Fix hyperparameter predictorbase

[c-w-feldmann]

Description

This PR fixes warnings from lightning, caused by setting parameters.
(Also removes a duplicate line)

Example / Current workflow

from chemprop.nn.predictors import BinaryClassificationFFN
from chemprop.nn.loss import BCELoss
from torch import nn

BinaryClassificationFFN(criterion=BCELoss(), output_transform=nn.Identity())

>> /my_python_path/lib/python3.11/site-packages/lightning/pytorch/utilities/parsing.py:199: Attribute 'criterion' is an instance of `nn.Module` and is already saved during checkpointing. It is recommended to ignore them using `self.save_hyperparameters(ignore=['criterion'])`.
>> /my_python_path/lib/python3.11/site-packages/lightning/pytorch/utilities/parsing.py:199: Attribute 'output_transform' is an instance of `nn.Module` and is already saved during checkpointing. It is recommended to ignore them using `self.save_hyperparameters(ignore=['output_transform'])`

Bugfix / Desired workflow

See PR

[JacksonBurns]

Thanks for the PR! @hwpang are these test failures because the saved checkpoints include the hyperparameters which are now ignored in this PR?

[c-w-feldmann]

Thanks for the PR! @hwpang are these test failures because the saved checkpoints include the hyperparameters which are now ignored in this PR?

When I run pytest locally on the original code I get:

2 failed, 660 passed, 15 skipped, 560 warnings in 940.00s (0:15:39)

My code produces:

17 failed, 645 passed, 15 skipped, 421 warnings in 930.48s (0:15:30) 

I guess it's not my code, but I also don't really understand why these errors have not occurred earlier.

Edit
I forgot to checkout the correct env. I fixed the obtained results. Most errors originate from my fix. Sorry for that. I will try to come up with a better solution.

Edit 2
I think I found a solution, but it is not really good looking. Help would be appreciated.

[JacksonBurns]

@hwpang could you review? Not sure which of these lines are actually required. Pretty sure we need to rebuild the checkpoint files with these hparams ignored.

[hwpang]

Thanks for the PR! This seems fine to me. Could you run this script? The script will regenerate all the checkpoint files we use for tests. After you run the script, please commit those new checkpoint files and push to this PR. This will run the tests with the newly generated checkpoints to ensure that they are indeed correct and compatible.

[c-w-feldmann]

@hwpang

Could you run this script?

Done

[hwpang]

LGTM

[KnathanM]

@hwpang could you review? Not sure which of these lines are actually required. Pretty sure we need to rebuild the checkpoint files with these hparams ignored.

Based on #898, I don't think the checkpoint files needed to be rebuilt because this PR just manually saves criterion and output_transform to hparams instead of having save_hyperparameters do it automatically. If I am wrong and the checkpoint files did need to be rebuilt, then models created in v2.0.0 might not be compatible with v2.0.1 and we should comment on that in the release notes for the next release.

@hwpang or @JacksonBurns can one of you check my thinking on this?

[JacksonBurns]

I'm not sure, deferring this to your judgement

[hwpang]

@KnathanM Thanks for tagging me! I looked into this and the tldr is that the checkpoint file has changed slightly, but they are equivalent.

Using the previous method self.save_hyperparameters() to save criterion and output_transform automatically, the checkpoint_auto['hyper_parameters']['predictor'] has

"activation":       RELU
"cls":              <class 'chemprop.nn.predictors.RegressionFFN'>
"criterion":        None
"dropout":          0.0
"hidden_dim":       300
"input_dim":        300
"n_layers":         1
"n_tasks":          1
"output_transform": UnscaleTransform()
"task_weights":     None
"threshold":        None

And checkpoint_auto['state_dict']['predictor.criterion.task_weights'] has

tensor([[1.]])

Using the current method by setting ignore and saving the criterion and output_transform manually, checkpoint_man['hyper_parameters']['predictor'] has

"activation":       RELU
"cls":              <class 'chemprop.nn.predictors.RegressionFFN'>
"criterion":        MSELoss(task_weights=[[1.0]])
"dropout":          0.0
"hidden_dim":       300
"input_dim":        300
"n_layers":         1
"n_tasks":          1
"output_transform": UnscaleTransform()
"task_weights":     None
"threshold":        None

And the state dict has checkpoint_man['state_dict']['predictor.criterion.task_weights']:

tensor([[1.]])

The only difference I have observed is that the criterion is saved explicitly when we use the manual method. But either of these model files should produce the same predictions when used in inference.