Proof of concept for command-line training #821
Conversation
|
Super cool, thanks @jscanvic ! |
Andrewwango
left a comment
There was a problem hiding this comment.
Some initial comments, this is super cool, thanks Jérémy!
| @@ -0,0 +1,23 @@ | ||
| dataset: |
There was a problem hiding this comment.
Do you think we should provide somewhere in the repo example/template configs? or perhaps even in examples, each example involving training can be used in either notebook mode or using the provided config that we provide?
| import argparse | ||
|
|
||
|
|
||
| class CommandLineTrainer: |
There was a problem hiding this comment.
For future extensibility, can we generalise this to just a CLIParser or something, that handles all the config ingestion, and then maybe have CommandLineTrainer as a subclass or something with only the core Trainer orchestration?
| num_workers=2, | ||
| ) | ||
| trainer = dinv.training.Trainer( | ||
| epochs=10, |
There was a problem hiding this comment.
What do you think of doing a one-to-one mapping between config params and trainer options, such that this can be reduced to just like **parsed_kwargs?
| import argparse | ||
|
|
||
|
|
||
| class CommandLineTrainer: |
| args: argparse.Namespace = parser.parse_args() | ||
|
|
||
| if args.command == "train": | ||
| trainer = CommandLineTrainer() |
There was a problem hiding this comment.
Maybe we could have a cli submodule which contains all the orchestrators e.g. CommandLineTrainer, CommandLineOptim etc. to keep this __main__ clean. And also keep this argparser clean, and the individual files in cli can contain the task-specific args
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #821 +/- ##
==========================================
- Coverage 83.79% 83.21% -0.58%
==========================================
Files 204 205 +1
Lines 20503 20643 +140
Branches 2807 2840 +33
==========================================
- Hits 17180 17179 -1
- Misses 2410 2550 +140
- Partials 913 914 +1 ☔ View full report in Codecov by Sentry. |
A proof of concept towards #496
Usage
Denoising
Super-resolution
Example