Compute custom video with COMEDIAN about eztorch

[edugzlez]

Yes @juliendenize, thank you very much. I have already been able to use the pre-trained network.

Thank you very much and congratulations on your results.

[juliendenize]

Hello @edugzlez, thanks for taking an interest in our work.

I think the model loaded properly, it might have been a better idea on my part to remove from the state_dict the transformation parameters as they are not pertinent and should be config-based. I think that you loaded it right.

Still, I would suggest that you filter the transform keys from the model state_dict if you do not wish to use any transforms. You can do something like this (I did not proofread it but I hope you get the idea):

config.model.train_transform=None # Remove train_transform initialization
config.model.val_transform=None # Remove val_transform initialization
config.model.test_transform=None # Remove test_transform initialization

model = ....

state_dict = torch.load(...)
filtered_state_dict = state_dict.copy()
for key in state_dict.keys():
   if key.startswith(("train_transform","val_transform","test_transform")): # Remove transforms from state_dict
      del filtered_state_dict[key]
model.load_state_dict(filtered_state_dict) # Only load model parameters

However, keep in mind you would have to resize and normalize the input data before passing them to the model.

If I am right the model._orig_mod.trunk layer directly extracts the features every 2 frames.

This is right indeed you should obtain 64 tokens at the output if you pass 128 frames.

Did I answer correctly to your questions?