I'm fortunate to be able to dedicate significant time and money of my own supporting this and other open source projects. However, as the projects increase in scope, outside support is needed to continue with the current trajectory of cloud services, hardware, and electricity costs.
*`resnet50` - 80.7 @ 224, 80.9 @ 288 (trained at 176, not replacing current a1 weights as default since these don't scale as well to higher res, [weights](https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-rsb-weights/resnet50_a1h2_176-001a1197.pth))
* ResNet strikes back (https://arxiv.org/abs/2110.00476) weights added, plus any extra training components used. Model weights and some more details here (https://github.com/rwightman/pytorch-image-models/releases/tag/v0.1-rsb-weights)
* BCE loss and Repeated Augmentation support for RSB paper
* 4 series of ResNet based attention model experiments being added (implemented across byobnet.py/byoanet.py). These include all sorts of attention, from channel attn like SE, ECA to 2D QKV self-attention layers such as Halo, Bottlneck, Lambda. Details here (https://github.com/rwightman/pytorch-image-models/releases/tag/v0.1-attn-weights)
* Working implementations of the following 2D self-attention modules (likely to be differences from paper or eventual official impl):
* A RegNetZ series of models with some attention experiments (being added to). These do not follow the paper (https://arxiv.org/abs/2103.06877) in any way other than block architecture, details of official models are not available. See more here (https://github.com/rwightman/pytorch-image-models/releases/tag/v0.1-attn-weights)
* Add LAMB and LARS optimizers, incl trust ratio clipping options. Tweaked to work properly in PyTorch XLA (tested on TPUs w/ `timm bits` [branch](https://github.com/rwightman/pytorch-image-models/tree/bits_and_tpu/timm/bits))
* Add MADGRAD from FB research w/ a few tweaks (decoupled decay option, step handling that works with PyTorch XLA)
* Some cleanup on all optimizers and factory. No more `.data`, a bit more consistency, unit tests for all!
* SGDP and AdamP still won't work with PyTorch XLA but others should (have yet to test Adabelief, Adafactor, Adahessian myself).
* EfficientNet-V2 XL TF ported weights added, but they don't validate well in PyTorch (L is better). The pre-processing for the V2 TF training is a bit diff and the fine-tuned 21k -> 1k weights are very sensitive and less robust than the 1k weights.
* Added PyTorch trained EfficientNet-V2 'Tiny' w/ GlobalContext attn weights. Only .1-.2 top-1 better than the SE so more of a curiosity for those interested.
* Add `efficientnetv2_rw_t` weights, a custom 'tiny' 13.6M param variant that is a bit better than (non NoisyStudent) B3 models. Both faster and better accuracy (at same or lower res)
* Add [SAM pretrained](https://arxiv.org/abs/2106.01548) in1k weight for ViT B/16 (`vit_base_patch16_sam_224`) and B/32 (`vit_base_patch32_sam_224`) models.
* Reproduce gMLP model training, `gmlp_s16_224` trained to 79.6 top-1, matching [paper](https://arxiv.org/abs/2105.08050). Hparams for this and other recent MLP training [here](https://gist.github.com/rwightman/d6c264a9001f9167e06c209f630b2cc6)
* Release Vision Transformer 'AugReg' weights from [How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers](https://arxiv.org/abs/2106.10270)
* .npz weight loading support added, can load any of the 50K+ weights from the [AugReg series](https://console.cloud.google.com/storage/browser/vit_models/augreg)
* See [example notebook](https://colab.research.google.com/github/google-research/vision_transformer/blob/master/vit_jax_augreg.ipynb) from [official impl](https://github.com/google-research/vision_transformer/) for navigating the augreg weights
* Remove my old small model, replace with DeiT compatible small w/ AugReg weights
* Add 1st training of my `gmixer_24_224` MLP /w GLU, 78.1 top-1 w/ 25M params.
* Add weights from official ResMLP release (https://github.com/facebookresearch/deit)
* Add `eca_nfnet_l2` weights from my 'lightweight' series. 84.7 top-1 at 384x384.
* Add distilled BiT 50x1 student and 152x2 Teacher weights from [Knowledge distillation: A good teacher is patient and consistent](https://arxiv.org/abs/2106.05237)
* NFNets and ResNetV2-BiT models work w/ Pytorch XLA now
* Add EfficientNet-V2 official model defs w/ ported weights from official [Tensorflow/Keras](https://github.com/google/automl/tree/master/efficientnetv2) impl.
* Add ECA-NFNet-L1 (slimmed down F1 w/ SiLU, 41M params) trained with this code. 84% top-1 @ 320x320. Trained at 256x256.
* Add EfficientNet-V2S model (unverified model definition) weights. 83.3 top-1 @ 288x288. Only trained single res 224. Working on progressive training.
* Add ByoaNet model definition (Bring-your-own-attention) w/ SelfAttention block and corresponding SA/SA-like modules and model defs
* Uses SiLU activation, approx 2x faster than `dm_nfnet_f0` and 50% faster than `nfnet_f0s` w/ 1/3 param count
* Integrate [Hugging Face model hub](https://huggingface.co/models) into timm create_model and default_cfg handling for pretrained weight and config sharing (more on this soon!)
* Merge HardCoRe NAS models contributed by https://github.com/yoniaflalo
* Merge PyTorch trained EfficientNet-EL and pruned ES/EL variants contributed by [DeGirum](https://github.com/DeGirum)
* Tested with PyTorch 1.8 release. Updated CI to use 1.8.
* Benchmarked several arch on RTX 3090, Titan RTX, and V100 across 1.7.1, 1.8, NGC 20.12, and 21.02. Some interesting performance variations to take note of https://gist.github.com/rwightman/bb59f9e245162cee0e38bd66bd8cd77f
* Add pretrained weights and model variants for NFNet-F* models from [DeepMind Haiku impl](https://github.com/deepmind/deepmind-research/tree/master/nfnets).
* Models are prefixed with `dm_`. They require SAME padding conv, skipinit enabled, and activation gains applied in act fn.
* These models are big, expect to run out of GPU memory. With the GELU activiation + other options, they are roughly 1/2 the inference speed of my SiLU PyTorch optimized `s` variants.
* Original model results are based on pre-processing that is not the same as all other models so you'll see different results in the results csv (once updated).
* Matching the original pre-processing as closely as possible I get these results:
* Add Adaptive Gradient Clipping (AGC) as per https://arxiv.org/abs/2102.06171. Integrated w/ PyTorch gradient clipping via mode arg that defaults to prev 'norm' mode. For backward arg compat, clip-grad arg must be specified to enable when using train.py.
* AGC performance is definitely sensitive to the clipping factor. More experimentation needed to determine good values for smaller batch sizes and optimizers besides those in paper. So far I've found .001-.005 is necessary for stable RMSProp training w/ NFNet/NF-ResNet.
* More model archs, incl a flexible ByobNet backbone ('Bring-your-own-blocks')
* GPU-Efficient-Networks (https://github.com/idstcv/GPU-Efficient-Networks), impl in `byobnet.py`
* RepVGG (https://github.com/DingXiaoH/RepVGG), impl in `byobnet.py`
* classic VGG (from torchvision, impl in `vgg.py`)
* Refinements to normalizer layer arg handling and normalizer+act layer handling in some models
* Default AMP mode changed to native PyTorch AMP instead of APEX. Issues not being fixed with APEX. Native works with `--channels-last` and `--torchscript` model training, APEX does not.
* Fix a few bugs introduced since last pypi release
* Remove separate tiered (`t`) vs tiered_narrow (`tn`) ResNet model defs, all `tn` changed to `t` and `t` models removed (`seresnext26t_32x4d` only model w/ weights that was removed).
* Support model default_cfgs with separate train vs test resolution `test_input_size` and remove extra `_320` suffix ResNet model defs that were just for test.
Py**T**orch **Im**age **M**odels (`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](https://rwightman.github.io/pytorch-image-models/training_hparam_examples) to get you started.
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:
* doing a forward pass on just the features - `forward_features` (see [documentation](https://rwightman.github.io/pytorch-image-models/feature_extraction/))
* All models support multi-scale feature map extraction (feature pyramids) via create_model (see [documentation](https://rwightman.github.io/pytorch-image-models/feature_extraction/))
*`out_indices` creation arg specifies which feature maps to return, these indices are 0 based and generally correspond to the `C(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.
* High performance [reference training, validation, and inference scripts](https://rwightman.github.io/pytorch-image-models/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)
* 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)
* Space-to-Depth by [mrT23](https://github.com/mrT23/TResNet/blob/master/src/models/tresnet/layers/space_to_depth.py) (https://arxiv.org/abs/1801.04590) -- original paper?
Model validation results can be found in the [documentation](https://rwightman.github.io/pytorch-image-models/results/) and in the [results tables](results/README.md)
My current [documentation](https://rwightman.github.io/pytorch-image-models/) for `timm` covers the basics.
[timmdocs](https://fastai.github.io/timmdocs/) is quickly becoming a much more comprehensive set of documentation for `timm`. A big thanks to [Aman Arora](https://github.com/amaarora) for his efforts creating timmdocs.
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](https://rwightman.github.io/pytorch-image-models/scripts/) for some basics and [training hparams](https://rwightman.github.io/pytorch-image-models/training_hparam_examples) 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.
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.
