Skip to content

Clinic 11

Augmentation Choice

Choose image transformations using the deployment slice that must work, as well as the overall score.

Back To Clinics Open Augmentation Topic Open The Full Track

Situation

An image classifier must support a dim scanner. Before comparing recipes, the team requires dim-scanner recall ≥ 0.70, then maximizes overall macro F1 among eligible recipes. The slice and floor come from deployment requirements. Training and validation images are separated by source where needed to avoid duplicate or scanner leakage.

Artifact Packet

These are fixed illustrative scores on the same validation images. Recipes apply to training images; evaluation uses a fixed preprocessing pipeline.

recipe macro F1 class 4 recall dim-scanner recall
no_aug 0.782 0.61 0.44
standard_flips 0.798 0.68 0.52
heavy_color_cutout 0.810 0.66 0.49
rotation_blur 0.803 0.72 0.74

Decision Prompt

  1. Which recipe passes the dim-scanner recall requirement?
  2. Why is the largest overall F1 insufficient here?
  3. Which transforms could change the label for this task?
  4. What validation would you do before release?

Strong Reasoning Looks Like

  • apply the predeclared slice requirement before ranking eligible recipes
  • check that rotations, flips, blur, and color changes preserve labels for this dataset
  • inspect class and scanner slices with their sample counts and uncertainty
  • keep the final test set untouched while choosing transforms

Run The Clinic In Browser

The runner prints this fixed illustrative packet and recalculates any derived columns. It does not run a new training experiment. Edit its PACKET values to explore the decision.

Reference Reveal

Open after writing your note Choose **`rotation_blur`**: its dim-scanner recall is 0.74, and it is the only recipe meeting the 0.70 floor in this packet. `heavy_color_cutout` has higher overall F1 but fails the deployment requirement. This does not establish that rotation and blur universally help scanner data. Inspect transformed examples for label changes and repeat the comparison using enough dim-scanner cases to quantify uncertainty. If the floor were removed or the deployment mix changed, the ranking could change.

What To Do Next

After this clinic:

  1. open Data Augmentation
  2. open Vision Augmentation and Shift Robustness
  3. use Vision and Audio Workflows for the full track where the augmentation decision is wired into a defended workflow