LightningModule.load_from_checkpoint not working with .ckpt from 0.7.6

[dscarmo]

🐛 Bug

Trying to use an old experiment .ckpt (generated with ModelCheckpoint(monitor="val_loss", mode="min') in 0.7.6) results in an error when trying to load.

To Reproduce

Steps to reproduce the behavior:

1. Train something with 0.7.6 and save checkpoint with the checkpoint callback
2. Try to load checkpoint with 0.8.1

Trying to load: ckpts/MNIST_0.7.6epoch=4.ckpt with PL version: 0.8.1
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-3-36a7290b43c6> in <module>()
     35 selected_ckpt = glob(os.path.join("ckpts", "*.ckpt"))[0]
     36 print(f"Trying to load: {selected_ckpt} with PL version: {pl.__version__}")
---> 37 LitClassifier.load_from_checkpoint(selected_ckpt)

1 frames
/usr/local/lib/python3.6/dist-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 

/usr/local/lib/python3.6/dist-packages/pytorch_lightning/core/saving.py in _load_model_state(cls, checkpoint, *args, **kwargs)
    184 
    185             if cls.CHECKPOINT_HYPER_PARAMS_TYPE in checkpoint:
--> 186                 model_args = checkpoint[cls.CHECKPOINT_HYPER_PARAMS_TYPE](model_args)
    187 
    188             args_name = checkpoint.get(cls.CHECKPOINT_HYPER_PARAMS_NAME)

TypeError: 'str' object is not callable

Here is a Colab notebook reproducing the issue: https://colab.research.google.com/drive/1s7SzG3EkLJZOmUOkb0mJ85zDcOxBcDLG?usp=sharing

Maybe this is related to the changes in hparams recently?
It happened in my local enviroment and in Google Colab's enviroment.
I was able to load the checkpoint after downgrading to 0.7.6 without problems.

Expected behavior

Being able to use old .ckpts.

Environment

• PyTorch Version (e.g., 1.0): 1.5
• OS (e.g., Linux): Ubuntu 18.04
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source):
• Python version: 3.6.9
• CUDA/cuDNN version: 10.2
• GPU models and configuration: 1060
• Any other relevant information:

Also happened in the Colab enviroment.

[dscarmo]

Added a Google Colab notebook reproducing the issue.

[williamFalcon]

@dscarmo no more bug :) @Borda fixed it!