MODEL_RUN_COMMANDS.md

# Per-model run commands — `foldsrunner_simplified_after_ablation.py`

Base config is already set in the file: `STRATEGIES=[2,3]`, 10 phases (`PHASE_EXECUTION_MODE="auto"`),
100% data, `BATCH_SIZE=48`, `NUM_WORKERS=12`, `STRATEGY_2_MAX_EPOCHS=100`, `STRATEGY_3_MAX_EPOCHS=120`,
`RUN_OPTUNA=False`, `STRATEGY2_CHECKPOINT_MODE="best"` (strategy 3 uses the strategy-2 base trained in
the same run — no checkpoint wiring needed).

**Validated at 128px:** DPT+ViT-tiny and UPerNet+Swin-tiny were dropped (weights locked to 224px) and
replaced with **PVTv2 + UPerNet** (b1 and b2), which run natively at 128px. Re-run
`python validate_models.py` to confirm all 6 pass on your box before launching.

---

## 1. SegFormer-B0  (transformer → SegFormer params, proj 192)

```bash
cd /workspace
sed -i -E \
  -e 's/^SMP_ENCODER_NAME = .*/SMP_ENCODER_NAME = "mit_b0"/' \
  -e 's/^SMP_DECODER_TYPE = .*/SMP_DECODER_TYPE = "Segformer"/' \
  -e 's/^SMP_ENCODER_PROJ_DIM = .*/SMP_ENCODER_PROJ_DIM = 192/' \
  -e 's/^SMP_ENCODER_DEPTH = .*/SMP_ENCODER_DEPTH = 5/' \
  -e 's/^MODEL_NAME = .*/MODEL_NAME = "Segformer_B0"/' \
  -e 's#"2:100": "[^"]*"#"2:100": "params/segformer_b0/best_params_strat2.json"#' \
  -e 's#"3:100": "[^"]*"#"3:100": "params/segformer_b0/best_params_strat3.json"#' \
  foldsrunner_simplified_after_ablation.py
python foldsrunner_simplified_after_ablation.py
```

## 2. U-Net + EfficientNet-B0  (CNN → U-Net params, proj 256)

```bash
cd /workspace
sed -i -E \
  -e 's/^SMP_ENCODER_NAME = .*/SMP_ENCODER_NAME = "efficientnet-b0"/' \
  -e 's/^SMP_DECODER_TYPE = .*/SMP_DECODER_TYPE = "Unet"/' \
  -e 's/^SMP_ENCODER_PROJ_DIM = .*/SMP_ENCODER_PROJ_DIM = 256/' \
  -e 's/^SMP_ENCODER_DEPTH = .*/SMP_ENCODER_DEPTH = 5/' \
  -e 's/^MODEL_NAME = .*/MODEL_NAME = "Unet_EffB0"/' \
  -e 's#"2:100": "[^"]*"#"2:100": "params/unet_effb0/best_params_strat2.json"#' \
  -e 's#"3:100": "[^"]*"#"3:100": "params/unet_effb0/best_params_strat3.json"#' \
  foldsrunner_simplified_after_ablation.py
python foldsrunner_simplified_after_ablation.py
```

## 3. DeepLabV3+ + ResNet34  (CNN → U-Net params, proj 256)

```bash
cd /workspace
sed -i -E \
  -e 's/^SMP_ENCODER_NAME = .*/SMP_ENCODER_NAME = "resnet34"/' \
  -e 's/^SMP_DECODER_TYPE = .*/SMP_DECODER_TYPE = "DeepLabV3Plus"/' \
  -e 's/^SMP_ENCODER_PROJ_DIM = .*/SMP_ENCODER_PROJ_DIM = 256/' \
  -e 's/^SMP_ENCODER_DEPTH = .*/SMP_ENCODER_DEPTH = 5/' \
  -e 's/^MODEL_NAME = .*/MODEL_NAME = "DeepLabV3Plus_ResNet34"/' \
  -e 's#"2:100": "[^"]*"#"2:100": "params/deeplabv3plus_r34/best_params_strat2.json"#' \
  -e 's#"3:100": "[^"]*"#"3:100": "params/deeplabv3plus_r34/best_params_strat3.json"#' \
  foldsrunner_simplified_after_ablation.py
python foldsrunner_simplified_after_ablation.py
```

## 4. LinkNet + MobileNetV3-Large  (CNN → U-Net params, proj 256)

```bash
cd /workspace
sed -i -E \
  -e 's/^SMP_ENCODER_NAME = .*/SMP_ENCODER_NAME = "timm-mobilenetv3_large_100"/' \
  -e 's/^SMP_DECODER_TYPE = .*/SMP_DECODER_TYPE = "Linknet"/' \
  -e 's/^SMP_ENCODER_PROJ_DIM = .*/SMP_ENCODER_PROJ_DIM = 256/' \
  -e 's/^SMP_ENCODER_DEPTH = .*/SMP_ENCODER_DEPTH = 5/' \
  -e 's/^MODEL_NAME = .*/MODEL_NAME = "Linknet_MobileNetV3"/' \
  -e 's#"2:100": "[^"]*"#"2:100": "params/linknet_mbv3/best_params_strat2.json"#' \
  -e 's#"3:100": "[^"]*"#"3:100": "params/linknet_mbv3/best_params_strat3.json"#' \
  foldsrunner_simplified_after_ablation.py
python foldsrunner_simplified_after_ablation.py
```

## 5. UPerNet + PVTv2-b1  (transformer → SegFormer params, proj 192)

```bash
cd /workspace
sed -i -E \
  -e 's/^SMP_ENCODER_NAME = .*/SMP_ENCODER_NAME = "tu-pvt_v2_b1"/' \
  -e 's/^SMP_DECODER_TYPE = .*/SMP_DECODER_TYPE = "UPerNet"/' \
  -e 's/^SMP_ENCODER_PROJ_DIM = .*/SMP_ENCODER_PROJ_DIM = 192/' \
  -e 's/^SMP_ENCODER_DEPTH = .*/SMP_ENCODER_DEPTH = 5/' \
  -e 's/^MODEL_NAME = .*/MODEL_NAME = "UPerNet_PVTv2_b1"/' \
  -e 's#"2:100": "[^"]*"#"2:100": "params/upernet_pvtv2_b1/best_params_strat2.json"#' \
  -e 's#"3:100": "[^"]*"#"3:100": "params/upernet_pvtv2_b1/best_params_strat3.json"#' \
  foldsrunner_simplified_after_ablation.py
python foldsrunner_simplified_after_ablation.py
```

## 6. UPerNet + PVTv2-b2  (transformer → SegFormer params, proj 192)

```bash
cd /workspace
sed -i -E \
  -e 's/^SMP_ENCODER_NAME = .*/SMP_ENCODER_NAME = "tu-pvt_v2_b2"/' \
  -e 's/^SMP_DECODER_TYPE = .*/SMP_DECODER_TYPE = "UPerNet"/' \
  -e 's/^SMP_ENCODER_PROJ_DIM = .*/SMP_ENCODER_PROJ_DIM = 192/' \
  -e 's/^SMP_ENCODER_DEPTH = .*/SMP_ENCODER_DEPTH = 5/' \
  -e 's/^MODEL_NAME = .*/MODEL_NAME = "UPerNet_PVTv2_b2"/' \
  -e 's#"2:100": "[^"]*"#"2:100": "params/upernet_pvtv2_b2/best_params_strat2.json"#' \
  -e 's#"3:100": "[^"]*"#"3:100": "params/upernet_pvtv2_b2/best_params_strat3.json"#' \
  foldsrunner_simplified_after_ablation.py
python foldsrunner_simplified_after_ablation.py
```

---

## After the runs — plots

```bash
cd /workspace
python plot_s2_vs_s3.py
```

One titled figure per model → `runs/_plots/<MODEL_NAME>__s2_vs_s3.png` (6 panels: BIoU band, BIoU
contour, inference time, total training time, time per epoch, time to best checkpoint — S2 vs S3,
mean ± std across the 10 phases, Δ annotated), plus `runs/_plots/summary_s2_vs_s3.csv`.