Grid search

That is an example of a grid_search configuration:

executors:
  train:
    type: catalyst
    args:
      config: catalyst.yml
    grid:
      - batch_size: [20, 40, 80]
      - - num_workers: 2
          lr: 0.01
        - num_workers: 3
          lr: 0.1

The user needs to specify parameter sets in a grid field.

MLComp considers all possible combinations of them and transfers each combination to an executor within a special parameter: additional_info.

Each parameter set can be specified with 4 different ways:

1. key: list

   for example, batch_size: [20, 40, 80]

2. list of dicts

   for example,

   - - num_workers: 2
       lr: 0.01
     - num_workers: 3
       lr: 0.1
   

3. _folder: path_to_folder

   for example,

   _folder: configs/
   

   Each config in the folder must be a valid yml file.

   File structure is the same as in number 4.

4. _file: list of files

   for example,

   _file: [file1.yml]
   

   File structure is an ordinary dictionary. For example,

   num_workers: 2
   lr: 0.01
   stages:
       stage2:
           optimizer_params:
             optimizer: Adam
             lr: 0.001
             weight_decay: 0.0001