Add first scratch of implementation for maestro CLI

[PawelPeczek-Roboflow]

Description

Did not have chance to fully test it e2e due to lack of GPU machines.

Failed to create the CLI as generic as I wanted:

• typer is quite sophisticated regarding generating command hints but it does not support some of the typings - like Unions or Literals (and I am not sure if other CLI libs will)
• unfortunately - it seems like the only option to get list of args for train / val commands is to declare them explicitly - initially my plan was to create function to assembly typings based on dataclass (or even better pydantic model) - and then what we would gain is the experience that you create training config and we generate command args based on the config class
• decided to declare training config in CLI entrypoint
• alternatively - we could accept path to config and kwargs to override, but then people would not be able to see nice command help that we have now - to cover for that, we could have additional command to list all config fields (which could happen automatically based on introspection of dataclass) - so you would have command maestro <model> train and maestro <model> list-train-parameters which would give you hints on what params the train command accept - not sure if that is better than explicit params

What I managed to do:

• put the foundation of CLI structure - I imagine maestro CLI have some generic commands - like maestro info - which are implemented at the top-level of CLI
• then each model recipe would implement sub-CLI app, which would expose its custom commands (we do not put any constraints on what that commands could be)
• we have introspection mechanism at top level of CLI which dynamically load recipes - now based on what can be imported

Examples of usage:

❯ python -m maestro.cli.main info
Welcome to maestro CLI. Let's train some VLM! 🏋

❯ python -m maestro.cli.main --help

 Usage: python -m maestro.cli.main [OPTIONS] COMMAND [ARGS]...

╭─ Options ────────────────────────────────────────────────────────────────────────────────────╮
│ --install-completion          Install completion for the current shell.                      │
│ --show-completion             Show completion for the current shell, to copy it or customize │
│                               the installation.                                              │
│ --help                        Show this message and exit.                                    │
╰──────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ Commands ───────────────────────────────────────────────────────────────────────────────────╮
│ florence2   Fine-tune and evaluate Florence 2 model                                          │
│ info        Display information about maestro                                                │
│ paligemma   Fine-tune and evaluate PaliGemma model                                           │
╰──────────────────────────────────────────────────────────────────────────────────────────────╯

❯ python -m maestro.cli.main florence2 --help

 Usage: python -m maestro.cli.main florence2 [OPTIONS] COMMAND [ARGS]...

 Fine-tune and evaluate Florence 2 model

╭─ Options ────────────────────────────────────────────────────────────────────────────────────╮
│ --help          Show this message and exit.                                                  │
╰──────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ Commands ───────────────────────────────────────────────────────────────────────────────────╮
│ evaluate   Evaluate Florence 2 model                                                         │
│ train      Train Florence 2 model                                                            │
╰──────────────────────────────────────────────────────────────────────────────────────────────╯

❯ python -m maestro.cli.main florence2 train --help

 Usage: python -m maestro.cli.main florence2 train [OPTIONS]

 Train Florence 2 model

╭─ Options ────────────────────────────────────────────────────────────────────────────────────╮
│ *  --dataset_location                               TEXT     Path to directory with dataset  │
│                                                              [default: None]                 │
│                                                              [required]                      │
│    --model_id_or_path                               TEXT     Model to be used or path to     │
│                                                              your checkpoint                 │
│                                                              [default:                       │
│                                                              microsoft/Florence-2-base-ft]   │
│    --revision                                       TEXT     Revision of Florence2 HF        │
│                                                              repository                      │
│                                                              [default: refs/pr/20]           │
│    --device                                         TEXT     CUDA device ID to be used (in   │
│                                                              format: 'cuda:0')               │
│                                                              [default: cuda:0]               │
│    --transformers_cache_dir                         TEXT     Cache dir for HF weights        │
│                                                              [default: None]                 │
[...]

❯ python -m maestro.cli.main florence2 train --dataset_location some
Training configuration
{
    'dataset_location': 'some',
    'model_id_or_path': 'microsoft/Florence-2-base-ft',
    'revision': 'refs/pr/20',
    'device': device(type='cuda', index=0),
    'transformers_cache_dir': None,
    'training_epochs': 10,
    'optimiser': 'adamw',
    'learning_rate': 1e-05,
    'lr_scheduler': 'linear',
[...]

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

How has this change been tested, please provide a testcase or example of how you tested the change?

YOUR_ANSWER

Any specific deployment considerations

For example, documentation changes, usability, usage/costs, secrets, etc.

Docs

• Docs updated? What were the changes: