Nathan Morgan

Here are some thoughts from our dev meeting today and @kevingreenman and @hwpang, correct me if I'm wrong. 1. Because we are using the lightning framework, the model needs to...

As I wrote this up, I also had two more thoughts. 1. While the training splits should be opt-in, the sizes of the splits are relevant for experiments that test...

I think these are great points and ideas. Two more thoughts/questions: 1. Currently specifying just the flag `--logfile` sends the logfile to `{LOG_DIR}/MODE/{NOW}.log` instead of `{LOG_DIR}/{NOW}.log`. It could be useful...

I like the framework of its more important for commands to be idiomatic than concise. > Probably the easiest solution right now is (2), promoting output to an argument in...

> Each mode can just override `--output`. That is when you have the following block of code: This isn't possible with argparse. Your example gives `argparse.ArgumentError: argument --foo: conflicting option...

In an effort to move towards closing this issue, here is my summary: Saving Needs: - [X] Model architecture and weights - [x] Input and ouput transforms #726 - [...

Hi Arijit, A couple months ago we decided that caching was [unnecessary](https://github.com/chemprop/chemprop/pull/455#discussion_r1325126598) in the new version. It sounds like you might have a use case for caching. Can you open...

@c-w-feldmann Can you remind me your specific use-case for wanting to change the behavior of `get_params()`? In scikit-learn, `get_params()` is used for standard API for estimators (models). In Chemprop, `get_params()`...

Thank you @c-w-feldmann for pinging us again on this. I have tested the changes and get the same `repr(obj)` behavior for the loss functions, so I think this PR would...

> I'm in favor of closing this PR and moving `nn.LossFunction` over to using [`extra_repr()`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module.extra_repr) rather than defining some custom `__repr__()` method for this class. This is the "official" way...