v0.6.7 Release

Minor bug fixes and a few more weights since 0.6.5

• A few more weights & model defs added:
  • darknetaa53 - 79.8 @ 256, 80.5 @ 288
  • convnext_nano - 80.8 @ 224, 81.5 @ 288
  • cs3sedarknet_l - 81.2 @ 256, 81.8 @ 288
  • cs3darknet_x - 81.8 @ 256, 82.2 @ 288
  • cs3sedarknet_x - 82.2 @ 256, 82.7 @ 288
  • cs3edgenet_x - 82.2 @ 256, 82.7 @ 288
  • cs3se_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

v0.6.5 Release

First official release in a long while (since 0.5.4). All change log since 0.5.4 below,

July 8, 2022

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.
• My own model weight results (all ImageNet-1k training)
  • resnet10t - 66.5 @ 176, 68.3 @ 224
  • resnet14t - 71.3 @ 176, 72.3 @ 224
  • resnetaa50 - 80.6 @ 224 , 81.6 @ 288
  • darknet53 - 80.0 @ 256, 80.5 @ 288
  • cs3darknet_m - 77.0 @ 256, 77.6 @ 288
  • cs3darknet_focus_m - 76.7 @ 256, 77.3 @ 288
  • cs3darknet_l - 80.4 @ 256, 80.9 @ 288
  • cs3darknet_focus_l - 80.3 @ 256, 80.9 @ 288
  • vit_srelpos_small_patch16_224 - 81.1 @ 224, 82.1 @ 320
  • vit_srelpos_medium_patch16_224 - 82.3 @ 224, 83.1 @ 320
  • vit_relpos_small_patch16_cls_224 - 82.6 @ 224, 83.6 @ 320
  • edgnext_small_rw - 79.6 @ 224, 80.4 @ 320
• cs3, darknet, and vit_*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/parsers. See (#1274 (comment))
• Default ConvNeXt LayerNorm impl to use F.layer_norm(x.permute(0, 2, 3, 1), ...).permute(0, 3, 1, 2) via LayerNorm2d 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 in models/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) ...

May 13, 2022

• 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 pool
  • vit_relpos_medium_patch16_rpn_224 - 82.3 @ 224, 83.1 @ 320 -- rel pos + res-post-norm, no class token, avg pool
  • vit_relpos_medium_patch16_224 - 82.5 @ 224, 83.3 @ 320 -- rel pos, layer scale, no class token, avg pool
  • vit_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)

May 2, 2022

• 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 pool
  • vit_relpos_base_patch16_224 - 82.5 @ 224, 83.6 @ 320 -- rel pos, layer scale, no class token, avg pool
  • vit_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).

April 22, 2022

• 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 @ 288
  • seresnextaa101d_32x8d (anti-aliased w/ AvgPool2d) - 83.85 @ 224, 84.57 @ 288

March 23, 2022

• Add ParallelBlock and LayerScale 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.

March 21, 2022

• 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 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 @ 288
  • regnety_064 - 83.0 @ 224, 83.65 @ 288
  • regnety_080 - 83.17 @ 224, 83.86 @ 288
  • regnetv_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 @ 320
  • regnetz_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 @ 299
  • xception65p - 83.14 @ 299 (timm pre-act)
  • resnext101_64x4d - 82.46 @ 224, 83.16 @ 288
  • seresnext101_32x8d - 83.57 @ 224, 84.270 @ 288
  • resnetrs200 - 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 of forward_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 in forward_head

Feb 2, 2022

• 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 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.

Swin Transformer V2 (CR) weights and experiments

This release holds weights for timm's variant of Swin V2 (from @ChristophReich1996 impl, https://github.com/ChristophReich1996/Swin-Transformer-V2)

NOTE: ns variants of the models have extra norms on the main branch at the end of each stage, this seems to help training. The current small model is not using this, but currently training one. Will have a non-ns tiny soon as well as a comparsion. in21k and 1k base models are also in the works...

small checkpoints trained on TPU-VM instances via the TPU-Research Cloud (https://sites.research.google/trc/about/)

• swin_v2_tiny_ns_224 - 81.80 top-1
• swin_v2_small_224 - 83.13 top-1
• swin_v2_small_ns_224 - 83.5 top-1

TPU VM trained weight release w/ PyTorch XLA

A wide range of mid-large sized models trained in PyTorch XLA on TPU VM instances. Demonstrating viability of the TPU + PyTorch combo for excellent image model results. All models trained w/ the bits_and_tpu branch of this codebase.

A big thanks to the TPU Research Cloud (https://sites.research.google/trc/about/) for the compute used in these experiments.

This set includes several novel weights, including EvoNorm-S RegNetZ (C/D timm variants) and ResNet-V2 model experiments, as well as custom pre-activation model variants of RegNet-Y (called RegNet-V) and Xception (Xception-P) models.

Many if not all of the included RegNet weights surpass original paper results by a wide margin and remain above other known results (e.g. recent torchvision updates) in ImageNet-1k validation and especially OOD test set / robustness performance and scaling to higher resolutions.

RegNets

• regnety_040 - 82.3 @ 224, 82.96 @ 288
• regnety_064 - 83.0 @ 224, 83.65 @ 288
• regnety_080 - 83.17 @ 224, 83.86 @ 288
• regnetv_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 @ 320
• regnetz_040h - 83.77 @ 256, 84.5 @ 320 (w/ extra fc in head)

Alternative norm layers (no BN!)

• 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)

Xception redux

• xception41p - 82 @ 299 (timm pre-act)
• xception65 - 83.17 @ 299
• xception65p - 83.14 @ 299 (timm pre-act)

ResNets (w/ SE and/or NeXT)

• resnext101_64x4d - 82.46 @ 224, 83.16 @ 288
• seresnext101_32x8d - 83.57 @ 224, 84.27 @ 288
• seresnext101d_32x8d - 83.69 @ 224, 84.35 @ 288
• seresnextaa101d_32x8d - 83.85 @ 224, 84.57 @ 288
• resnetrs200 - 83.85 @ 256, 84.44 @ 320

Vision transformer experiments -- relpos, residual-post-norm, layer-scale, fc-norm, and GAP

• vit_relpos_base_patch32_plus_rpn_256 - 79.5 @ 256, 80.6 @ 320 -- rel pos + extended width + res-post-norm, no class token, avg pool
• vit_relpos_small_patch16_224 - 81.5 @ 224, 82.5 @ 320 -- rel pos, layer scale, no class token, avg pool
• vit_relpos_medium_patch16_rpn_224 - 82.3 @ 224, 83.1 @ 320 -- rel pos + res-post-norm, no class token, avg pool
• vit_base_patch16_rpn_224 - 82.3 @ 224 -- rel pos + res-post-norm, no class token, avg pool
• vit_relpos_medium_patch16_224 - 82.5 @ 224, 83.3 @ 320 -- rel pos, layer scale, no class token, avg pool
• vit_relpos_base_patch16_224 - 82.5 @ 224, 83.6 @ 320 -- rel pos, layer scale, no class token, avg pool
• vit_relpos_base_patch16_gapcls_224 - 82.8 @ 224, 83.9 @ 320 -- rel pos, layer scale, class token, avg pool (by mistake)

MobileViT weights

Pretrained weights for MobileViT and MobileViT-V2 adapted from Apple impl at https://github.com/apple/ml-cvnets

Checkpoints remapped to timm impl of the model with BGR corrected to RGB (for V1).

v0.5.4 - More weights, models. ResNet strikes back, self-attn - convnet hybrids, optimizers and more

Default conv_mlp to False across the board for ConvNeXt, causing issu…

v0.1-rsb-weights

Weights for ResNet Strikes Back

Paper: https://arxiv.org/abs/2110.00476

More details on weights and hparams to come...

v0.1-attn-weights

A collection of weights I've trained comparing various types of SE-like (SE, ECA, GC, etc), self-attention (bottleneck, halo, lambda) blocks, and related non-attn baselines.

ResNet-26-T series

• [2, 2, 2, 2] repeat Bottlneck block ResNet architecture
• ReLU activations
• 3 layer stem with 24, 32, 64 chs, max-pool
• avg pool in shortcut downsample
• self-attn blocks replace 3x3 in both blocks for last stage, and second block of penultimate stage

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
botnet26t_256	79.246	20.754	94.53	5.47	12.49	256	0.95	bicubic
halonet26t	79.13	20.87	94.314	5.686	12.48	256	0.95	bicubic
lambda_resnet26t	79.112	20.888	94.59	5.41	10.96	256	0.94	bicubic
lambda_resnet26rpt_256	78.964	21.036	94.428	5.572	10.99	256	0.94	bicubic
resnet26t	77.872	22.128	93.834	6.166	16.01	256	0.94	bicubic

Details:

• HaloNet - 8 pixel block size, 2 pixel halo (overlap), relative position embedding
• BotNet - relative position embedding
• Lambda-ResNet-26-T - 3d lambda conv, kernel = 9
• Lambda-ResNet-26-RPT - relative position embedding

Benchmark - RTX 3090 - AMP - NCHW - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnet26t	2967.55	86.252	256	256	857.62	297.984	256	256	16.01
botnet26t_256	2642.08	96.879	256	256	809.41	315.706	256	256	12.49
halonet26t	2601.91	98.375	256	256	783.92	325.976	256	256	12.48
lambda_resnet26t	2354.1	108.732	256	256	697.28	366.521	256	256	10.96
lambda_resnet26rpt_256	1847.34	138.563	256	256	644.84	197.892	128	256	10.99

Benchmark - RTX 3090 - AMP - NHWC - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnet26t	3691.94	69.327	256	256	1188.17	214.96	256	256	16.01
botnet26t_256	3291.63	77.76	256	256	1126.68	226.653	256	256	12.49
halonet26t	3230.5	79.232	256	256	1077.82	236.934	256	256	12.48
lambda_resnet26rpt_256	2324.15	110.133	256	256	864.42	147.485	128	256	10.99
lambda_resnet26t	Not Supported

ResNeXT-26-T series

• [2, 2, 2, 2] repeat Bottlneck block ResNeXt architectures
• SiLU activations
• grouped 3x3 convolutions in bottleneck, 32 channels per group
• 3 layer stem with 24, 32, 64 chs, max-pool
• avg pool in shortcut downsample
• channel attn (active in non self-attn blocks) between 3x3 and last 1x1 conv
• when active, self-attn blocks replace 3x3 conv in both blocks for last stage, and second block of penultimate stage

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
eca_halonext26ts	79.484	20.516	94.600	5.400	10.76	256	0.94	bicubic
eca_botnext26ts_256	79.270	20.730	94.594	5.406	10.59	256	0.95	bicubic
bat_resnext26ts	78.268	21.732	94.1	5.9	10.73	256	0.9	bicubic
seresnext26ts	77.852	22.148	93.784	6.216	10.39	256	0.9	bicubic
gcresnext26ts	77.804	22.196	93.824	6.176	10.48	256	0.9	bicubic
eca_resnext26ts	77.446	22.554	93.57	6.43	10.3	256	0.9	bicubic
resnext26ts	76.764	23.236	93.136	6.864	10.3	256	0.9	bicubic

Benchmark - RTX 3090 - AMP - NCHW - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnext26ts	3006.57	85.134	256	256	864.4	295.646	256	256	10.3
seresnext26ts	2931.27	87.321	256	256	836.92	305.193	256	256	10.39
eca_resnext26ts	2925.47	87.495	256	256	837.78	305.003	256	256	10.3
gcresnext26ts	2870.01	89.186	256	256	818.35	311.97	256	256	10.48
eca_botnext26ts_256	2652.03	96.513	256	256	790.43	323.257	256	256	10.59
eca_halonext26ts	2593.03	98.705	256	256	766.07	333.541	256	256	10.76
bat_resnext26ts	2469.78	103.64	256	256	697.21	365.964	256	256	10.73

Benchmark - RTX 3090 - AMP - NHWC - NGC 21.09

NOTE: there are performance issues with certain grouped conv configs with channels last layout, backwards pass in particular is really slow. Also causing issues for RegNet and NFNet networks.

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnext26ts	3952.37	64.755	256	256	608.67	420.049	256	256	10.3
eca_resnext26ts	3815.77	67.074	256	256	594.35	430.146	256	256	10.3
seresnext26ts	3802.75	67.304	256	256	592.82	431.14	256	256	10.39
gcresnext26ts	3626.97	70.57	256	256	581.83	439.119	256	256	10.48
eca_botnext26ts_256	3515.84	72.8	256	256	611.71	417.862	256	256	10.59
eca_halonext26ts	3410.12	75.057	256	256	597.52	427.789	256	256	10.76
bat_resnext26ts	3053.83	83.811	256	256	533.23	478.839	256	256	10.73

ResNet-33-T series.

• [2, 3, 3, 2] repeat Bottlneck block ResNet architecture
• SiLU activations
• 3 layer stem with 24, 32, 64 chs, no max-pool, 1st and 3rd conv stride 2
• avg pool in shortcut downsample
• channel attn (active in non self-attn blocks) between 3x3 and last 1x1 conv
• when active, self-attn blocks replace 3x3 conv last block of stage 2 and 3, and both blocks of final stage
• FC 1x1 conv between last block and classifier

The 33-layer models have an extra 1x1 FC layer between last conv block and classifier. There is both a non-attenion 33 layer baseline and a 32 layer without the extra FC.

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
sehalonet33ts	80.986	19.014	95.272	4.728	13.69	256	0.94	bicubic
seresnet33ts	80.388	19.612	95.108	4.892	19.78	256	0.94	bicubic
eca_resnet33ts	80.132	19.868	95.054	4.946	19.68	256	0.94	bicubic
gcresnet33ts	79.99	20.01	94.988	5.012	19.88	256	0.94	bicubic
resnet33ts	79.352	20.648	94.596	5.404	19.68	256	0.94	bicubic
resnet32ts	79.028...

v0.4.12. Vision Transformer AugReg support and more

• Vision Transformer AugReg weights and model defs (https://arxiv.org/abs/2106.10270)
• ResMLP official weights
• ECA-NFNet-L2 weights
• gMLP-S weights
• ResNet51-Q
• Visformer, LeViT, ConViT, Twins
• Many fixes, improvements, better test coverage

3rd Party Vision Transformer Weights

A catch-all (ish) release for storing vision transformer weights adapted/rehosted from 3rd parties. Too many incoming models for one release per source...

Containing weights from:

• Twins - https://github.com/Meituan-AutoML/Twins
• Visformer - danczs/Visformer#2
• NesT (Aggregated Nested Transformer) - weights converted from https://github.com/google-research/nested-transformer by @alexander-soare ' script