in22k_in1k | 89.8 | 1014.4 | 1906.8 | 2577.2 | link |
eva_giant_patch14_336.m30m_ft_in22k_in1k | 89.6 | 1013 | 620.6 | 550.7 | link |
eva_giant_patch14_336.clip_ft_in1k | 89.4 | 1013 | 620.6 | 550.7 | link |
eva_giant_patch14_224.clip_ft_in1k | 89.1 | 1012.6 | 267.2 | 192.6 | link |
- Pre-release (
0.8.0dev0
) of multi-weight support (model_arch.pretrained_tag
). Install withpip install --pre timm
- vision_transformer, maxvit, convnext are the first three model impl w/ support
- model names are changing with this (previous _21k, etc. fn will merge), still sorting out deprecation handling
- bugs are likely, but I need feedback so please try it out
- if stability is needed, please use 0.6.x pypi releases or clone from 0.6.x branch
- Support for PyTorch 2.0 compile is added in train/validate/inference/benchmark, use
--torchcompile
argument - Inference script allows more control over output, select k for top-class index + prob json, csv or parquet output
- Add a full set of fine-tuned CLIP image tower weights from both LAION-2B and original OpenAI CLIP models
model | top1 | param_count | gmac | macts | hub |
---|---|---|---|---|---|
vit_huge_patch14_clip_336.laion2b_ft_in12k_in1k | 88.6 | 632.5 | 391 | 407.5 | link |
vit_large_patch14_clip_336.openai_ft_in12k_in1k | 88.3 | 304.5 | 191.1 | 270.2 | link |
vit_huge_patch14_clip_224.laion2b_ft_in12k_in1k | 88.2 | 632 | 167.4 | 139.4 | link |
vit_large_patch14_clip_336.laion2b_ft_in12k_in1k | 88.2 | 304.5 | 191.1 | 270.2 | link |
vit_large_patch14_clip_224.openai_ft_in12k_in1k | 88.2 | 304.2 | 81.1 | 88.8 | link |
vit_large_patch14_clip_224.laion2b_ft_in12k_in1k | 87.9 | 304.2 | 81.1 | 88.8 | link |
vit_large_patch14_clip_224.openai_ft_in1k | 87.9 | 304.2 | 81.1 | 88.8 | link |
vit_large_patch14_clip_336.laion2b_ft_in1k | 87.9 | 304.5 | 191.1 | 270.2 | link |
vit_huge_patch14_clip_224.laion2b_ft_in1k | 87.6 | 632 | 167.4 | 139.4 | link |
vit_large_patch14_clip_224.laion2b_ft_in1k | 87.3 | 304.2 | 81.1 | 88.8 | link |
vit_base_patch16_clip_384.laion2b_ft_in12k_in1k | 87.2 | 86.9 | 55.5 | 101.6 | link |
vit_base_patch16_clip_384.openai_ft_in12k_in1k | 87 | 86.9 | 55.5 | 101.6 | link |
vit_base_patch16_clip_384.laion2b_ft_in1k | 86.6 | 86.9 | 55.5 | 101.6 | link |
vit_base_patch16_clip_384.openai_ft_in1k | 86.2 | 86.9 | 55.5 | 101.6 | link |
vit_base_patch16_clip_224.laion2b_ft_in12k_in1k | 86.2 | 86.6 | 17.6 | 23.9 | link |
vit_base_patch16_clip_224.openai_ft_in12k_in1k | 85.9 | 86.6 | 17.6 | 23.9 | link |
vit_base_patch32_clip_448.laion2b_ft_in12k_in1k | 85.8 | 88.3 | 17.9 | 23.9 | link |
vit_base_patch16_clip_224.laion2b_ft_in1k | 85.5 | 86.6 | 17.6 | 23.9 | link |
vit_base_patch32_clip_384.laion2b_ft_in12k_in1k | 85.4 | 88.3 | 13.1 | 16.5 | link |
vit_base_patch16_clip_224.openai_ft_in1k | 85.3 | 86.6 | 17.6 | 23.9 | link |
vit_base_patch32_clip_384.openai_ft_in12k_in1k | 85.2 | 88.3 | 13.1 | 16.5 | link |
vit_base_patch32_clip_224.laion2b_ft_in12k_in1k | 83.3 | 88.2 | 4.4 | 5 | link |
vit_base_patch32_clip_224.laion2b_ft_in1k | 82.6 | 88.2 | 4.4 | 5 | link |
vit_base_patch32_clip_224.openai_ft_in1k | 81.9 | 88.2 | 4.4 | 5 | link |
- Port of MaxViT Tensorflow Weights from official impl at https://github.com/google-research/maxvit
- There was larger than expected drops for the upscaled 384/512 in21k fine-tune weights, possible detail missing, but the 21k FT did seem sensitive to small preprocessing
model | top1 | param_count | gmac | macts | hub |
---|---|---|---|---|---|
maxvit_xlarge_tf_512.in21k_ft_in1k | 88.5 | 475.8 | 534.1 | 1413.2 | link |
maxvit_xlarge_tf_384.in21k_ft_in1k | 88.3 | 475.3 | 292.8 | 668.8 | link |
maxvit_base_tf_512.in21k_ft_in1k | 88.2 | 119.9 | 138 | 704 | link |
maxvit_large_tf_512.in21k_ft_in1k | 88 | 212.3 | 244.8 | 942.2 | link |
maxvit_large_tf_384.in21k_ft_in1k | 88 | 212 | 132.6 | 445.8 | link |
maxvit_base_tf_384.in21k_ft_in1k | 87.9 | 119.6 | 73.8 | 332.9 | link |
maxvit_base_tf_512.in1k | 86.6 | 119.9 | 138 | 704 | link |
maxvit_large_tf_512.in1k | 86.5 | 212.3 | 244.8 | 942.2 | link |
maxvit_base_tf_384.in1k | 86.3 | 119.6 | 73.8 | 332.9 | link |
maxvit_large_tf_384.in1k | 86.2 | 212 | 132.6 | 445.8 | link |
maxvit_small_tf_512.in1k | 86.1 | 69.1 | 67.3 | 383.8 | link |
maxvit_tiny_tf_512.in1k | 85.7 | 31 | 33.5 | 257.6 | link |
maxvit_small_tf_384.in1k | 85.5 | 69 | 35.9 | 183.6 | link |
maxvit_tiny_tf_384.in1k | 85.1 | 31 | 17.5 | 123.4 | link |
maxvit_large_tf_224.in1k | 84.9 | 211.8 | 43.7 | 127.4 | link |
maxvit_base_tf_224.in1k | 84.9 | 119.5 | 24 | 95 | link |
maxvit_small_tf_224.in1k | 84.4 | 68.9 | 11.7 | 53.2 | link |
maxvit_tiny_tf_224.in1k | 83.4 | 30.9 | 5.6 | 35.8 | link |
- Train and validation script enhancements
- Non-GPU (ie CPU) device support
- SLURM compatibility for train script
- HF datasets support (via ReaderHfds)
- TFDS/WDS dataloading improvements (sample padding/wrap for distributed use fixed wrt sample count estimate)
- in_chans !=3 support for scripts / loader
- Adan optimizer
- Can enable per-step LR scheduling via args
- Dataset 'parsers' renamed to 'readers', more descriptive of purpose
- AMP args changed, APEX via
--amp-impl apex
, bfloat16 supportedf via--amp-dtype bfloat16
- main branch switched to 0.7.x version, 0.6x forked for stable release of weight only adds
- master -> main branch rename
- More weights in
maxxvit
series, incl first ConvNeXt block basedcoatnext
andmaxxvit
experiments:coatnext_nano_rw_224
- 82.0 @ 224 (G) -- (uses ConvNeXt conv block, no BatchNorm)maxxvit_rmlp_nano_rw_256
- 83.0 @ 256, 83.7 @ 320 (G) (uses ConvNeXt conv block, no BN)maxvit_rmlp_small_rw_224
- 84.5 @ 224, 85.1 @ 320 (G)maxxvit_rmlp_small_rw_256
- 84.6 @ 256, 84.9 @ 288 (G) -- could be trained better, hparams need tuning (uses ConvNeXt block, no BN)coatnet_rmlp_2_rw_224
- 84.6 @ 224, 85 @ 320 (T)- NOTE: official MaxVit weights (in1k) have been released at https://github.com/google-research/maxvit -- some extra work is needed to port and adapt since my impl was created independently of theirs and has a few small differences + the whole TF same padding fun.
- LAION-2B CLIP image towers supported as pretrained backbones for fine-tune or features (no classifier)
- vit_base_patch32_224_clip_laion2b
- vit_large_patch14_224_clip_laion2b
- vit_huge_patch14_224_clip_laion2b
- vit_giant_patch14_224_clip_laion2b
- Hugging Face
timm
docs home now exists, look for more here in the future - Add BEiT-v2 weights for base and large 224x224 models from https://github.com/microsoft/unilm/tree/master/beit2
- Add more weights in
maxxvit
series incl apico
(7.5M params, 1.9 GMACs), twotiny
variants:maxvit_rmlp_pico_rw_256
- 80.5 @ 256, 81.3 @ 320 (T)maxvit_tiny_rw_224
- 83.5 @ 224 (G)maxvit_rmlp_tiny_rw_256
- 84.2 @ 256, 84.8 @ 320 (T)
- MaxVit window size scales with img_size by default. Add new RelPosMlp MaxViT weight that leverages this:
maxvit_rmlp_nano_rw_256
- 83.0 @ 256, 83.6 @ 320 (T)
- CoAtNet (https://arxiv.org/abs/2106.04803) and MaxVit (https://arxiv.org/abs/2204.01697)
timm
original models- both found in
maxxvit.py
model def, contains numerous experiments outside scope of original papers - an unfinished Tensorflow version from MaxVit authors can be found https://github.com/google-research/maxvit
- both found in
- Initial CoAtNet and MaxVit timm pretrained weights (working on more):
coatnet_nano_rw_224
- 81.7 @ 224 (T)coatnet_rmlp_nano_rw_224
- 82.0 @ 224, 82.8 @ 320 (T)coatnet_0_rw_224
- 82.4 (T) -- NOTE timm '0' coatnets have 2 more 3rd stage blockscoatnet_bn_0_rw_224
- 82.4 (T)maxvit_nano_rw_256
- 82.9 @ 256 (T)coatnet_rmlp_1_rw_224
- 83.4 @ 224, 84 @ 320 (T)coatnet_1_rw_224
- 83.6 @ 224 (G)- (T) = TPU trained with
bits_and_tpu
branch training code, (G) = GPU trained
- GCVit (weights adapted from https://github.com/NVlabs/GCVit, code 100%
timm
re-write for license purposes) - MViT-V2 (multi-scale vit, adapted from https://github.com/facebookresearch/mvit)
- EfficientFormer (adapted from https://github.com/snap-research/EfficientFormer)
- PyramidVisionTransformer-V2 (adapted from https://github.com/whai362/PVT)
- 'Fast Norm' support for LayerNorm and GroupNorm that avoids float32 upcast w/ AMP (uses APEX LN if available for further boost)
- ConvNeXt atto weights added
convnext_atto
- 75.7 @ 224, 77.0 @ 288convnext_atto_ols
- 75.9 @ 224, 77.2 @ 288
- More custom ConvNeXt smaller model defs with weights
convnext_femto
- 77.5 @ 224, 78.7 @ 288convnext_femto_ols
- 77.9 @ 224, 78.9 @ 288convnext_pico
- 79.5 @ 224, 80.4 @ 288convnext_pico_ols
- 79.5 @ 224, 80.5 @ 288convnext_nano_ols
- 80.9 @ 224, 81.6 @ 288
- Updated EdgeNeXt to improve ONNX export, add new base variant and weights from original (https://github.com/mmaaz60/EdgeNeXt)
- Add freshly minted DeiT-III Medium (width=512, depth=12, num_heads=8) model weights. Thanks Hugo Touvron!
- All runtime benchmark and validation result csv files are finally up-to-date!
- A few more weights & model defs added:
darknetaa53
- 79.8 @ 256, 80.5 @ 288convnext_nano
- 80.8 @ 224, 81.5 @ 288cs3sedarknet_l
- 81.2 @ 256, 81.8 @ 288cs3darknet_x
- 81.8 @ 256, 82.2 @ 288cs3sedarknet_x
- 82.2 @ 256, 82.7 @ 288cs3edgenet_x
- 82.2 @ 256, 82.7 @ 288cs3se_edgenet_x
- 82.8 @ 256, 83.5 @ 320
cs3*
weights above all trained on TPU w/bits_and_tpu
branch. Thanks to TRC program!- Add output_stride=8 and 16 support to ConvNeXt (dilation)
- deit3 models not being able to resize pos_emb fixed
- Version 0.6.7 PyPi release (/w above bug fixes and new weighs since 0.6.5)
More models, more fixes
- Official research models (w/ weights) added:
- EdgeNeXt from (https://github.com/mmaaz60/EdgeNeXt)
- MobileViT-V2 from (https://github.com/apple/ml-cvnets)
- DeiT III (Revenge of the ViT) from (https://github.com/facebookresearch/deit)
- My own models:
- Small
ResNet
defs added by request with 1 block repeats for both basic and bottleneck (resnet10 and resnet14) CspNet
refactored with dataclass config, simplified CrossStage3 (cs3
) option. These are closer to YOLO-v5+ backbone defs.- More relative position vit fiddling. Two
srelpos
(shared relative position) models trained, and a medium w/ class token. - Add an alternate downsample mode to EdgeNeXt and train a
small
model. Better than original small, but not their new USI trained weights.
- Small
- My own model weight results (all ImageNet-1k training)
resnet10t
- 66.5 @ 176, 68.3 @ 224resnet14t
- 71.3 @ 176, 72.3 @ 224resnetaa50
- 80.6 @ 224 , 81.6 @ 288darknet53
- 80.0 @ 256, 80.5 @ 288cs3darknet_m
- 77.0 @ 256, 77.6 @ 288cs3darknet_focus_m
- 76.7 @ 256, 77.3 @ 288cs3darknet_l
- 80.4 @ 256, 80.9 @ 288cs3darknet_focus_l
- 80.3 @ 256, 80.9 @ 288vit_srelpos_small_patch16_224
- 81.1 @ 224, 82.1 @ 320vit_srelpos_medium_patch16_224
- 82.3 @ 224, 83.1 @ 320vit_relpos_small_patch16_cls_224
- 82.6 @ 224, 83.6 @ 320edgnext_small_rw
- 79.6 @ 224, 80.4 @ 320
cs3
,darknet
, andvit_*relpos
weights above all trained on TPU thanks to TRC program! Rest trained on overheating GPUs.- Hugging Face Hub support fixes verified, demo notebook TBA
- Pretrained weights / configs can be loaded externally (ie from local disk) w/ support for head adaptation.
- Add support to change image extensions scanned by
timm
datasets/readers. See (huggingface#1274 (comment)) - Default ConvNeXt LayerNorm impl to use
F.layer_norm(x.permute(0, 2, 3, 1), ...).permute(0, 3, 1, 2)
viaLayerNorm2d
in all cases.- a bit slower than previous custom impl on some hardware (ie Ampere w/ CL), but overall fewer regressions across wider HW / PyTorch version ranges.
- previous impl exists as
LayerNormExp2d
inmodels/layers/norm.py
- Numerous bug fixes
- Currently testing for imminent PyPi 0.6.x release
- LeViT pretraining of larger models still a WIP, they don't train well / easily without distillation. Time to add distill support (finally)?
- ImageNet-22k weight training + finetune ongoing, work on multi-weight support (slowly) chugging along (there are a LOT of weights, sigh) ...
- Official Swin-V2 models and weights added from (https://github.com/microsoft/Swin-Transformer). Cleaned up to support torchscript.
- Some refactoring for existing
timm
Swin-V2-CR impl, will likely do a bit more to bring parts closer to official and decide whether to merge some aspects. - More Vision Transformer relative position / residual post-norm experiments (all trained on TPU thanks to TRC program)
vit_relpos_small_patch16_224
- 81.5 @ 224, 82.5 @ 320 -- rel pos, layer scale, no class token, avg poolvit_relpos_medium_patch16_rpn_224
- 82.3 @ 224, 83.1 @ 320 -- rel pos + res-post-norm, no class token, avg poolvit_relpos_medium_patch16_224
- 82.5 @ 224, 83.3 @ 320 -- rel pos, layer scale, no class token, avg poolvit_relpos_base_patch16_gapcls_224
- 82.8 @ 224, 83.9 @ 320 -- rel pos, layer scale, class token, avg pool (by mistake)
- Bring 512 dim, 8-head 'medium' ViT model variant back to life (after using in a pre DeiT 'small' model for first ViT impl back in 2020)
- Add ViT relative position support for switching btw existing impl and some additions in official Swin-V2 impl for future trials
- Sequencer2D impl (https://arxiv.org/abs/2205.01972), added via PR from author (https://github.com/okojoalg)
- Vision Transformer experiments adding Relative Position (Swin-V2 log-coord) (
vision_transformer_relpos.py
) and Residual Post-Norm branches (from Swin-V2) (vision_transformer*.py
)vit_relpos_base_patch32_plus_rpn_256
- 79.5 @ 256, 80.6 @ 320 -- rel pos + extended width + res-post-norm, no class token, avg poolvit_relpos_base_patch16_224
- 82.5 @ 224, 83.6 @ 320 -- rel pos, layer scale, no class token, avg poolvit_base_patch16_rpn_224
- 82.3 @ 224 -- rel pos + res-post-norm, no class token, avg pool
- Vision Transformer refactor to remove representation layer that was only used in initial vit and rarely used since with newer pretrain (ie
How to Train Your ViT
) vit_*
models support removal of class token, use of global average pool, use of fc_norm (ala beit, mae).
timm
models are now officially supported in fast.ai! Just in time for the new Practical Deep Learning course.timmdocs
documentation link updated to timm.fast.ai.- Two more model weights added in the TPU trained series. Some In22k pretrain still in progress.
seresnext101d_32x8d
- 83.69 @ 224, 84.35 @ 288seresnextaa101d_32x8d
(anti-aliased w/ AvgPool2d) - 83.85 @ 224, 84.57 @ 288
- Add
ParallelBlock
andLayerScale
option to base vit models to support model configs in Three things everyone should know about ViT convnext_tiny_hnf
(head norm first) weights trained with (close to) A2 recipe, 82.2% top-1, could do better with more epochs.
- Merge
norm_norm_norm
. IMPORTANT this update for a coming 0.6.x release will likely de-stabilize the master branch for a while. Branch0.5.x
or a previous 0.5.x release can be used if stability is required. - Significant weights update (all TPU trained) as described in this release
regnety_040
- 82.3 @ 224, 82.96 @ 288regnety_064
- 83.0 @ 224, 83.65 @ 288regnety_080
- 83.17 @ 224, 83.86 @ 288regnetv_040
- 82.44 @ 224, 83.18 @ 288 (timm pre-act)regnetv_064
- 83.1 @ 224, 83.71 @ 288 (timm pre-act)regnetz_040
- 83.67 @ 256, 84.25 @ 320regnetz_040h
- 83.77 @ 256, 84.5 @ 320 (w/ extra fc in head)resnetv2_50d_gn
- 80.8 @ 224, 81.96 @ 288 (pre-act GroupNorm)resnetv2_50d_evos
80.77 @ 224, 82.04 @ 288 (pre-act EvoNormS)regnetz_c16_evos
- 81.9 @ 256, 82.64 @ 320 (EvoNormS)regnetz_d8_evos
- 83.42 @ 256, 84.04 @ 320 (EvoNormS)xception41p
- 82 @ 299 (timm pre-act)xception65
- 83.17 @ 299xception65p
- 83.14 @ 299 (timm pre-act)resnext101_64x4d
- 82.46 @ 224, 83.16 @ 288seresnext101_32x8d
- 83.57 @ 224, 84.270 @ 288resnetrs200
- 83.85 @ 256, 84.44 @ 320
- HuggingFace hub support fixed w/ initial groundwork for allowing alternative 'config sources' for pretrained model definitions and weights (generic local file / remote url support soon)
- SwinTransformer-V2 implementation added. Submitted by Christoph Reich. Training experiments and model changes by myself are ongoing so expect compat breaks.
- Swin-S3 (AutoFormerV2) models / weights added from https://github.com/microsoft/Cream/tree/main/AutoFormerV2
- MobileViT models w/ weights adapted from https://github.com/apple/ml-cvnets
- PoolFormer models w/ weights adapted from https://github.com/sail-sg/poolformer
- VOLO models w/ weights adapted from https://github.com/sail-sg/volo
- Significant work experimenting with non-BatchNorm norm layers such as EvoNorm, FilterResponseNorm, GroupNorm, etc
- Enhance support for alternate norm + act ('NormAct') layers added to a number of models, esp EfficientNet/MobileNetV3, RegNet, and aligned Xception
- Grouped conv support added to EfficientNet family
- Add 'group matching' API to all models to allow grouping model parameters for application of 'layer-wise' LR decay, lr scale added to LR scheduler
- Gradient checkpointing support added to many models
forward_head(x, pre_logits=False)
fn added to all models to allow separate calls offorward_features
+forward_head
- All vision transformer and vision MLP models update to return non-pooled / non-token selected features from
foward_features
, for consistency with CNN models, token selection or pooling now applied inforward_head
- Chris Hughes posted an exhaustive run through of
timm
on his blog yesterday. Well worth a read. Getting Started with PyTorch Image Models (timm): A Practitioner’s Guide - I'm currently prepping to merge the
norm_norm_norm
branch back to master (ver 0.6.x) in next week or so.- The changes are more extensive than usual and may destabilize and break some model API use (aiming for full backwards compat). So, beware
pip install git+https://github.com/rwightman/pytorch-image-models
installs! 0.5.x
releases and a0.5.x
branch will remain stable with a cherry pick or two until dust clears. Recommend sticking to pypi install for a bit if you want stable.
- The changes are more extensive than usual and may destabilize and break some model API use (aiming for full backwards compat). So, beware
- Version 0.5.4 w/ release to be pushed to pypi. It's been a while since last pypi update and riskier changes will be merged to main branch soon....
- Add ConvNeXT models /w weights from official impl (https://github.com/facebookresearch/ConvNeXt), a few perf tweaks, compatible with timm features
- Tried training a few small (~1.8-3M param) / mobile optimized models, a few are good so far, more on the way...
mnasnet_small
- 65.6 top-1mobilenetv2_050
- 65.9lcnet_100/075/050
- 72.1 / 68.8 / 63.1semnasnet_075
- 73fbnetv3_b/d/g
- 79.1 / 79.7 / 82.0
- TinyNet models added by rsomani95
- LCNet added via MobileNetV3 architecture
PyTorch Image Models (timm
) is a collection of image models, layers, utilities, optimizers, schedulers, data-loaders / augmentations, and reference training / validation scripts that aim to pull together a wide variety of SOTA models with ability to reproduce ImageNet training results.
The work of many others is present here. I've tried to make sure all source material is acknowledged via links to github, arxiv papers, etc in the README, documentation, and code docstrings. Please let me know if I missed anything.
All model architecture families include variants with pretrained weights. There are specific model variants without any weights, it is NOT a bug. Help training new or better weights is always appreciated. Here are some example training hparams to get you started.
A full version of the list below with source links can be found in the documentation.
- Aggregating Nested Transformers - https://arxiv.org/abs/2105.12723
- BEiT - https://arxiv.org/abs/2106.08254
- Big Transfer ResNetV2 (BiT) - https://arxiv.org/abs/1912.11370
- Bottleneck Transformers - https://arxiv.org/abs/2101.11605
- CaiT (Class-Attention in Image Transformers) - https://arxiv.org/abs/2103.17239
- CoaT (Co-Scale Conv-Attentional Image Transformers) - https://arxiv.org/abs/2104.06399
- CoAtNet (Convolution and Attention) - https://arxiv.org/abs/2106.04803
- ConvNeXt - https://arxiv.org/abs/2201.03545
- ConvNeXt-V2 - http://arxiv.org/abs/2301.00808
- ConViT (Soft Convolutional Inductive Biases Vision Transformers)- https://arxiv.org/abs/2103.10697
- CspNet (Cross-Stage Partial Networks) - https://arxiv.org/abs/1911.11929
- DeiT - https://arxiv.org/abs/2012.12877
- DeiT-III - https://arxiv.org/pdf/2204.07118.pdf
- DenseNet - https://arxiv.org/abs/1608.06993
- DLA - https://arxiv.org/abs/1707.06484
- DPN (Dual-Path Network) - https://arxiv.org/abs/1707.01629
- EdgeNeXt - https://arxiv.org/abs/2206.10589
- EfficientFormer - https://arxiv.org/abs/2206.01191
- EfficientNet (MBConvNet Family)
- EfficientNet NoisyStudent (B0-B7, L2) - https://arxiv.org/abs/1911.04252
- EfficientNet AdvProp (B0-B8) - https://arxiv.org/abs/1911.09665
- EfficientNet (B0-B7) - https://arxiv.org/abs/1905.11946
- EfficientNet-EdgeTPU (S, M, L) - https://ai.googleblog.com/2019/08/efficientnet-edgetpu-creating.html
- EfficientNet V2 - https://arxiv.org/abs/2104.00298
- FBNet-C - https://arxiv.org/abs/1812.03443
- MixNet - https://arxiv.org/abs/1907.09595
- MNASNet B1, A1 (Squeeze-Excite), and Small - https://arxiv.org/abs/1807.11626
- MobileNet-V2 - https://arxiv.org/abs/1801.04381
- Single-Path NAS - https://arxiv.org/abs/1904.02877
- TinyNet - https://arxiv.org/abs/2010.14819
- EVA - https://arxiv.org/abs/2211.07636
- FlexiViT - https://arxiv.org/abs/2212.08013
- GCViT (Global Context Vision Transformer) - https://arxiv.org/abs/2206.09959
- GhostNet - https://arxiv.org/abs/1911.11907
- gMLP - https://arxiv.org/abs/2105.08050
- GPU-Efficient Networks - https://arxiv.org/abs/2006.14090
- Halo Nets - https://arxiv.org/abs/2103.12731
- HRNet - https://arxiv.org/abs/1908.07919
- Inception-V3 - https://arxiv.org/abs/1512.00567
- Inception-ResNet-V2 and Inception-V4 - https://arxiv.org/abs/1602.07261
- Lambda Networks - https://arxiv.org/abs/2102.08602
- LeViT (Vision Transformer in ConvNet's Clothing) - https://arxiv.org/abs/2104.01136
- MaxViT (Multi-Axis Vision Transformer) - https://arxiv.org/abs/2204.01697
- MLP-Mixer - https://arxiv.org/abs/2105.01601
- MobileNet-V3 (MBConvNet w/ Efficient Head) - https://arxiv.org/abs/1905.02244
- FBNet-V3 - https://arxiv.org/abs/2006.02049
- HardCoRe-NAS - https://arxiv.org/abs/2102.11646
- LCNet - https://arxiv.org/abs/2109.15099
- MobileViT - https://arxiv.org/abs/2110.02178
- MobileViT-V2 - https://arxiv.org/abs/2206.02680
- MViT-V2 (Improved Multiscale Vision Transformer) - https://arxiv.org/abs/2112.01526
- NASNet-A - https://arxiv.org/abs/1707.07012
- NesT - https://arxiv.org/abs/2105.12723
- NFNet-F - https://arxiv.org/abs/2102.06171
- NF-RegNet / NF-ResNet - https://arxiv.org/abs/2101.08692
- PNasNet - https://arxiv.org/abs/1712.00559
- PoolFormer (MetaFormer) - https://arxiv.org/abs/2111.11418
- Pooling-based Vision Transformer (PiT) - https://arxiv.org/abs/2103.16302
- PVT-V2 (Improved Pyramid Vision Transformer) - https://arxiv.org/abs/2106.13797
- RegNet - https://arxiv.org/abs/2003.13678
- RegNetZ - https://arxiv.org/abs/2103.06877
- RepVGG - https://arxiv.org/abs/2101.03697
- ResMLP - https://arxiv.org/abs/2105.03404
- ResNet/ResNeXt
- ResNet (v1b/v1.5) - https://arxiv.org/abs/1512.03385
- ResNeXt - https://arxiv.org/abs/1611.05431
- 'Bag of Tricks' / Gluon C, D, E, S variations - https://arxiv.org/abs/1812.01187
- Weakly-supervised (WSL) Instagram pretrained / ImageNet tuned ResNeXt101 - https://arxiv.org/abs/1805.00932
- Semi-supervised (SSL) / Semi-weakly Supervised (SWSL) ResNet/ResNeXts - https://arxiv.org/abs/1905.00546
- ECA-Net (ECAResNet) - https://arxiv.org/abs/1910.03151v4
- Squeeze-and-Excitation Networks (SEResNet) - https://arxiv.org/abs/1709.01507
- ResNet-RS - https://arxiv.org/abs/2103.07579
- Res2Net - https://arxiv.org/abs/1904.01169
- ResNeSt - https://arxiv.org/abs/2004.08955
- ReXNet - https://arxiv.org/abs/2007.00992
- SelecSLS - https://arxiv.org/abs/1907.00837
- Selective Kernel Networks - https://arxiv.org/abs/1903.06586
- Sequencer2D - https://arxiv.org/abs/2205.01972
- Swin S3 (AutoFormerV2) - https://arxiv.org/abs/2111.14725
- Swin Transformer - https://arxiv.org/abs/2103.14030
- Swin Transformer V2 - https://arxiv.org/abs/2111.09883
- Transformer-iN-Transformer (TNT) - https://arxiv.org/abs/2103.00112
- TResNet - https://arxiv.org/abs/2003.13630
- Twins (Spatial Attention in Vision Transformers) - https://arxiv.org/pdf/2104.13840.pdf
- Visformer - https://arxiv.org/abs/2104.12533
- Vision Transformer - https://arxiv.org/abs/2010.11929
- VOLO (Vision Outlooker) - https://arxiv.org/abs/2106.13112
- VovNet V2 and V1 - https://arxiv.org/abs/1911.06667
- Xception - https://arxiv.org/abs/1610.02357
- Xception (Modified Aligned, Gluon) - https://arxiv.org/abs/1802.02611
- Xception (Modified Aligned, TF) - https://arxiv.org/abs/1802.02611
- XCiT (Cross-Covariance Image Transformers) - https://arxiv.org/abs/2106.09681
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
andreset_classifier
- doing a forward pass on just the features -
forward_features
(see documentation) - these makes it easy to write consistent network wrappers that work with any of the models
- accessing/changing the classifier -
- All models support multi-scale feature map extraction (feature pyramids) via create_model (see documentation)
create_model(name, features_only=True, out_indices=..., output_stride=...)
out_indices
creation arg specifies which feature maps to return, these indices are 0 based and generally correspond to theC(i + 1)
feature level.output_stride
creation arg controls output stride of the network by using dilated convolutions. Most networks are stride 32 by default. Not all networks support this.- feature map channel counts, reduction level (stride) can be queried AFTER model creation via the
.feature_info
member
- All models have a consistent pretrained weight loader that adapts last linear if necessary, and from 3 to 1 channel input if desired
- High performance reference training, validation, and inference scripts that work 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 provides 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)
- Learning rate schedulers
- Ideas adopted from
- AllenNLP schedulers
- FAIRseq lr_scheduler
- SGDR: Stochastic Gradient Descent with Warm Restarts (https://arxiv.org/abs/1608.03983)
- Schedulers include
step
,cosine
w/ restarts,tanh
w/ restarts,plateau
- Ideas adopted from
- Optimizers:
rmsprop_tf
adapted from PyTorch RMSProp by myself. Reproduces much improved Tensorflow RMSProp behaviour.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)fused<name>
optimizers by name with NVIDIA Apex installedadamp
andsgdp
by Naver ClovAI (https://arxiv.org/abs/2006.08217)adafactor
adapted from FAIRSeq impl (https://arxiv.org/abs/1804.04235)adahessian
by David Samuel (https://arxiv.org/abs/2006.00719)
- Random Erasing from Zhun Zhong (https://arxiv.org/abs/1708.04896)
- Mixup (https://arxiv.org/abs/1710.09412)
- CutMix (https://arxiv.org/abs/1905.04899)
- 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
- DropPath aka "Stochastic Depth" (https://arxiv.org/abs/1603.09382)
- DropBlock (https://arxiv.org/abs/1810.12890)
- Blur Pooling (https://arxiv.org/abs/1904.11486)
- Space-to-Depth by mrT23 (https://arxiv.org/abs/1801.04590) -- original paper?
- Adaptive Gradient Clipping (https://arxiv.org/abs/2102.06171, https://github.com/deepmind/deepmind-research/tree/master/nfnets)
- An extensive selection of channel and/or spatial attention modules:
- Bottleneck Transformer - https://arxiv.org/abs/2101.11605
- CBAM - https://arxiv.org/abs/1807.06521
- Effective Squeeze-Excitation (ESE) - https://arxiv.org/abs/1911.06667
- Efficient Channel Attention (ECA) - https://arxiv.org/abs/1910.03151
- Gather-Excite (GE) - https://arxiv.org/abs/1810.12348
- Global Context (GC) - https://arxiv.org/abs/1904.11492
- Halo - https://arxiv.org/abs/2103.12731
- Involution - https://arxiv.org/abs/2103.06255
- Lambda Layer - https://arxiv.org/abs/2102.08602
- Non-Local (NL) - https://arxiv.org/abs/1711.07971
- Squeeze-and-Excitation (SE) - https://arxiv.org/abs/1709.01507
- Selective Kernel (SK) - (https://arxiv.org/abs/1903.06586
- Split (SPLAT) - https://arxiv.org/abs/2004.08955
- Shifted Window (SWIN) - https://arxiv.org/abs/2103.14030
Model validation results can be found in the documentation and in the results tables
My current documentation for timm
covers the basics.
Hugging Face timm
docs will be the documentation focus going forward and will eventually replace the github.io
docs above.
Getting Started with PyTorch Image Models (timm): A Practitioner’s Guide by Chris Hughes is an extensive blog post covering many aspects of timm
in detail.
timmdocs is quickly becoming a much more comprehensive set of documentation for timm
. A big thanks to Aman Arora for his efforts creating timmdocs.
paperswithcode is a good resource for browsing the models within timm
.
The root folder of the repository contains reference train, validation, and inference scripts that work with the included models and other features of this repository. They are adaptable for other datasets and use cases with a little hacking. See documentation for some basics and training hparams for some train examples that produce SOTA ImageNet results.
One of the greatest assets of PyTorch is the community and their contributions. A few of my favourite resources that pair well with the models and components here are listed below.
- Detectron2 - https://github.com/facebookresearch/detectron2
- Segmentation Models (Semantic) - https://github.com/qubvel/segmentation_models.pytorch
- EfficientDet (Obj Det, Semantic soon) - https://github.com/rwightman/efficientdet-pytorch
- Albumentations - https://github.com/albumentations-team/albumentations
- Kornia - https://github.com/kornia/kornia
- RepDistiller - https://github.com/HobbitLong/RepDistiller
- torchdistill - https://github.com/yoshitomo-matsubara/torchdistill
- PyTorch Metric Learning - https://github.com/KevinMusgrave/pytorch-metric-learning
- fastai - https://github.com/fastai/fastai
The code here is licensed Apache 2.0. I've taken care to make sure any third party code included or adapted has compatible (permissive) licenses such as MIT, BSD, etc. I've made an effort to avoid any GPL / LGPL conflicts. That said, it is your responsibility to ensure you comply with licenses here and conditions of any dependent licenses. Where applicable, I've linked the sources/references for various components in docstrings. If you think I've missed anything please create an issue.
So far all of the pretrained weights available here are pretrained on ImageNet with a select few that have some additional pretraining (see extra note below). ImageNet was released for non-commercial research purposes only (https://image-net.org/download). It's not clear what the implications of that are for the use of pretrained weights from that dataset. Any models I have trained with ImageNet are done for research purposes and one should assume that the original dataset license applies to the weights. It's best to seek legal advice if you intend to use the pretrained weights in a commercial product.
Several weights included or references here were pretrained with proprietary datasets that I do not have access to. These include the Facebook WSL, SSL, SWSL ResNe(Xt) and the Google Noisy Student EfficientNet models. The Facebook models have an explicit non-commercial license (CC-BY-NC 4.0, https://github.com/facebookresearch/semi-supervised-ImageNet1K-models, https://github.com/facebookresearch/WSL-Images). The Google models do not appear to have any restriction beyond the Apache 2.0 license (and ImageNet concerns). In either case, you should contact Facebook or Google with any questions.
@misc{rw2019timm,
author = {Ross Wightman},
title = {PyTorch Image Models},
year = {2019},
publisher = {GitHub},
journal = {GitHub repository},
doi = {10.5281/zenodo.4414861},
howpublished = {\url{https://github.com/rwightman/pytorch-image-models}}
}