longlong-jing / 3drotnet

Code for Self-supervised Spatiotemporal Feature Learning by Video Geometric Transformations

3DRotNet

Code for Self-supervised Spatiotemporal Feature Learning by Video Geometric Transformations

3drotnet's Issues

Very good job and when is open source? Thank you!

Very good job and when is open source? Thank you!

Video dataloader for finetuning

There's no code available for finetuning. Dataloader should be different as well as fine-tuning code file should be separate. Isn't that so?

The transforms during fine-tuning are same as training for pretext task?

For testing, what are the transforms used?

The naming of the datasets

Hi,

Great work! Thanks for releasing the code!
When I am trying to run this code, I noticed there're some list files like: './MITS_split/MITS.lst' or './KITS_split/KITS.lst'.
Is it possible that you could share these files with me? Or maybe the preprocessing code. It would be of great help.
Thanks :)

Yutong

pre-training model

Would you please provide the pre-training model in Kinetics? Thank you

Problems in reproducing results

Thanks for the amazing work in self-supervision. I was trying to
reproduce the results and ran into some problems. I
modified the code from
https://github.com/kenshohara/3D-ResNets-PyTorch for self-supervision
task as mentioned in your paper. In particular, I am trying to
reproduce the result of Kinetics dataset with 100000 training videos.
However, my model seems to converge quickly and results in
overfitting. I am unable to find any bug. Did you make any other
changes (like momentum, optimizer parameters, weight decay etc) to the code of 3D Resnet? Could you share some more details about the implementation?

Here are the training and validation logs for 7 epochs. For validation, I am using ucf101 dataset.

Training logs
epoch loss acc lr
1 0.801 0.60 0.1
2 0.553 0.68 0.1
3 0.487 0.70 0.1
4 0.461 0.71 0.1
5 0.443 0.71 0.1
6 0.434 0.72 0.1
7 0.430 0.72 0.1

Validaiton loss
epoch loss acc
1 1.44 0.282
2 1.41 0.282
3 1.78 0.246
4 1.41 0.247
5 1.49 0.243
6 1.68 0.250
7 1.45 0.256
~