Mvitv2 dataset

[xqc-qc]

What is the format of the MVITV2 dataset and how to make it?

[xqc-qc]

Traceback (most recent call last):
File "tools/run_net.py", line 52, in
main()
File "tools/run_net.py", line 27, in main
launch_job(cfg=cfg, init_method=args.init_method, func=train)
File "/root/SlowFast/slowfast/utils/misc.py", line 430, in launch_job
func(cfg=cfg)
File "/root/SlowFast/tools/train_net.py", line 555, in train
checkpoint_epoch = cu.load_checkpoint(
File "/root/SlowFast/slowfast/utils/checkpoint.py", line 558, in load_checkpoint
optimizer.load_state_dict(checkpoint["optimizer_state"])
File "/root/miniconda3/envs/slowfast/lib/python3.8/site-packages/torch/optim/optimizer.py", line 146, in load_state_dict
raise ValueError("loaded state dict contains a parameter group "
ValueError: loaded state dict contains a parameter group that doesn't match the size of optimizer's group

What I did was K400 data format, why is it wrong?

[alpargun]

Can you share your experiment config file?

You can refer to provided MViTv2 experiment config files:

• MViTv2 supervised training on Kinetics
• MViTv2 masked self-supervised training on Kinetics

However, looking at your error, it seems like the problem is due to loading the model and the checkpoint, so could be due to the head network or incorrect config file. Providing your config file will also help track the error.

[xqc-qc]

Can you share your experiment config file?

You can refer to provided MViTv2 experiment config files:

• MViTv2 supervised training on Kinetics
• MViTv2 masked self-supervised training on Kinetics

However, looking at your error, it seems like the problem is due to loading the model and the checkpoint, so could be due to the head network or incorrect config file. Providing your config file will also help track the error.
config.txt

[AbrarKhan009]

hello @xqc-qc did your problem resolve? actually i am also working on mvitv2 model and my first task is to run mvitv2 and replicate the same results by using same dataset.

[alpargun]

@xqc-qc,

config.txt

Looking at your config file, I understand that you would like to train on AVA dataset.

Your NUM_CLASSES is 7, but the config file for AVA in configs/AVA/SLOWFAST_32x2_R50_SHORT.yaml has NUM_CLASSES: 80

So, the model checkpoint you downloaded should have 80 output neurons but here you are initializing a model with 7 output neurons. You need to set NUM_CLASSES as it appears in the original config file to restore the checkpoint and the state dictionary successfully.

If you did that on purpose, i.e., you want 7 outputs for your application, so you want to train a new last layer (linear probing), then you can check out fine-tuning in SlowFast.

[xqc-qc]

@xqc-qc,

config.txt

Looking at your config file, I understand that you would like to train on AVA dataset.

Your NUM_CLASSES is 7, but the config file for AVA in [configs/AVA/SLOWFAST_32x2_R50_SHORT.yaml](https://github.com/facebookresearch/SlowFast/blob/main/configs/AVA/SLOWFAST_32x2_R50_SHORT.yaml) has NUM_CLASSES: 80 So, the model checkpoint you downloaded should have 80 output neurons but here you are initializing a model with 7 output neurons. You need to set NUM_CLASSES as it appears in the original [config file](https://github.com/facebookresearch/SlowFast/blob/main/configs/AVA/SLOWFAST_32x2_R50_SHORT.yaml) to restore the checkpoint and the state dictionary successfully.

If you did that on purpose, i.e., you want 7 outputs for your application, so you want to train a new last layer (linear probing), then you can check out fine-tuning in SlowFast.

Thank you very much, thank you for your answer！