BaseCompose support

[dtronmans]

Purpose

Tested that training runs with this config:

model:
  name: detection_nested_augs_test
  predefined_model:
    name: DetectionModel
    variant: light
    params:
      loss_params:
        iou_type: "siou"
        iou_loss_weight: 20
        class_loss_weight: 8

loader:
  params:
    dataset_name: fruit_dataset

trainer:
  precision: "16-mixed"
  preprocessing:
    train_image_size: [384, 512]
    keep_aspect_ratio: true
    normalize:
      active: true
      params:
        mean: [0., 0., 0.]
        std: [1, 1, 1]

    augmentations:
      # --- Regular transforms ---
      - name: HorizontalFlip
        params:
          p: 0.5

      - name: RandomRotate90
        params:
          p: 0.3

      # --- OneOf ---
      - name: OneOf
        params:
          transforms:
            - name: GaussianBlur
              params:
                blur_limit: [3, 7]
            - name: MotionBlur
              params:
                blur_limit: [3, 7]
            - name: Defocus
              params:
                radius: [3, 5]
          p: 0.4

      # --- SomeOf ---
      - name: SomeOf
        params:
          n: 2
          transforms:
            - name: ColorJitter
              params:
                brightness: 0.3
                contrast: 0.3
                saturation: 0.3
                hue: 0.05
            - name: CLAHE
              params:
                clip_limit: [1, 4]
            - name: Sharpen
              params:
                alpha: [0.2, 0.5]
                lightness: [0.5, 1.0]
            - name: Posterize
              params:
                num_bits: [4, 6]
          p: 0.5

      # --- Sequential ---
      - name: Sequential
        params:
          transforms:
            - name: GaussNoise
              params:
                std_range: [0.01, 0.05]
            - name: Sharpen
              params:
                alpha: [0.1, 0.3]
                lightness: [0.7, 1.0]
          p: 0.3

      # --- Batch augmentations ---
      - name: Mosaic4
        params:
          out_height: 384
          out_width: 512
          p: 0.5

      - name: MixUp
        params:
          alpha: [0.3, 0.7]
          p: 0.3

  batch_size: 8
  epochs: 50
  accumulate_grad_batches: 8
  n_workers: 4
  validation_interval: 10
  n_log_images: 8

  callbacks:
    - name: EMACallback
      params:
        decay: 0.9999
        use_dynamic_decay: true
        decay_tau: 2000

  training_strategy:
    name: "TripleLRSGDStrategy"
    params:
      warmup_epochs: 2
      warmup_bias_lr: 0.05
      warmup_momentum: 0.5
      lr: 0.0032
      lre: 0.000384
      momentum: 0.843
      weight_decay: 0.00036
      nesterov: true

Specification

Dependencies & Potential Impact

Deployment Plan

Testing & Validation

[klemen1999]

I found some strange behaviour. We should check if this is due to our implementation or is it a bug in the original Albumentations library by creating a MRE outside of luxonis-train/luxonis-ml stack.

• (Sequential or OneOf or SomeOf) + HorizontalFlip produces a mask that is not flipped

augmentations:
  - name: Sequential
    params:
      transforms:
        - name: HorizontalFlip
          params:
            p: 1
      p: 1

• (Sequential or OneOf or SomeOf) + BatchTransform (e.g. Mosaic or MixUp) don't play well together. The below config yields in an error:

- name: OneOf
  params:
    transforms:
      - name: Mosaic4
        params:
          out_height: 384
          out_width: 512
          p: 1

Error:

ValueError: could not broadcast input array from shape (384,4,1) into shape (384,4)

(and a different issue if MixUp is used).
IMO it's fine that this combination is not supported. But in this case we should try to catch it earlier, in some validation step and error out with in a more informative way (e.g. X and Y combination is not supported).

[dtronmans]

I found some strange behaviour. We should check if this is due to our implementation or is it a bug in the original Albumentations library by creating a MRE outside of luxonis-train/luxonis-ml stack.

• (Sequential or OneOf or SomeOf) + HorizontalFlip produces a mask that is not flipped

augmentations:
  - name: Sequential
    params:
      transforms:
        - name: HorizontalFlip
          params:
            p: 1
      p: 1

* (Sequential or OneOf or SomeOf) + BatchTransform (e.g. Mosaic or MixUp) don't play well together. The below config yields in an error:

- name: OneOf
  params:
    transforms:
      - name: Mosaic4
        params:
          out_height: 384
          out_width: 512
          p: 1

Error:

ValueError: could not broadcast input array from shape (384,4,1) into shape (384,4)

(and a different issue if MixUp is used). IMO it's fine that this combination is not supported. But in this case we should try to catch it earlier, in some validation step and error out with in a more informative way (e.g. X and Y combination is not supported).

instance segmentation augmentations were handled in a special way in the albumentations engine that I hadn't noticed, and yes indeed it is true that batch transforms like Mosaic4 cannot be handled by nested augmentations.

I actually didn't know what to add as tests before so I just had "passing augmentations does not fail" as a test, but your comment gave me the idea to include tests that assert that a single transformation inside SomeOf/Sequential/OneOf gives the same result as the augmentation applied alone (both with p=1). Otherwise the default albumentations library works correctly this (and the batch augmentations not working) were just oversights from my part

[klemen1999]

LGTM