Jit script does not contain hparams

[simeneide]

Hi,
when I try to script my pl module it does not seem to compile the hparams attributes, which I need. An simple example:

import pytorch_lightning as pl
import torch
class Model(pl.LightningModule):
    def __init__(self, *args, **kwargs):
        super().__init__()
        self.save_hyperparameters()

    def forward(self):
        val = self.hparams.par
        return torch.tensor([val])

model = Model(par=1)
model.to_torchscript()

gives the error

RuntimeError: 
Module 'Model' has no attribute 'hparams' :
  File "/tmp/ipykernel_60324/2140052930.py", line 8
    def forward(self):
        val = self.hparams.par
              ~~~~~~~~~~~~ <--- HERE
        return torch.tensor([val])

I can make it work by saving a separate hparams_dict, but this feels a bit clumsy:

#%%
class Model(pl.LightningModule):
    def __init__(self, *args, **kwargs):
        super().__init__()
        self.save_hyperparameters()
        self.hparams_dict = dict(self.hparams)

    def forward(self):
        val = self.hparams_dict['par']
        return torch.tensor([val])

model = Model(par=2)
model.to_torchscript()
#Works

I am running the following on debian:

pytorch-lightning             1.5.3
torch                         1.10.0
torchmetrics                  0.6.0

Is this how it is supposed to work? Or is it possible to add support for this? :)

[rohitgr7]

not sure, but I think I read somewhere that while scripting, pytorch ignores all the custom data structures during compilation to torchscript module.