|
|
|
@ -1,130 +1,130 @@
|
|
|
|
# Recent Changes
|
|
|
|
# Recent Changes
|
|
|
|
|
|
|
|
|
|
|
|
### June 8, 2021
|
|
|
|
|
|
|
|
* Add first ResMLP weights, trained in PyTorch XLA on TPU-VM w/ my XLA branch. 24 block variant, 79.2 top-1.
|
|
|
|
|
|
|
|
* Add ResNet51-Q model w/ pretrained weights at 82.36 top-1.
|
|
|
|
|
|
|
|
* NFNet inspired block layout with quad layer stem and no maxpool
|
|
|
|
|
|
|
|
* Same param count (35.7M) and throughput as ResNetRS-50 but +1.5 top-1 @ 224x224 and +2.5 top-1 at 288x288
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
### May 25, 2021
|
|
|
|
### March 23, 2022
|
|
|
|
* Add LeViT, Visformer, Convit (PR by Aman Arora), Twins (PR by paper authors) transformer models
|
|
|
|
* Add `ParallelBlock` and `LayerScale` option to base vit models to support model configs in [Three things everyone should know about ViT](https://arxiv.org/abs/2203.09795)
|
|
|
|
* Cleanup input_size/img_size override handling and testing for all vision transformer models
|
|
|
|
* `convnext_tiny_hnf` (head norm first) weights trained with (close to) A2 recipe, 82.2% top-1, could do better with more epochs.
|
|
|
|
* Add `efficientnetv2_rw_m` model and weights (started training before official code). 84.8 top-1, 53M params.
|
|
|
|
|
|
|
|
|
|
|
|
### March 21, 2022
|
|
|
|
### May 14, 2021
|
|
|
|
* Merge `norm_norm_norm`. **IMPORTANT** this update for a coming 0.6.x release will likely de-stabilize the master branch for a while. Branch [`0.5.x`](https://github.com/rwightman/pytorch-image-models/tree/0.5.x) or a previous 0.5.x release can be used if stability is required.
|
|
|
|
* Add EfficientNet-V2 official model defs w/ ported weights from official [Tensorflow/Keras](https://github.com/google/automl/tree/master/efficientnetv2) impl.
|
|
|
|
* Significant weights update (all TPU trained) as described in this [release](https://github.com/rwightman/pytorch-image-models/releases/tag/v0.1-tpu-weights)
|
|
|
|
* 1k trained variants: `tf_efficientnetv2_s/m/l`
|
|
|
|
* `regnety_040` - 82.3 @ 224, 82.96 @ 288
|
|
|
|
* 21k trained variants: `tf_efficientnetv2_s/m/l_in21k`
|
|
|
|
* `regnety_064` - 83.0 @ 224, 83.65 @ 288
|
|
|
|
* 21k pretrained -> 1k fine-tuned: `tf_efficientnetv2_s/m/l_in21ft1k`
|
|
|
|
* `regnety_080` - 83.17 @ 224, 83.86 @ 288
|
|
|
|
* v2 models w/ v1 scaling: `tf_efficientnetv2_b0` through `b3`
|
|
|
|
* `regnetv_040` - 82.44 @ 224, 83.18 @ 288 (timm pre-act)
|
|
|
|
* Rename my prev V2 guess `efficientnet_v2s` -> `efficientnetv2_rw_s`
|
|
|
|
* `regnetv_064` - 83.1 @ 224, 83.71 @ 288 (timm pre-act)
|
|
|
|
* Some blank `efficientnetv2_*` models in-place for future native PyTorch training
|
|
|
|
* `regnetz_040` - 83.67 @ 256, 84.25 @ 320
|
|
|
|
|
|
|
|
* `regnetz_040h` - 83.77 @ 256, 84.5 @ 320 (w/ extra fc in head)
|
|
|
|
### May 5, 2021
|
|
|
|
* `resnetv2_50d_gn` - 80.8 @ 224, 81.96 @ 288 (pre-act GroupNorm)
|
|
|
|
* Add MLP-Mixer models and port pretrained weights from [Google JAX impl](https://github.com/google-research/vision_transformer/tree/linen)
|
|
|
|
* `resnetv2_50d_evos` 80.77 @ 224, 82.04 @ 288 (pre-act EvoNormS)
|
|
|
|
* Add CaiT models and pretrained weights from [FB](https://github.com/facebookresearch/deit)
|
|
|
|
* `regnetz_c16_evos` - 81.9 @ 256, 82.64 @ 320 (EvoNormS)
|
|
|
|
* Add ResNet-RS models and weights from [TF](https://github.com/tensorflow/tpu/tree/master/models/official/resnet/resnet_rs). Thanks [Aman Arora](https://github.com/amaarora)
|
|
|
|
* `regnetz_d8_evos` - 83.42 @ 256, 84.04 @ 320 (EvoNormS)
|
|
|
|
* Add CoaT models and weights. Thanks [Mohammed Rizin](https://github.com/morizin)
|
|
|
|
* `xception41p` - 82 @ 299 (timm pre-act)
|
|
|
|
* Add new ImageNet-21k weights & finetuned weights for TResNet, MobileNet-V3, ViT models. Thanks [mrT](https://github.com/mrT23)
|
|
|
|
* `xception65` - 83.17 @ 299
|
|
|
|
* Add GhostNet models and weights. Thanks [Kai Han](https://github.com/iamhankai)
|
|
|
|
* `xception65p` - 83.14 @ 299 (timm pre-act)
|
|
|
|
* Update ByoaNet attention modles
|
|
|
|
* `resnext101_64x4d` - 82.46 @ 224, 83.16 @ 288
|
|
|
|
* Improve SA module inits
|
|
|
|
* `seresnext101_32x8d` - 83.57 @ 224, 84.270 @ 288
|
|
|
|
* Hack together experimental stand-alone Swin based attn module and `swinnet`
|
|
|
|
* `resnetrs200` - 83.85 @ 256, 84.44 @ 320
|
|
|
|
* Consistent '26t' model defs for experiments.
|
|
|
|
* 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)
|
|
|
|
* Add improved Efficientnet-V2S (prelim model def) weights. 83.8 top-1.
|
|
|
|
* SwinTransformer-V2 implementation added. Submitted by [Christoph Reich](https://github.com/ChristophReich1996). Training experiments and model changes by myself are ongoing so expect compat breaks.
|
|
|
|
* WandB logging support
|
|
|
|
* 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
|
|
|
|
### April 13, 2021
|
|
|
|
* PoolFormer models w/ weights adapted from https://github.com/sail-sg/poolformer
|
|
|
|
* Add Swin Transformer models and weights from https://github.com/microsoft/Swin-Transformer
|
|
|
|
* 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
|
|
|
|
### April 12, 2021
|
|
|
|
* Enhance support for alternate norm + act ('NormAct') layers added to a number of models, esp EfficientNet/MobileNetV3, RegNet, and aligned Xception
|
|
|
|
* Add ECA-NFNet-L1 (slimmed down F1 w/ SiLU, 41M params) trained with this code. 84% top-1 @ 320x320. Trained at 256x256.
|
|
|
|
* Grouped conv support added to EfficientNet family
|
|
|
|
* Add EfficientNet-V2S model (unverified model definition) weights. 83.3 top-1 @ 288x288. Only trained single res 224. Working on progressive training.
|
|
|
|
* 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
|
|
|
|
* Add ByoaNet model definition (Bring-your-own-attention) w/ SelfAttention block and corresponding SA/SA-like modules and model defs
|
|
|
|
* Gradient checkpointing support added to many models
|
|
|
|
* Lambda Networks - https://arxiv.org/abs/2102.08602
|
|
|
|
* `forward_head(x, pre_logits=False)` fn added to all models to allow separate calls of `forward_features` + `forward_head`
|
|
|
|
* Bottleneck Transformers - https://arxiv.org/abs/2101.11605
|
|
|
|
* 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 in `forward_head`
|
|
|
|
* Halo Nets - https://arxiv.org/abs/2103.12731
|
|
|
|
|
|
|
|
* Adabelief optimizer contributed by Juntang Zhuang
|
|
|
|
### Feb 2, 2022
|
|
|
|
|
|
|
|
* [Chris Hughes](https://github.com/Chris-hughes10) 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](https://towardsdatascience.com/getting-started-with-pytorch-image-models-timm-a-practitioners-guide-4e77b4bf9055)
|
|
|
|
### April 1, 2021
|
|
|
|
* I'm currently prepping to merge the `norm_norm_norm` branch back to master (ver 0.6.x) in next week or so.
|
|
|
|
* Add snazzy `benchmark.py` script for bulk `timm` model benchmarking of train and/or inference
|
|
|
|
* 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!
|
|
|
|
* Add Pooling-based Vision Transformer (PiT) models (from https://github.com/naver-ai/pit)
|
|
|
|
* `0.5.x` releases and a `0.5.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.
|
|
|
|
* Merged distilled variant into main for torchscript compatibility
|
|
|
|
|
|
|
|
* Some `timm` cleanup/style tweaks and weights have hub download support
|
|
|
|
### Jan 14, 2022
|
|
|
|
* Cleanup Vision Transformer (ViT) models
|
|
|
|
* 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....
|
|
|
|
* Merge distilled (DeiT) model into main so that torchscript can work
|
|
|
|
* Add ConvNeXT models /w weights from official impl (https://github.com/facebookresearch/ConvNeXt), a few perf tweaks, compatible with timm features
|
|
|
|
* Support updated weight init (defaults to old still) that closer matches original JAX impl (possibly better training from scratch)
|
|
|
|
* Tried training a few small (~1.8-3M param) / mobile optimized models, a few are good so far, more on the way...
|
|
|
|
* Separate hybrid model defs into different file and add several new model defs to fiddle with, support patch_size != 1 for hybrids
|
|
|
|
* `mnasnet_small` - 65.6 top-1
|
|
|
|
* Fix fine-tuning num_class changes (PiT and ViT) and pos_embed resizing (Vit) with distilled variants
|
|
|
|
* `mobilenetv2_050` - 65.9
|
|
|
|
* nn.Sequential for block stack (does not break downstream compat)
|
|
|
|
* `lcnet_100/075/050` - 72.1 / 68.8 / 63.1
|
|
|
|
* TnT (Transformer-in-Transformer) models contributed by author (from https://gitee.com/mindspore/mindspore/tree/master/model_zoo/research/cv/TNT)
|
|
|
|
* `semnasnet_075` - 73
|
|
|
|
* Add RegNetY-160 weights from DeiT teacher model
|
|
|
|
* `fbnetv3_b/d/g` - 79.1 / 79.7 / 82.0
|
|
|
|
* Add new NFNet-L0 w/ SE attn (rename `nfnet_l0b`->`nfnet_l0`) weights 82.75 top-1 @ 288x288
|
|
|
|
* TinyNet models added by [rsomani95](https://github.com/rsomani95)
|
|
|
|
* Some fixes/improvements for TFDS dataset wrapper
|
|
|
|
* LCNet added via MobileNetV3 architecture
|
|
|
|
|
|
|
|
|
|
|
|
### March 7, 2021
|
|
|
|
### Nov 22, 2021
|
|
|
|
* First 0.4.x PyPi release w/ NFNets (& related), ByoB (GPU-Efficient, RepVGG, etc).
|
|
|
|
* A number of updated weights anew new model defs
|
|
|
|
* Change feature extraction for pre-activation nets (NFNets, ResNetV2) to return features before activation.
|
|
|
|
* `eca_halonext26ts` - 79.5 @ 256
|
|
|
|
|
|
|
|
* `resnet50_gn` (new) - 80.1 @ 224, 81.3 @ 288
|
|
|
|
### Feb 18, 2021
|
|
|
|
* `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))
|
|
|
|
* Add pretrained weights and model variants for NFNet-F* models from [DeepMind Haiku impl](https://github.com/deepmind/deepmind-research/tree/master/nfnets).
|
|
|
|
* `resnext50_32x4d` - 81.1 @ 224, 82.0 @ 288
|
|
|
|
* Models are prefixed with `dm_`. They require SAME padding conv, skipinit enabled, and activation gains applied in act fn.
|
|
|
|
* `sebotnet33ts_256` (new) - 81.2 @ 224
|
|
|
|
* 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.
|
|
|
|
* `lamhalobotnet50ts_256` - 81.5 @ 256
|
|
|
|
* 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).
|
|
|
|
* `halonet50ts` - 81.7 @ 256
|
|
|
|
* Matching the original pre-processing as closely as possible I get these results:
|
|
|
|
* `halo2botnet50ts_256` - 82.0 @ 256
|
|
|
|
* `dm_nfnet_f6` - 86.352
|
|
|
|
* `resnet101` - 82.0 @ 224, 82.8 @ 288
|
|
|
|
* `dm_nfnet_f5` - 86.100
|
|
|
|
* `resnetv2_101` (new) - 82.1 @ 224, 83.0 @ 288
|
|
|
|
* `dm_nfnet_f4` - 85.834
|
|
|
|
* `resnet152` - 82.8 @ 224, 83.5 @ 288
|
|
|
|
* `dm_nfnet_f3` - 85.676
|
|
|
|
* `regnetz_d8` (new) - 83.5 @ 256, 84.0 @ 320
|
|
|
|
* `dm_nfnet_f2` - 85.178
|
|
|
|
* `regnetz_e8` (new) - 84.5 @ 256, 85.0 @ 320
|
|
|
|
* `dm_nfnet_f1` - 84.696
|
|
|
|
* `vit_base_patch8_224` (85.8 top-1) & `in21k` variant weights added thanks [Martins Bruveris](https://github.com/martinsbruveris)
|
|
|
|
* `dm_nfnet_f0` - 83.464
|
|
|
|
* Groundwork in for FX feature extraction thanks to [Alexander Soare](https://github.com/alexander-soare)
|
|
|
|
|
|
|
|
* models updated for tracing compatibility (almost full support with some distlled transformer exceptions)
|
|
|
|
### Feb 16, 2021
|
|
|
|
|
|
|
|
* 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.
|
|
|
|
### Oct 19, 2021
|
|
|
|
* AGC w/ default clipping factor `--clip-grad .01 --clip-mode agc`
|
|
|
|
* 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)
|
|
|
|
* PyTorch global norm of 1.0 (old behaviour, always norm), `--clip-grad 1.0`
|
|
|
|
* BCE loss and Repeated Augmentation support for RSB paper
|
|
|
|
* PyTorch value clipping of 10, `--clip-grad 10. --clip-mode value`
|
|
|
|
* 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)
|
|
|
|
* 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.
|
|
|
|
* Working implementations of the following 2D self-attention modules (likely to be differences from paper or eventual official impl):
|
|
|
|
|
|
|
|
* Halo (https://arxiv.org/abs/2103.12731)
|
|
|
|
### Feb 12, 2021
|
|
|
|
* Bottleneck Transformer (https://arxiv.org/abs/2101.11605)
|
|
|
|
* Update Normalization-Free nets to include new NFNet-F (https://arxiv.org/abs/2102.06171) model defs
|
|
|
|
* LambdaNetworks (https://arxiv.org/abs/2102.08602)
|
|
|
|
|
|
|
|
* 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)
|
|
|
|
### Feb 10, 2021
|
|
|
|
* ConvMixer (https://openreview.net/forum?id=TVHS5Y4dNvM), CrossVit (https://arxiv.org/abs/2103.14899), and BeiT (https://arxiv.org/abs/2106.08254) architectures + weights added
|
|
|
|
* More model archs, incl a flexible ByobNet backbone ('Bring-your-own-blocks')
|
|
|
|
* freeze/unfreeze helpers by [Alexander Soare](https://github.com/alexander-soare)
|
|
|
|
* GPU-Efficient-Networks (https://github.com/idstcv/GPU-Efficient-Networks), impl in `byobnet.py`
|
|
|
|
|
|
|
|
* RepVGG (https://github.com/DingXiaoH/RepVGG), impl in `byobnet.py`
|
|
|
|
### Aug 18, 2021
|
|
|
|
* classic VGG (from torchvision, impl in `vgg`)
|
|
|
|
* Optimizer bonanza!
|
|
|
|
* Refinements to normalizer layer arg handling and normalizer+act layer handling in some models
|
|
|
|
* 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))
|
|
|
|
* 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.
|
|
|
|
* Add MADGRAD from FB research w/ a few tweaks (decoupled decay option, step handling that works with PyTorch XLA)
|
|
|
|
* Fix a few bugs introduced since last pypi release
|
|
|
|
* 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).
|
|
|
|
### Feb 8, 2021
|
|
|
|
* 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.
|
|
|
|
* Add several ResNet weights with ECA attention. 26t & 50t trained @ 256, test @ 320. 269d train @ 256, fine-tune @320, test @ 352.
|
|
|
|
* 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.
|
|
|
|
* `ecaresnet26t` - 79.88 top-1 @ 320x320, 79.08 @ 256x256
|
|
|
|
|
|
|
|
* `ecaresnet50t` - 82.35 top-1 @ 320x320, 81.52 @ 256x256
|
|
|
|
### July 12, 2021
|
|
|
|
* `ecaresnet269d` - 84.93 top-1 @ 352x352, 84.87 @ 320x320
|
|
|
|
* Add XCiT models from [official facebook impl](https://github.com/facebookresearch/xcit). Contributed by [Alexander Soare](https://github.com/alexander-soare)
|
|
|
|
* 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.
|
|
|
|
### July 5-9, 2021
|
|
|
|
|
|
|
|
* 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)
|
|
|
|
### Jan 30, 2021
|
|
|
|
* top-1 82.34 @ 288x288 and 82.54 @ 320x320
|
|
|
|
* Add initial "Normalization Free" NF-RegNet-B* and NF-ResNet model definitions based on [paper](https://arxiv.org/abs/2101.08692)
|
|
|
|
* 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.
|
|
|
|
|
|
|
|
* Add 'Aggregating Nested Transformer' (NesT) w/ weights converted from official [Flax impl](https://github.com/google-research/nested-transformer). Contributed by [Alexander Soare](https://github.com/alexander-soare).
|
|
|
|
### Jan 25, 2021
|
|
|
|
* `jx_nest_base` - 83.534, `jx_nest_small` - 83.120, `jx_nest_tiny` - 81.426
|
|
|
|
* Add ResNetV2 Big Transfer (BiT) models w/ ImageNet-1k and 21k weights from https://github.com/google-research/big_transfer
|
|
|
|
|
|
|
|
* Add official R50+ViT-B/16 hybrid models + weights from https://github.com/google-research/vision_transformer
|
|
|
|
### June 23, 2021
|
|
|
|
* ImageNet-21k ViT weights are added w/ model defs and representation layer (pre logits) support
|
|
|
|
* 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)
|
|
|
|
* NOTE: ImageNet-21k classifier heads were zero'd in original weights, they are only useful for transfer learning
|
|
|
|
|
|
|
|
* Add model defs and weights for DeiT Vision Transformer models from https://github.com/facebookresearch/deit
|
|
|
|
### June 20, 2021
|
|
|
|
* Refactor dataset classes into ImageDataset/IterableImageDataset + dataset specific parser classes
|
|
|
|
* Release Vision Transformer 'AugReg' weights from [How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers](https://arxiv.org/abs/2106.10270)
|
|
|
|
* Add Tensorflow-Datasets (TFDS) wrapper to allow use of TFDS image classification sets with train script
|
|
|
|
* .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)
|
|
|
|
* Ex: `train.py /data/tfds --dataset tfds/oxford_iiit_pet --val-split test --model resnet50 -b 256 --amp --num-classes 37 --opt adamw --lr 3e-4 --weight-decay .001 --pretrained -j 2`
|
|
|
|
* 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
|
|
|
|
* Add improved .tar dataset parser that reads images from .tar, folder of .tar files, or .tar within .tar
|
|
|
|
* Replaced all default weights w/ best AugReg variant (if possible). All AugReg 21k classifiers work.
|
|
|
|
* Run validation on full ImageNet-21k directly from tar w/ BiT model: `validate.py /data/fall11_whole.tar --model resnetv2_50x1_bitm_in21k --amp`
|
|
|
|
* Highlights: `vit_large_patch16_384` (87.1 top-1), `vit_large_r50_s32_384` (86.2 top-1), `vit_base_patch16_384` (86.0 top-1)
|
|
|
|
* Models in this update should be stable w/ possible exception of ViT/BiT, possibility of some regressions with train/val scripts and dataset handling
|
|
|
|
* `vit_deit_*` renamed to just `deit_*`
|
|
|
|
|
|
|
|
* Remove my old small model, replace with DeiT compatible small w/ AugReg weights
|
|
|
|
### Jan 3, 2021
|
|
|
|
* Add 1st training of my `gmixer_24_224` MLP /w GLU, 78.1 top-1 w/ 25M params.
|
|
|
|
* Add SE-ResNet-152D weights
|
|
|
|
* Add weights from official ResMLP release (https://github.com/facebookresearch/deit)
|
|
|
|
* 256x256 val, 0.94 crop top-1 - 83.75
|
|
|
|
* Add `eca_nfnet_l2` weights from my 'lightweight' series. 84.7 top-1 at 384x384.
|
|
|
|
* 320x320 val, 1.0 crop - 84.36
|
|
|
|
* 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)
|
|
|
|
* Update results files
|
|
|
|
* NFNets and ResNetV2-BiT models work w/ Pytorch XLA now
|
|
|
|
|
|
|
|
* weight standardization uses F.batch_norm instead of std_mean (std_mean wasn't lowered)
|
|
|
|
|
|
|
|
* eps values adjusted, will be slight differences but should be quite close
|
|
|
|
|
|
|
|
* Improve test coverage and classifier interface of non-conv (vision transformer and mlp) models
|
|
|
|
|
|
|
|
* Cleanup a few classifier / flatten details for models w/ conv classifiers or early global pool
|
|
|
|
|
|
|
|
* Please report any regressions, this PR touched quite a few models.
|
|
|
|
|
Ross Wightman
3 years ago