[hparams] save_hyperparameters doesn't save kwargs

[xiadingZ]

❓ Questions and Help

when I use hyperparemeters like docs:

class LitMNIST(LightningModule):

    def __init__(self, layer_1_dim=128, learning_rate=1e-2, **kwargs):
        super().__init__()
        # call this to save (layer_1_dim=128, learning_rate=1e-4) to the checkpoint
        self.save_hyperparameters()

model checkpoint doesn't save args in kwargs. But kwargs is important. Args such as num_frames, img_size, img_std ... must be used in creating dataloader, but it will be tedious if writes them in __init__ explicitly . it can make code clean if hides them in kwargs.

Before I use hparams, it's ok. But now it's not recommended to use hparams, is there any good idea to deal with this problem?

[xiadingZ]

if don't use hparams, it will put all args of model, dataset, dataloader... in a LightnModule' s __init__ method, and save_hyperparameters doesn't save args in kwargs. It this really a good idea?

[Borda]

Could you please share your model example?

[xiadingZ]

Could you please share your model example?

class LitMNIST(LightningModule):

    def __init__(self, layer_1_dim=128, learning_rate=1e-2, **kwargs):
        super().__init__()
        # call this to save (layer_1_dim=128, learning_rate=1e-4) to the checkpoint
        self.save_hyperparameters()
        self.kwargs = kwargs
        ...
    
    def train_dataloader(self):
        img_size = self.kwargs['img_size']
        ...

I can train this model, but when I load from checkpoint, it says kwargs hasn't img_size

[Borda]

I can train this model, but when I load from checkpoint, it says kwargs hasn't img_size

I see, we need to ignore kwargs from the model hparams saving...
@xiadingZ mind adding PR with a test for this case and I ll finish it with a patch?

[s-rog]

Traceback (most recent call last):
  File "./training.py", line 101, in <module>
    main(hparam_trial)
  File "./training.py", line 86, in main
    model = module(hparams, fold_train, fold_val, data_dir+img_dir)
  File "../main/module.py", line 18, in __init__
    self.hparams = hparams
  File "/opt/conda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 638, in __setattr__
    object.__setattr__(self, name, value)
  File "/opt/conda/lib/python3.6/site-packages/pytorch_lightning/core/lightning.py", line 1695, in hparams
    self.save_hyperparameters(hp, frame=inspect.currentframe().f_back.f_back)
  File "/opt/conda/lib/python3.6/site-packages/pytorch_lightning/core/lightning.py", line 1662, in save_hyperparameters
    cand_names = [k for k, v in init_args.items() if v == hp]
  File "/opt/conda/lib/python3.6/site-packages/pytorch_lightning/core/lightning.py", line 1662, in <listcomp>
    cand_names = [k for k, v in init_args.items() if v == hp]
  File "/opt/conda/lib/python3.6/site-packages/pandas/core/generic.py", line 1479, in __nonzero__
    f"The truth value of a {type(self).__name__} is ambiguous. "
ValueError: The truth value of a DataFrame is ambiguous. Use a.empty, a.bool(), a.item(), a.any() or a.all().

I'm guessing this is why 0.8.0 causes this error, it's trying to save all args (including dataframes in my case) outside of hparams?

Edit: #2250