yaml_inputs.md

Instruction for YAML input forms

There are 2 fields in The App which have to be entered by hand in YAML format: train transforms and optimizer params.

Train transforms

This field's input is based on albumentations Compose and OneOf methods.

If you want to enter transforms, you need to choose transforms from transforms API and to combine two methods mentioned above, including probability of each transform wrt its direct parent. For example, this config

- ShiftScaleRotate:
    shift_limit_y: 0.0625
    shift_limit_x: 0.0225
    p: 0.3
- OneOf:
    args:
      - Blur:
          blur_limit: 
            - 3
            - 3
          p: 1.0
      - GaussianBlur:
          p: 1.0
      - ImageCompression:
          p: 1.0
          quality_lower: 45
          quality_upper: 99
    p: 0.5

will produce Compose of ShiftScaleRotate and OneOf (made from Blur, GaussianBlur and ImageCompression).

Optimizer params

This field is pretty simple. You choose optimizer from torch.optim and enter key-value pairs to YAML form. For example, if you choose optimizer Adadelta, you can enter its params like this:

rho: 0.91
eps: 0.00001
weight_decay: 0.999