How to write OTX Configuration (recipe)#

Configuration#

Example of recipe/classification/multi_class_cls

model:
    class_path: otx.algo.classification.mobilenet_v3.MobileNetV3ForMulticlassCls
    init_args:
        label_info: 1000
        light: True

optimizer:
    class_path: torch.optim.SGD
    init_args:
        lr: 0.0058
        momentum: 0.9
        weight_decay: 0.0001

scheduler:
  class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
  init_args:
    num_warmup_steps: 10
    main_scheduler_callable:
      class_path: lightning.pytorch.cli.ReduceLROnPlateau
      init_args:
        mode: max
        factor: 0.5
        patience: 1
        monitor: val/accuracy

engine:
    task: MULTI_CLASS_CLS
    device: auto

callback_monitor: val/accuracy
data: ../../_base_/data/torchvision_base.yaml

We can use the ~.yaml with the above values configured.

  • engine

  • model, optimizer, scheduler

  • data

  • callback_monitor

The basic configuration is the same as the configuration configuration format for jsonargparse. Jsonargparse Documentation <https://jsonargparse.readthedocs.io/en/v4.27.4/#configuration-files>_

Configuration Overrides#

Here we provide a feature called overrides. This feature allows you to override the values need from the default configuration. Currently, you can overrides data, callbacks, logger, and other single value configurations. Also, you can use reset to reset the configurations from the default values to the new values.

To update single value configurations, just set them in the overrides.

...
overrides:
  max_epochs: 10 # update to 10
  data:
    image_color_channel: BGR # update to BGR
...

If you want to add new configuration which isn’t set before, just set them in the overrides and it will be appended.

...
overrides:
  new_configuration: 1 # add new_configuration with 1
...

Note

You can see the final configuration with the command below.

$ otx train --config <config-file-path> --print_config

Data overrides#

data can currently be provided as a list of different transforms. The way to override this is as follows.

Let’s try to change the size of Resize and the prob of RandomFlip which are already set in base data configuration of instance segmentation. To change them, you can just set the values in the overrides.

...
overrides:
  data:
    train_subset:
      transforms:
        - class_path: otx.core.data.transform_libs.torchvision.Resize
          init_args:
            size: # update `size` from 1024 to 512
              - 512
              - 512
        # Pad is used as is because it is not set here
        - class_path: otx.core.data.transform_libs.torchvision.RandomFlip
          init_args:
            prob: 0 # update `prob` from 0.5 to 0
        # ToDtype and Normalize are used as is because they are not set here
...

Like single value configurations, when adding new transforms in overrides it will be appended.

...
overrides:
  data:
    train_subset:
      transforms:
        - class_path: new_transform
...

Callbacks & Logger overrides#

callbacks and logger can currently be provided as a list of different callbacks and loggers. The way to override this is as follows.

For example, if you want to change the patience of EarlyStopping, you can configure the overrides like this

overrides:
...
    callbacks:
      - class_path: ligthning.pytorch.callbacks.EarlyStopping
        init_args:
          patience: 3

Reset overrides#

If you want to reset the configurations to the default values, especially data, callbacks, or logger that are difficult to be reset, you can use the reset keyword. The way to override this is as follows.

Let’s try to reset all transforms which are already set in base data configuration of instance segmentation. To reset them, you can just add the keys in reset in the overrides. reset also supports both types, string and list. If you want to reset single one, string or list can be used. But if you want to reset multiple ones, list should be used.

...
overrides:
  reset:
    - data.train_subset.transforms
  # or
  # reset: data.train_subset.transforms
  data:
    train_subset:
      transforms:
        # previous ones are not used anymore
        - class_path: new_transform_1
        - class_path: new_transform_2
...

...
overrides:
  reset:
    - data.train_subset.transforms
    - data.val_subset.transforms
    - data.test_subset.transforms
    - callbacks
  # reset: data.train_subset.transforms cannot be used because there are multiple resets
  data:
    train_subset:
      transforms:
        # previous ones are not used anymore
        - class_path: new_transform_1
        - class_path: new_transform_2
    val_subset:
      transforms:
        # previous ones are not used anymore
        - class_path: new_transform_1
        - class_path: new_transform_2
    test_subset:
      transforms:
        # previous ones are not used anymore
        - class_path: new_transform_1
        - class_path: new_transform_2
  ...
  callbacks:
    # previous ones are not used anymore
    - class_path: new_callback_1
    - class_path: new_callback_2
...