[rllib] Train script overwrites framework field in experiment config argument

[konradkalita]

[rllib]

What is the problem?

Framework setting passed to training script via config argument is ignored, because training script uses separate arguments for this field (--torch, --eager). If none of these args is passed to the script the framework is set to tf, regardless of options passed in --config argument.

Expected behavior

Script should use default value for framework field only if it is not already present in config dict (it can be passed as cmd argument or be loaded from file).

Reproduction (REQUIRED)

Run rllib train --run DQN --env CartPole-v0 --config "{\"framework\": \"tfe\"}" - script uses tf instead of tfe

• I have verified my script runs in a clean environment and reproduces the issue.
• I have verified the issue also occurs with the latest wheels.

[sven1977]

This PR should fix it. #9572
To patch for now (until this is merged): Remove the entire else block in line 188 (ray/rllib/train.py). This way, it'll:

1. Use torch iff --torch
2. Use tf-eager iff --eager
3. Use config framework (or default tf if not specified in config).