Error when loading with LitModel.load_from_checkpoint(path)?

[MicPie]

❓ Questions and Help

Before asking:

1. search the issues. Done
2. search the docs. Done

What is your question?

When I'm loading a Lightning model with LitModel.load_from_checkpoint(path) I always get this error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-11-5cfc485836df> in <module>
      1 # OPTION 2:
      2 # test after loading weights
----> 3 model = LitModel.load_from_checkpoint("lightning_logs/version_0/checkpoints/epoch=2.ckpt")

~/anaconda3/envs/pytorch-lightning/lib/python3.7/site-packages/pytorch_lightning/core/saving.py in load_from_checkpoint(cls, checkpoint_path, map_location, hparams_file, tags_csv, *args, **kwargs)
    169         checkpoint[cls.CHECKPOINT_HYPER_PARAMS_KEY].update(kwargs)
    170 
--> 171         model = cls._load_model_state(checkpoint, *args, **kwargs)
    172         return model
    173 

~/anaconda3/envs/pytorch-lightning/lib/python3.7/site-packages/pytorch_lightning/core/saving.py in _load_model_state(cls, checkpoint, *args, **kwargs)
    195 
    196         # load the state_dict on the model automatically
--> 197         model = cls(*args, **kwargs)
    198         model.load_state_dict(checkpoint['state_dict'])
    199 

TypeError: __init__() takes 1 positional argument but 2 were given

Code

Simple example is from https://pytorch-lightning.readthedocs.io/en/latest/new-project.html.
Custom LightModul is very similar but I am using other data, sets, and loaders and a torchvision ResNet.

What have you tried?

I tried different checkpoints as well as relative and absolute paths.
Checkpoints are loaded from lightning_logs/version_0/checkpoints/ which where saved automatically after stopping training.

Setup and training itself works without problems.

I also get the same error with the more complex setup.

Maybe I don't see something super obvious?

Thank you very much & kind regards

What's your environment?

• OS: Ubuntu 18.04
• Packaging: Conda
• Version:
  pytorch_select 0.2
  pytorch 1.3.1
  pytorch-lightning 0.8.0

[github-actions]

Hi! thanks for your contribution!, great first issue!

[dscarmo]

Might be related to #2334
If you put an hparams argument in your LitModel does it work? What about using 0.8.1?

[MicPie]

Thank you for your reply!

This code snippet gives me the same error:

model = LitModel.load_from_checkpoint(checkpoint_path="lightning_logs/version_0/checkpoints/epoch=2.ckpt",
                                      hparams_file="lightning_logs/version_0/hparams.yaml")

Via conda I get no update and also tried to upgrade with pip but still 0.8.
Do I need for 0.8.1 an pip editable install via the GitHub repo?

[dscarmo]

Sorry, I meant changing the init to (self, hparams).

To me it looks the load_from_checkpoint is trying to force hparams (2 arguments given)
But this might have been fixed in 0.8.1.

Try to pip uninstall pytorch-lightning and/or pip install pytorch-lightning --upgrade

[MicPie]

Thank you, I was able to install the 0.8.1 version with this trick.

When I add hparams to the def __init__(self, hparams) to my lightning module I have to supply it during object creation too?

When I load the checkpoint with torch.load everything seems to be there.

I suspect the args supplied to cls._load_model_state(checkpoint, *args, **kwargs) are the reason this happens (kwargs are empty), will check later in detail.

[Borda]

I suspect the args supplied to cls._load_model_state(checkpoint, *args, **kwargs) are the reason this happens (kwargs are empty), will check later in detail.

yes, you need to allow your class to pass arguments to the parent class, otherwise, it seems duplicate to #2386