Can you share the pretrained models?

Hi,
First of all, Thank you so much for your work and sharing your code and pretrained weights!
I would appreciate if you could help me with a peroblem regarding loading the network

so i'm looking at the code an experimenting with it (more specifically the demo code)
there I see that the network is loaded in lines (320-322 demo-> run.py)

        encoder, regressor = nn_output()
        encoder.load_state_dict(torch.load('/home/daehakim/Desktop/FER_int_system/AVCE_demo/weights/Sparse_cont_enc.t7'), strict=False)
        regressor.load_state_dict(torch.load('/home/daehakim/Desktop/FER_int_system/AVCE_demo/weights/Sparse_cont_reg.t7'), strict=False)

but when loading the weights i get missing keys warning in the reggressor line

_IncompatibleKeys(missing_keys=['lin2.weight', 'lin2.bias', 'pw_conv.weight', 'pw_conv.bias'], unexpected_keys=[])

could it be that the weights file Sparse_cont_reg.t7 is corrupted / not up to date? could you please update it?

Thanks

Do you know how to annotate valence and arousal value for each image ?

May I inquire about the dataset on which the model "Sparse_..._.t7" in AVCE_demo is trained? I have noticed a significant disparity between the predicted outcomes and the labels when evaluating it on datasets like AFEW-VA. Could you kindly provide some insights into this matter?

Original line was : y.register_hook(lambda grad,b=batch : torch.solve(grad[:,None], J[b].transpose(0,1))[0][:,0])

I changed to torch.linalg.solve, but still got error as
RuntimeError: linalg.solve: A must be batches of square matrices, but they are 65 by 1 matrices

I changed the code to the
y.register_hook(lambda grad,b=batch : torch.linalg.solve(J[b].transpose(0,1),grad[:,None])[:,0])
By this change, code works but I am not sure if operations are right ?