Printing out model layer weights

[rgknox]

Hello. I've generated a trained neural network model in pytorch, and loaded it with Ftorch with the intention of performing inference. But I'd like to verify that I am writing and loading what I think I am, by printing the weights on both sides of the process. In my case it is just a couple of small linear layers and a relu.

In pytorch, I can loop through the items in the state_dict, like so:

    state_dict = model.state_dict()

    for name, tensor in state_dict.items():
        print(f"Parameter: {name}")
        print(tensor)

Are there wrapper functions on the FTorch side that can help with this? I see there is a wrapper for printing out the contents of a tensor, but I'm not sure how to loop through the contents of the model and access those tensors. Thank you for your help!

[joewallwork]

Hi @rgknox. Unfortunately, we don't currently have a way to do this. It seems like it'd be a useful thing for us to have, though! I'll open an issue from this discussion and we'll aim to get something implemented soon.

[rgknox]

Thanks for checking in @jwallwork23 , I do appreciate you all making ftorch public, and I'm making use of it.

[joewallwork]

Hi @rgknox, I've implemented something for this in #488, adding a subroutine called torch_model_print_parameters.

With the proposed changes, running the SimpleNet demo in Fortran will now print

 Model parameters:
_fwd_seq.0.weight:
 2  0  0  0  0
 0  2  0  0  0
 0  0  2  0  0
 0  0  0  2  0
 0  0  0  0  2
[ CPUFloatType{5,5} ]

to indicate that there is one parameter associated with the SimpleNet model, which is a 5x5 weights matrix of 32-bit reals with the values shown.

Is this in line with what you had in mind?