❗Updates after Oct 10, 2022 are available in version >= 0.9❗
* Many changes since the last 0.6.x stable releases. They were previewed in 0.8.x dev releases but not everyone transitioned.
*`timm.models.layers` moved to `timm.layers`:
*`from timm.models.layers import name` will still work via deprecation mapping (but please transition to `timm.layers`).
*`import timm.models.layers.module` or `from timm.models.layers.module import name` needs to be changed now.
* Builder, helper, non-model modules in `timm.models` have a `_` prefix added, ie `timm.models.helpers` -> `timm.models._helpers`, there are temporary deprecation mapping files but those will be removed.
* All models now support `architecture.pretrained_tag` naming (ex `resnet50.rsb_a1`).
* The pretrained_tag is the specific weight variant (different head) for the architecture.
* In adding pretrained tags, many model names that existed to differentiate were renamed to use the tag (ex: `vit_base_patch16_224_in21k` -> `vit_base_patch16_224.augreg_in21k`). There are deprecation mappings for these.
* A number of models had their checkpoints remaped to match architecture changes needed to better support `features_only=True`, there are `checkpoint_filter_fn` methods in any model module that was remapped. These can be passed to `timm.models.load_checkpoint(..., filter_fn=timm.models.swin_transformer_v2.checkpoint_filter_fn)` to remap your existing checkpoint.
* The Hugging Face Hub (https://huggingface.co/timm) is now the primary source for `timm` weights. Model cards include link to papers, original source, license.
* Previous 0.6.x can be cloned from [0.6.x](https://github.com/rwightman/pytorch-image-models/tree/0.6.x) branch or installed via pip with version.
* Prototype of `set_input_size()` added to vit and swin v1/v2 models to allow changing image size, patch size, window size after model creation.
* Improved support in swin for different size handling, in addition to `set_input_size`, `always_partition` and `strict_img_size` args have been added to `__init__` to allow more flexible input size constraints
* Fix out of order indices info for intermediate 'Getter' feature wrapper, check out or range indices for same.
* Add several `tiny`<.5MparammodelsfortestingthatareactuallytrainedonImageNet-1k
* AttentionExtract helper added to extract attention maps from `timm` models. See example in https://github.com/huggingface/pytorch-image-models/discussions/1232#discussioncomment-9320949
*`forward_intermediates()` API refined and added to more models including some ConvNets that have other extraction methods.
* 1017 of 1047 model architectures support `features_only=True` feature extraction. Remaining 34 architectures can be supported but based on priority requests.
* Remove torch.jit.script annotated functions including old JIT activations. Conflict with dynamo and dynamo does a much better job when used.
* Prepping for a long overdue 1.0 release, things have been stable for a while now.
* Significant feature that's been missing for a while, `features_only=True` support for ViT models with flat hidden states or non-std module layouts (so far covering `'vit_*', 'twins_*', 'deit*', 'beit*', 'mvitv2*', 'eva*', 'samvit_*', 'flexivit*'`)
* Above feature support achieved through a new `forward_intermediates()` API that can be used with a feature wrapping module or direclty.
* Add `--bce-sum` (sum over class dim) and `--bce-pos-weight` (positive weighting) args for training as they're common BCE loss tweaks I was often hard coding
* Added significant flexibility for Hugging Face Hub based timm models via `model_args` config entry. `model_args` will be passed as kwargs through to models on creation.
* See example at https://huggingface.co/gaunernst/vit_base_patch16_1024_128.audiomae_as2m_ft_as20k/blob/main/config.json
* Add dynamic img size support to models in `vision_transformer.py`, `vision_transformer_hybrid.py`, `deit.py`, and `eva.py` w/o breaking backward compat.
* Add `dynamic_img_size=True` to args at model creation time to allow changing the grid size (interpolate abs and/or ROPE pos embed each forward pass).
* Add `dynamic_img_pad=True` to allow image sizes that aren't divisible by patch size (pad bottom right to patch size each forward pass).
* Enabling either dynamic mode will break FX tracing unless PatchEmbed module added as leaf.
* Existing method of resizing position embedding by passing different `img_size` (interpolate pretrained embed weights once) on creation still works.
* Existing method of changing `patch_size` (resize pretrained patch_embed weights once) on creation still works.
* Add `--reparam` arg to `benchmark.py`, `onnx_export.py`, and `validate.py` to trigger layer reparameterization / fusion for models with any one of `reparameterize()`, `switch_to_deploy()` or `fuse()`
* Including FastViT, MobileOne, RepGhostNet, EfficientViT (MSRA), RepViT, RepVGG, and LeViT
* Added timm trained `seresnextaa201d_32x8d.sw_in12k_ft_in1k_384` weights (and `.sw_in12k` pretrain) with 87.3% top-1 on ImageNet-1k, best ImageNet ResNet family model I'm aware of.
* RepViT model and weights (https://arxiv.org/abs/2307.09283) added by [wangao](https://github.com/jameslahm)
* I-JEPA ViT feature weights (no classifier) added by [SeeFun](https://github.com/seefun)
* SAM-ViT (segment anything) feature weights (no classifier) added by [SeeFun](https://github.com/seefun)
* Add support for alternative feat extraction methods and -ve indices to EfficientNet
* Experimental `get_intermediate_layers` function on vit/deit models for grabbing hidden states (inspired by DINO impl). This is WIP and may change significantly... feedback welcome.
* Model creation throws error if `pretrained=True` and no weights exist (instead of continuing with random initialization)
* Fix regression with inception / nasnet TF sourced weights with 1001 classes in original classifiers
* 97% of `timm` models uploaded to HF Hub and almost all updated to support multi-weight pretrained configs
* Minor cleanup and refactoring of another batch of models as multi-weight added. More fused_attn (F.sdpa) and features_only support, and torchscript fixes.
* Add ONNX export script, validate script, helpers that I've had kicking around for along time. Tweak 'same' padding for better export w/ recent ONNX + pytorch.
* Refactor dropout args for vit and vit-like models, separate drop_rate into `drop_rate` (classifier dropout), `proj_drop_rate` (block mlp / out projections), `pos_drop_rate` (position embedding drop), `attn_drop_rate` (attention dropout). Also add patch dropout (FLIP) to vit and eva models.
* fused F.scaled_dot_product_attention support to more vit models, add env var (TIMM_FUSED_ATTN) to control, and config interface to enable/disable
* Add EVA-CLIP backbones w/ image tower weights, all the way up to 4B param 'enormous' model, and 336x336 OpenAI ViT mode that was missed.
* Add EVA-02 MIM pretrained and fine-tuned weights, push to HF hub and update model cards for all EVA models. First model over 90% top-1 (99% top-5)! Check out the original code & weights at https://github.com/baaivision/EVA for more details on their work blending MIM, CLIP w/ many model, dataset, and train recipe tweaks.
* More weights pushed to HF hub along with multi-weight support, including: `regnet.py`, `rexnet.py`, `byobnet.py`, `resnetv2.py`, `swin_transformer.py`, `swin_transformer_v2.py`, `swin_transformer_v2_cr.py`
* Swin Transformer models support feature extraction (NCHW feat maps for `swinv2_cr_*`, and NHWC for all others) and spatial embedding outputs.
* FocalNet (from https://github.com/microsoft/FocalNet) models and weights added with significant refactoring, feature extraction, no fixed resolution / sizing constraint
* RegNet weights increased with HF hub push, SWAG, SEER, and torchvision v2 weights. SEER is pretty poor wrt to performance for model size, but possibly useful.
* More ImageNet-12k pretrained and 1k fine-tuned `timm` weights:
* Add 320x320 `convnext_large_mlp.clip_laion2b_ft_320` and `convnext_lage_mlp.clip_laion2b_ft_soup_320` CLIP image tower weights for features & fine-tune
* 0.8.13dev0 pypi release for latest changes w/ move to huggingface org
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.
Included optimizers available via `create_optimizer` / `create_optimizer_v2` factory methods:
*`adabelief` an implementation of AdaBelief adapted from https://github.com/juntang-zhuang/Adabelief-Optimizer - https://arxiv.org/abs/2010.07468
*`adafactor` adapted from [FAIRSeq impl](https://github.com/pytorch/fairseq/blob/master/fairseq/optim/adafactor.py) - https://arxiv.org/abs/1804.04235
*`adahessian` by [David Samuel](https://github.com/davda54/ada-hessian) - https://arxiv.org/abs/2006.00719
*`adamp` and `sgdp` by [Naver ClovAI](https://github.com/clovaai) - https://arxiv.org/abs/2006.08217
*`adan` an implementation of Adan adapted from https://github.com/sail-sg/Adan - https://arxiv.org/abs/2208.06677
*`lamb` an implementation of Lamb and LambC (w/ trust-clipping) cleaned up and modified to support use with XLA - https://arxiv.org/abs/1904.00962
*`lars` an implementation of LARS and LARC (w/ trust-clipping) - https://arxiv.org/abs/1708.03888
*`lion` and implementation of Lion adapted from https://github.com/google/automl/tree/master/lion - https://arxiv.org/abs/2302.06675
*`lookahead` adapted from impl by [Liam](https://github.com/alphadl/lookahead.pytorch) - https://arxiv.org/abs/1907.08610
*`madgrad` - and implementation of MADGRAD adapted from https://github.com/facebookresearch/madgrad - https://arxiv.org/abs/2101.11075
*`nadam` an implementation of Adam w/ Nesterov momentum
*`nadamw` an impementation of AdamW (Adam w/ decoupled weight-decay) w/ Nesterov momentum. A simplified impl based on https://github.com/mlcommons/algorithmic-efficiency
*`novograd` by [Masashi Kimura](https://github.com/convergence-lab/novograd) - https://arxiv.org/abs/1905.11286
*`radam` by [Liyuan Liu](https://github.com/LiyuanLucasLiu/RAdam) - https://arxiv.org/abs/1908.03265
*`rmsprop_tf` adapted from PyTorch RMSProp by myself. Reproduces much improved Tensorflow RMSProp behaviour
*`sgdw` and implementation of SGD w/ decoupled weight-decay
*`fused<name>` optimizers by name with [NVIDIA Apex](https://github.com/NVIDIA/apex/tree/master/apex/optimizers) installed
*`bits<name>` optimizers by name with [BitsAndBytes](https://github.com/TimDettmers/bitsandbytes) installed
### Augmentations
* Random Erasing from [Zhun Zhong](https://github.com/zhunzhong07/Random-Erasing/blob/master/transforms.py) - 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, JSD w/ clean + augmented mixing support works with AutoAugment and RandAugment as well - https://arxiv.org/abs/1912.02781
* SplitBachNorm - allows splitting batch norm layers between clean and augmented (auxiliary batch norm) data
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 support multi-scale feature map extraction (feature pyramids) via create_model (see [documentation](https://huggingface.co/docs/timm/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://huggingface.co/docs/timm/training_script) 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)
* 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?
[Getting Started with PyTorch Image Models (timm): A Practitioner’s Guide](https://towardsdatascience.com/getting-started-with-pytorch-image-models-timm-a-practitioners-guide-4e77b4bf9055) by [Chris Hughes](https://github.com/Chris-hughes10) is an extensive blog post covering many aspects of `timm` in detail.
[timmdocs](http://timm.fast.ai/) is an alternate 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://huggingface.co/docs/timm/training_script).
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.
