Remove trainer args from conf/defaults.yaml

[ashnair1]

Closes #347

[ashnair1]

1. While defaults.yaml pinned torchgeo to use newer versions of pytorch-lightning it was useful in documenting what values could be set. Now that it's being removed, users don't really have a reference to set up their training configs. For training args I guess they can refer https://pytorch-lightning.readthedocs.io/en/<version>/common/trainer.html#init (replace version with whatever version of pytorch-lightning they're using).

2. The experiment level arguments (seed, output_dir, log_dir) must be set in all task yamls for training to run. I can add something like this to all the task_defaults configs.

program:  # These are experiment level arguments
  seed: 0
  experiment_name: blah
  overwrite: True
  output_dir: output
  data_dir: data
  log_dir: logs

Edit: only seed, output_dir and log_dir are required in above config

[adamjstewart]

2. Hmm, we may want to keep a defaults.yaml that only contains these defaults. @calebrob6 thoughts?

[ashnair1]

Keeping a defaults.yaml with just these defaults seem like a good idea since they're not related to any library. The issue in #347 was only with trainer args which was dependent on the version of PL.

Plus copying these settings to all configs seems unnecessary

[adamjstewart]

Some of those do actually make sense to override on a dataset-level. Like experiment_name and *_dir could be dataset-specific so all output doesn't end up in the same directory.

[adamjstewart]

Another option is to make all settings (not just the ones in defaults.yaml) optional using something like getattr.

[ashnair1]

Some of those do actually make sense to override on a dataset-level. Like experiment_name and *_dir could be dataset-specific so all output doesn't end up in the same directory.

Basically I was thinking to leave the first half of conf/defaults.yaml more or less as is. So this is all it would contain:

config_file: null  # This lets the user pass a config filename to load other arguments from

program:  # These are the arguments that define how the train.py script works
  seed: 0
  output_dir: output
  log_dir: logs
  overwrite: False

We only need to specify the seed, output_dir and log_dir because they're explicitly used in train.py. experiment_name would go in the specific task.yaml. I made a mistake including it before.

For further customisation of *_dir users would need to write a new train.py right?