Slightly incorrect documentation for VAE

[jeremyadamsfisher]

The example for a VAE indicates that from_pretrained is a static method

https://github.com/PyTorchLightning/pytorch-lightning-bolts/blob/4e1c1502b70e4be59a9ac80878ec5ad5a212f87e/pl_bolts/models/autoencoders/basic_vae/basic_vae_module.py#L30

Indeed, running this line causes as error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-64-4c7d0e8ca436> in <module>
----> 1 vae = VAE.from_pretrained('cifar10-resnet18')

TypeError: from_pretrained() missing 1 required positional argument: 'checkpoint_name'

This is because the method is actually implemented nonstatically:

def from_pretrained(self, checkpoint_name):
    if checkpoint_name not in VAE.pretrained_urls:
        raise KeyError(str(checkpoint_name) + ' not present in pretrained weights.')

    return self.load_from_checkpoint(VAE.pretrained_urls[checkpoint_name], strict=False)

Seems to me either:

• the documentation should to reflect this (vae = VAE(32).from_pretrained(...))
• the method should be rewritten

@classmethod
def from_pretrained(cls, checkpoint_name):
    # snip
    return self.load_from_checkpoint(cls(32).pretrained_urls[checkpoint_name], strict=False)

Here is a minimal failing example on colab: https://colab.research.google.com/drive/1quvrQUyCIMC7Cq9QCjaJfip5hsi9jwSJ?usp=sharing

Happy to write a PR, if desired

[github-actions]

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

[akihironitta]

@jeremyadamsfisher Hi, thank you for reporting the issue! Following the suggestion by @ananyahjha93 in #200 (comment), let's just fix the docs for now.

2. from_pretrained() needs to be an instance method and not a static method. In most cases, you will initialize the lightning module with specific params according the the weights being loaded.

[akihironitta]

@jeremyadamsfisher Would you be interested in submitting a PR?

[jeremyadamsfisher]

Yeah, no problem