more pretrained checkpoints for ipCSN irCSN

[richardkxu]

Hi,

First of all, thanks for making this great repo!

Describe the feature I am wondering if you can provide more checkpoints for ipCSN and irCSN? Specifically, I am looking for:

• ipcsn_ig65m_pretrained_r152_32x2x1_58e_kinetics400_rgb.py
• ipcsn_ig65m_pretrained_r50_32x2x1_58e_kinetics400_rgb.py
• ircsn_ig65m_pretrained_r50_32x2x1_58e_kinetics400_rgb.py

There seems to be one ipcsn152 checkpoint in the old mmaction1 model zoo, which I tried to load. However, I got state_dict naming mismatch for all the layers.

Motivation Although ipCSN and irCSN have similar performance, ipCSN preserved the channel interaction, which is useful for research that models the temporal dimension. Resnet50 versions of both ipCSN and irCSN are convenient for finetuning on smaller datasets such as ucf101, since 152 might overfit on smaller datasets.

Related resources There is a model zoo from FB vision team here. However, all the checkpoints are in caffe2 and their scripts are buggy.

Thank you very much!

[richardkxu]

It would also be great if you can provide ipcsn_ig65m_pretrained_r152 and ircsn_ig65m_pretrained_r152, the IG65M pretrained versions without finetuning. These would be very helpful for finetuning on other datasets like ucf101. Thank you!