search

facebookresearch / vissl

VISSL is FAIR's library of extensible, modular and scalable components for SOTA Self-Supervised Learning with images.
https://vissl.ai
MIT License
3.25k stars 331 forks source link

ResNet101 weights it seems? #482

Open sayakpaul opened 2 years ago

sayakpaul commented 2 years ago

📚 VISSL Documentation

Documentation: https://vissl.ai/tutorials/Using_a_pretrained_model_for_inference_V0_1_6.

I am referring to this section (Download the ResNet-50 Simclr weights from the Model Zoo) of the above-mentioned tutorial. Even though it explicitly mentions ResNet50 the download link seems to be indicating the weights of ResNet101:

wget -q -O /content/resnet_simclr.torch https://dl.fbaipublicfiles.com/vissl/model_zoo/simclr_rn101_1000ep_simclr_8node_resnet_16_07_20.35063cea/model_final

Possible to ensure the validity i.e., which model architecture do the weights correspond to?

iseessel commented 2 years ago

Indeed we should just change ResNet50 to ResNet101 in the tutorail -- nice catch!

sayakpaul commented 2 years ago

Thanks! So if we change the identifier to rn50, should it be ResNet50? Or is there a different handle?

sayakpaul commented 2 years ago

@iseessel just a gentle ping.

iseessel commented 2 years ago

Sorry about the delay -- thanks for pinging me, missed this!

There are two options:

  1. Download the Resnet-50 weights instead of the Resnet-101 weights. You can see these here: https://github.com/facebookresearch/vissl/blob/main/MODEL_ZOO.md#SimCLR

  2. Continue to use the Resnet-101 weights and change the config to use Resnet-101 model here: https://github.com/facebookresearch/vissl/blob/main/configs/config/pretrain/simclr/models/resnext101.yaml

sayakpaul commented 2 years ago

Thanks much. Do you want me to keep the issue open until the tutorial is updated? I can also do a PR with the correction.

iseessel commented 2 years ago

Yeah if you don't mind sending a correction PR that would be great!

Let's do option 2 to give the tutorials some variety in terms of types of models used.

sayakpaul commented 2 years ago

Alright. Will do it.