M5: add OSCD Sentinel-2 Track-B 4-band bundle (test F1 0.453; portfolio artifact)

oscd_s2_baseline/config.yaml

@@ -0,0 +1,45 @@
+run_id: oscd_s2_baseline
+seed: 1337
+data:
+  name: oscd
+  root: /tmp/w/sat-change-detection/data/oscd
+  tile_size: 256
+  num_workers: 4
+  bands:
+  - R
+  - G
+  - B
+  - NIR
+model:
+  name: fc_siam_diff
+  in_channels: 4
+  fusion: diff
+  base_channels: 16
+  out_channels: 1
+loss:
+  type: bce_dice
+  bce_weight: 1.0
+  dice_weight: 1.0
+train:
+  epochs: 100
+  batch_size: 16
+  lr: 0.001
+  lr_reference_batch: 16
+  weight_decay: 0.0001
+  optimizer: adamw
+  scheduler: cosine
+  amp: true
+  ddp: false
+  grad_checkpointing: false
+  ckpt_every_min: 30
+  resume_if_exists: true
+eval:
+  threshold: 0.5
+  metrics:
+  - precision
+  - recall
+  - f1
+  - iou
+logging:
+  backend: tensorboard
+  log_dir: /tmp/w/sat-change-detection/results

oscd_s2_baseline/metrics_card.md

@@ -0,0 +1,22 @@
+# Model card — oscd_s2_baseline
+
+- **Architecture:** `fc_siam_diff` (encoder `None`, fusion `diff`)
+- **Dataset:** OSCD (Onera Satellite CD; Sentinel-2 ~10 m, RGB+NIR, binary change)
+- **Checkpoint:** `oscd_s2_baseline_best.pt` (epoch 19)
+- **Intended use:** portfolio/demo only; trained weights inherit the dataset's research/non-commercial terms.
+
+## Metrics (test split; threshold selected on val, applied to test)
+
+| F1 | IoU | Precision | Recall | AP | trainable params |
+|---|---|---|---|---|---|
+| 0.453 | 0.293 | 0.492 | 0.420 | 0.413 | 834705 |
+
+Per-scene F1 mean±std: 0.378 ± 0.175 (n=10). Overall pixel accuracy is intentionally NOT reported (change is a tiny pixel fraction).
+
+## Honest framing — read this
+
+Sentinel-2 is a **coarse 10 m/px** domain. This model is **directionally correct on large real-world change, not a high-accuracy detector** (test F1 ~0.45). OSCD is tiny (24 scenes) and genuinely hard, so scores are modest by design — not dressed up. The high-resolution **aerial LEVIR-CD track is the high-accuracy showcase**; this Sentinel-2 track exists to run on any real-world location, plainly caveated.
+
+## Export parity (PyTorch ↔ ONNXRuntime)
+- opset 17, tol 1e-03, input 256px, dynamic_hw=True
+- max |logit diff| = 1.43e-06 (post-sigmoid 2.68e-07) → **PASS**

oscd_s2_baseline/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ccb6229a5c83de2d2ccd2b4512e93eab3c8f3a5a163ce92b5871a629b78717a
+size 3352304

oscd_s2_baseline/parity.json

@@ -0,0 +1,34 @@
+{
+  "passed": true,
+  "tol": 0.001,
+  "opset": 17,
+  "input_size": 256,
+  "dynamic_hw": true,
+  "primary": {
+    "input_shape": [
+      1,
+      2,
+      4,
+      256,
+      256
+    ],
+    "max_abs": 1.430511474609375e-06,
+    "mean_abs": 1.1048496162402444e-07,
+    "max_prob_abs": 2.682209014892578e-07
+  },
+  "secondary": {
+    "desc": "dynamic H/W at 288px",
+    "input_shape": [
+      1,
+      2,
+      4,
+      288,
+      288
+    ],
+    "max_abs": 1.3113021850585938e-06,
+    "mean_abs": 1.07262344783976e-07,
+    "max_prob_abs": 2.682209014892578e-07
+  },
+  "torch_version": "2.5.1",
+  "onnxruntime_version": "1.27.0"
+}

oscd_s2_baseline/preprocessing.json

@@ -0,0 +1,46 @@
+{
+  "input_name": "input",
+  "output_name": "logits",
+  "input_shape": [
+    "batch",
+    2,
+    4,
+    256,
+    256
+  ],
+  "input_layout": "(batch, 2 dates, 4 bands [R,G,B,NIR], H, W)",
+  "band_order": [
+    "R",
+    "G",
+    "B",
+    "NIR"
+  ],
+  "value_range": "float32 reflectance; divide uint16 Sentinel-2 L2A (bands B04,B03,B02,B08) by 10000 BEFORE normalization",
+  "normalization": {
+    "mean": [
+      0.13559,
+      0.13234,
+      0.13953,
+      0.2011
+    ],
+    "std": [
+      0.08096,
+      0.05684,
+      0.04359,
+      0.08868
+    ]
+  },
+  "input_size": 256,
+  "dynamic_hw": true,
+  "resize_to_input": "feed 256px tiles directly (fully-convolutional; dynamic H/W also allowed)",
+  "tiling": {
+    "tile_size": 256,
+    "overlap": 0
+  },
+  "output": {
+    "activation": "sigmoid",
+    "meaning": "per-pixel change probability (channel 0)",
+    "recommended_threshold": 0.46875,
+    "threshold_source": "val-selected (max-F1)"
+  }
+}