gridsearch.md

ABnet3 cli documentation

This file explains how to use the ABNet3 cli abnet3-gridsearch

The abnet3 cli is designed to take all its parameters as yaml file. You can see examples of such yaml files in test/data/buckeye.yaml and test/data/test_embedding.yaml.

Run one training

To run one training, you must create your yaml file containing the training parameters, and then run the abnet3-gridsearch parameters.yaml command.

The yaml file is designed as following:

default_params:
  pathname_experience: path/to/experience
  features:
    class: FeaturesGenerator
    arguments:
      run: never
      ...
  dataloader:
    class: OriginalDataLoader
    arguments:
      ...
  sampler:
    class: SamplerClusterSiamese
    arguments:
      run: always
      ...
  model:
    class: SiameseNetwork
    arguments:
      input_dim: 280
      ...
  loss:
    class: coscos2
    arguments:
      ...
  trainer:
    class: TrainerSiamese
    arguments:
      ...
  embedder:
    class: EmbedderSiamese
    arguments:
      ...

You must define all the arguments for the features, dataloader, sampler, model, trainer, loss and embedder classes (and their class name as well).

Inputs and outputs

• The wav files have to be defined in features.arguments.files. This is not mandatory if the features are already generated.
• You can define the path to the output features in features.arguments.output_path
• The cluster file has to be defined in the sampler params
• The input features of the dataloader will automatically be defined as the output of the features generator. But this can be overridden in the arguments
• As well, the input to the embedding will automatically be defined as the output features of the features generator. This can also be overridden.

features.run and sampler.run

The run arguments are special : they allow you to control the feature generation or the sampling.

You can use never, always, once or if_none.

Run a gridsearch over a certain parameter

You can add at the end of the yaml file the following portion :

grid_params:
    sampler:
        arguments:
          type_sampling_mode: ['log','fcube','f','f2','1']

This will run a gridsearch over the parameter type_sampling_mode of the sampler. The gridsearch can only loop over one argument. If you put several arguments in the grid_params, they will be launched one by one.