Support model selection in tools

[corneliusboehm]

Description

Support using different model versions in all tools.

General Changes

• Extracted features will now be stored per model, so the directory structure looks like this:

• features_train/
  • StridedInflatedEfficientNet-pro/
    • num_layers_to_finetune=X/
      • video_0.npy
      • video_1.npy

• Methods for obtaining all relevant directories have been added to tools/utils.py
• The method get_weights was moved into the ModelConfig class so that it's also possible to obtain the weights for a specific model config

SenseStudio

• The best available model is automatically discovered and used for annotation
• In the future we should add a dropdown menu on the UI for selecting from the available models

Training

• The training script can now be provided with a model_name and model_version. If those are not provided, the best available model will automatically be used.
• The parameters for a training run will be stored in a config.json file. This can later for example be used to look up the correct backbone model.

  {
  'backbone_model_name': selected_config.model_name,
  'backbone_model_version': selected_config.version,
  'num_layers_to_finetune': num_layers_to_finetune,
  'classifier': str(gesture_classifier),
  'temporal_training': temporal_training,
  'lr_schedule': lr_schedule,
  'num_epochs': num_epochs,
  }

Running custom classifier

• The backbone network that was used for training is now looked up from the config.json file

Conversion

• Now requires backbone_name and backbone_version for pre-trained classifiers or path_in for a custom classifier
• For a custom classifier, the corresponding backbone model will be looked up from the training config

[corneliusboehm]

Ready for a final review, so we can merge what we have so far.

These questions are still open and can be addressed later:

• Call --backbone_version --backbone_variant instead to avoid confusion with a numerical version? Or merge version into --backbone_name, so that users only have to indicate one?
• In the example scripts, the command line parameters are called --model_name and --model_version, while the tools call it --backbone_name. Should we unify those?
• Should we shorten all model names to SIEN/SI-EN and SIMN/SI-MN?
• "Fine-tune" instead of "finetune"?

[guillaumebrg]

I haven't finished reviewing yet... but regarding your questions:

Call --backbone_version --backbone_variant instead to avoid confusion with a numerical version? Or merge version into --backbone_name, so that users only have to indicate one?

No strong opinion here. I'm fine with all 3 solutions (leave it like this, rename to "backbone_variant" or merge into "backbone_name"). I suggest we leave it like this in this PR and change that in another PR if anyone has a strong opinion about this and wants to do it.

In the example scripts, the command line parameters are called --model_name and --model_version, while the tools call it --backbone_name. Should we unify those?

Yes, I think that would be good.

Should we shorten all model names to SIEN/SI-EN and SIMN/SI-MN?

I vote "yes".

"Fine-tune" instead of "finetune"?

No strong opinion