Segmentation

[balthazarneveu]

https://wandb.ai/balthazarneveu/geosciences-segmentation

• [x] LR scheduler (plateau)
• [x] Validation accuracy
• [x] Validation Metric IoU, dice coeff
• [x] Augmentations (vertical reverse, horizontal wrap)
• [x] Wrapped convolutions
• [x] Add pos_weight in BCE to give stronger importance to segmented areas
• [x] Flexible Stacked Conv
• [x] Flexible Stacked Conv + Residual connections
• [x] Review input dynamics (remove NaN)
• [x] New metrics logging (Recall Precision F1 score)
• [x] Add pretrained models
• [x] New dataset + update kaggle version
• [x] Could not optimize Dice loss (1-dice coeff)
• [x] Fix DICE loss computation
• [x] Show dice/metrics in interactive pipe
• [x] Control zero labels images.
• [x] UNET+circular repeat = pad in + crop out
• [ ] Re-Split
• [x] Fix metrics normalization
• [ ] Augmentations (add noise, shift dynamic range, blur, sharpen, shear?)
• [ ] Review input dynamics (seems like clipping values to (-0.05 ,0.05) would be a good idea)

---

Lightweight models

• [x] Knowledge distillation
• [x] Report on KD

---

Model performances

• [ ] Refine annotations "super pixels"
• [ ] Refine annotations SAM

---

Bonus

• [ ] Mine difficult examples
• [ ] Attention layer

[balthazarneveu]

Deal with wrapping

All ones out: Loss: 1.217e+00 Accuracy: 9.639% All minus one: Loss 4.096e-01 Accuracy: 90.361%

[balthazarneveu]

• [x] Static quantization
• [ ] Knowledge distillation on Kaggle requires loading models out there!
• [x] Knowledge distillation framework
• [x] temperature scaling
• [ ] QAT

[balthazarneveu]