david graff

I think what we might want to consider is abolishing the CLI concept of a GPU entirely and moving our parameters entirely over to the Lightning spec: `--accelerator` and `--devices`....

> This is nice on UNIX but can't be done on Windows AFAIK. Windows does not allow per-line environment variables and requires setting them globally, calling the command, and then...

I agree with @JacksonBurns's last point. V1 was highly flexible and supported many features (like these) at the cost of the internal code-base being almost illegible and a real mess...

What about adding a split_column to the input dataset? If the argument is specified, then we use that column in the dataset map a datapoint to a split. The column...

The current model builds splits and loads data in one function, [`build_splits`](https://github.com/chemprop/chemprop/blob/fc036fb9fb3b7d9eda72341d719cc27eb4df0a1f/chemprop/cli/train.py#L521), which (1) loads in an input file and then (2) splits it into train/val/test depending on the values...

The problem with adding additional features to this input file as that arrays don't play nicely within CSVs and strings don't play nicely with binary data files meant for storing...

- probably want to mirror the arg names, i.e., `--model-dir` and `--checkpoint-path` rather than just `--checkpoint` - a `model_dir` contains a model's serialized structure, weights, and scalers, whereas a `checkpoint`...

I don't think it's a simple drop-in replacement as the APIs for `torch_scatter.scatter_sum` and `torch.scatter_add` are fairly different, but it seems it can be done: ```python import torch from torch_scatter...

for MacOS I have to install `torch-scatter` like so: ``` $ TORCH=2.0.0 $ CUDA=cpu $ pip install -f https://data.pyg.org/whl/torch-${TORCH}+${CUDA}.html --no-index ```

some brief notes (excuse the mixing of atom/node and bond/edge): - the primary difference of polymeric message passing is the introduction of atom and bond weights into the `BatchMolGraph` ([here](https://github.com/coleygroup/polymer-chemprop/blob/ae1ae1a86776cbf0860fde5d4823938acd0a794d/chemprop/models/mpn.py#L81C33-L81C33))...