pytorch-image-models

pytorch-image-models copied to clipboard

[BUG] Truncated normal initialization

2

I'm a bit confused with usage of trunc_normal here https://github.com/rwightman/pytorch-image-models/search?p=2&q=trunc_normal_&type=code I'm wondering why you are not using a and b parameters. Looking into TF implementation https://www.tensorflow.org/api_docs/python/tf/random/truncated_normal > The values are...

NProkoptsev

The models available in this repository are highly satisfactory, but is there a possibility to explore the [CoAtNet](https://proceedings.neurips.cc/paper/2021/hash/20568692db622456cc42a2e853ca21f8-Abstract.html) model in some benchmark dataset? (e.g: ImageNet [1k, 21k], COCO, ...)

calusbr

Video dataset support

4

Thanks for your awesome work, is there any plan to support video dataset such as kinetics or further video tasks?

Epiphqny

Add CUDA Graph and AOT Autograd support

7

Add CUDA Graph support with `--cuda-graph` and AOT Autograd support with `--aot-autograd` to **benchmark.py** and **train.py** The workflow for cuda graph in train.py might be a bit overcomplicated. Related: https://github.com/rwightman/pytorch-image-models/issues/1244

xwang233

Add `list_benchmarks` to compare models easily

2

This is a simple function for listing ImageNet benchmarks found [here](https://github.com/rwightman/pytorch-image-models/blob/master/results/results-imagenet.csv) for users to see

MohamedAliRashad

[FEATURE] Add segformer as feature extractors ...

1

https://github.com/NVlabs/SegFormer/blob/master/mmseg/models/backbones/mix_transformer.py#L374 https://drive.google.com/drive/folders/1b7bwrInTW4VLEm27YawHOAMSMikga2Ia I think segformer is a good backbone / feature extractor. It provides pretrained weights so I think maybe it is good to introduce into timm.

jiarenyf

Is there a way to load huggingface models directly using timm, it's not working ? Background: SEER combines a recent architecture family, RegNet, with an online self-supervised training to scale...

PrithivirajDamodaran

**Is your feature request related to a problem? Please describe.** ViT-S_16 SAM can be adopted as described in https://console.cloud.google.com/storage/browser/vit_models/sam;tab=objects?pli=1&prefix=&forceOnObjectsSortingFiltering=false. **Describe the solution you'd like** Add the model information in timm/models/vision_transformer.py...

Harry24k

**Is your feature request related to a problem? Please describe.** Models that are invariant under non-differentiable impairments like JPEG compression, video compression (H265, VP9, etc), etc. **Describe the solution you'd...

pfeatherstone

Cosine decay schedule changes with different accumulate_grad_batches values

1

When using the accumulate_grad_batches flag in a Pytorch Lightning trainer, I noticed it changes my scheduler LR differently. I am using TIMM's cosine decay scheduler. And when I use 1...

exnx