Fix command line argument

[wiederm]

Description

There were two separate bugs affecting command-line arguments:

• the device can be an int or a List[int], this was not correctly handled by the argparse type check
• command line arguments should override parameters provided in the TOML file. The read_config function was doing the opposite (the docstring described the correct behavior)

Status

• Ready to go

[chrisiacovella]

This looks good. Good catch on parsing the devices; I had this initially as just an int when I checked, then realized it could be a list and didn't properly test apparently (incorrectly assuming it would properly coerce the input). I pushed a minor change that looks for the presence of a comma instead of square brackets, so the input can be either "[0,1]" or "(0,1)" or "0,1"

[wiederm]

Unfortunately, I think that won't work! It's kind of counterintuitive how this parameter works: when you pass an int, it means grab that number of GPUs, if you pass a List[int], it means grab that specific GPU device. So, a common use case is to pass [1] to grab GPU:1 on a multi-GPU node. I think checking for , will miss that use case.

[chrisiacovella]

Unfortunately, I think that won't work! It's kind of counterintuitive how this parameter works: when you pass an int, it means grab that number of GPUs, if you pass a List[int], it means grab that specific GPU device. So, a common use case is to pass [1] to grab GPU:1 on a multi-GPU node. I think checking for , will miss that use case.

Hmm, I'll switch it back, I guess I misunderstood what this is doing. Maybe we need to eventually change the input syntax to make it clearer. Like two separate arguments, number_of_devices (automatically picks) and selected_devices (you pick which ones). Even if internally we don't differentiate the functionality of those, it might be clearer to convey what is being done.

[wiederm]

Yeah, this is confusing! THat wasn't my idea, but is what PyTorch Lighning is expecting

[chrisiacovella]

I also commented out the test figshare download test as reading from their API seems to randomly be timing out/failing on some of the CI nodes (inconsistent failures). I plan to refactor the code and the download from figshare function will be removed.