Add exported Track-A ONNX bundles (SegFormer, DINOv2+LoRA): model.onnx + preprocessing + config + parity + model card

levircd_dinov2/config.yaml

@@ -0,0 +1,61 @@
+run_id: levircd_dinov2
+seed: 1337
+data:
+  name: levircd
+  root: ${WORK}/sat-change-detection/data/levircd
+  tile_size: 256
+  num_workers: 8
+  bands:
+  - R
+  - G
+  - B
+model:
+  name: dinov2_cd
+  encoder: facebook/dinov2-base
+  pretrained: true
+  image_size: 448
+  num_feature_layers: 4
+  fusion: diff
+  decoder_dim: 256
+  dropout: 0.1
+  in_channels: 3
+  out_channels: 1
+  lora: true
+  lora_r: 16
+  lora_alpha: 32
+  lora_dropout: 0.05
+  lora_targets:
+  - query
+  - key
+  - value
+  - dense
+  freeze_encoder: true
+  grad_checkpointing: true
+loss:
+  type: bce_dice
+  bce_weight: 1.0
+  dice_weight: 1.0
+train:
+  epochs: 200
+  batch_size: 8
+  lr: 0.0001
+  lr_reference_batch: 8
+  weight_decay: 0.01
+  optimizer: adamw
+  scheduler: cosine
+  amp: true
+  ddp: true
+  ddp_find_unused_parameters: false
+  grad_checkpointing: true
+  ckpt_every_min: 30
+  resume_if_exists: true
+eval:
+  threshold: 0.5
+  metrics:
+  - precision
+  - recall
+  - f1
+  - iou
+logging:
+  backend: tensorboard
+  log_dir: ${WORK}/sat-change-detection/results

levircd_dinov2/metrics_card.md

@@ -0,0 +1,18 @@
+# Model card — levircd_dinov2
+
+- **Architecture:** `dinov2_cd` (encoder `facebook/dinov2-base`, fusion `diff`)
+- **Dataset:** LEVIR-CD (binary building change, 0.5 m RGB aerial)
+- **Checkpoint:** `best.pt` (epoch 198)
+- **Intended use:** portfolio/demo only; trained weights inherit LEVIR-CD research/non-commercial terms.
+
+## Metrics (LEVIR-CD test; threshold selected on val, applied to test)
+
+| F1 | IoU | Precision | Recall | AP | trainable params |
+|---|---|---|---|---|---|
+| 0.912 | 0.839 | 0.924 | 0.901 | 0.946 | 2821377 |
+
+Per-scene F1 mean±std: 0.767 ± 0.312 (n=128). Overall pixel accuracy is intentionally NOT reported (change is a tiny pixel fraction).
+
+## Export parity (PyTorch ↔ ONNXRuntime)
+- opset 17, tol 1e-03, input 448px, dynamic_hw=False
+- max |logit diff| = 8.01e-05 (post-sigmoid 6.52e-07) → **PASS**

levircd_dinov2/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6dceef41ed53de5e3af441e36e22fec4b7d8a747b9f0a32124bef589044d0b1
+size 358192702

levircd_dinov2/parity.json

@@ -0,0 +1,34 @@
+{
+  "passed": true,
+  "tol": 0.001,
+  "opset": 17,
+  "input_size": 448,
+  "dynamic_hw": false,
+  "primary": {
+    "input_shape": [
+      1,
+      2,
+      3,
+      448,
+      448
+    ],
+    "max_abs": 8.0108642578125e-05,
+    "mean_abs": 1.568812149344012e-05,
+    "max_prob_abs": 6.51925802230835e-07
+  },
+  "secondary": {
+    "desc": "dynamic batch=2",
+    "input_shape": [
+      2,
+      2,
+      3,
+      448,
+      448
+    ],
+    "max_abs": 0.00012254714965820312,
+    "mean_abs": 1.7884782209875993e-05,
+    "max_prob_abs": 5.736947059631348e-07
+  },
+  "torch_version": "2.5.1",
+  "onnxruntime_version": "1.27.0"
+}

levircd_dinov2/preprocessing.json

@@ -0,0 +1,44 @@
+{
+  "input_name": "input",
+  "output_name": "logits",
+  "input_shape": [
+    "batch",
+    2,
+    3,
+    448,
+    448
+  ],
+  "input_layout": "(batch, 2 dates, 3 RGB channels, H, W)",
+  "band_order": [
+    "R",
+    "G",
+    "B"
+  ],
+  "value_range": "float32; divide 8-bit RGB by 255 BEFORE normalization",
+  "normalization": {
+    "mean": [
+      0.485,
+      0.456,
+      0.406
+    ],
+    "std": [
+      0.229,
+      0.224,
+      0.225
+    ]
+  },
+  "input_size": 448,
+  "dynamic_hw": false,
+  "resize_to_input": "resize each 256px tile to 448px (bilinear) before inference; resize the 448px output mask back to display resolution",
+  "tiling": {
+    "tile_size": 256,
+    "overlap": 0
+  },
+  "output": {
+    "activation": "sigmoid",
+    "meaning": "per-pixel change probability (channel 0)",
+    "recommended_threshold": 0.5078125,
+    "threshold_source": "val-selected (max-F1)"
+  },
+  "dinov2_note": "encoder runs on a FIXED 448px / 32x32 patch grid with interpolate_pos_encoding baked in; do NOT feed a different size."
+}

levircd_segformer/config.yaml

@@ -0,0 +1,47 @@
+run_id: levircd_segformer
+seed: 1337
+data:
+  name: levircd
+  root: ${WORK}/sat-change-detection/data/levircd
+  tile_size: 256
+  num_workers: 8
+  bands:
+  - R
+  - G
+  - B
+model:
+  name: siamese_segformer
+  encoder: mit_b2
+  pretrained: true
+  fusion: diff
+  decoder_dim: 256
+  dropout: 0.1
+  in_channels: 3
+  out_channels: 1
+loss:
+  type: bce_dice
+  bce_weight: 1.0
+  dice_weight: 1.0
+train:
+  epochs: 200
+  batch_size: 8
+  lr: 4.0e-05
+  lr_reference_batch: 8
+  weight_decay: 0.01
+  optimizer: adamw
+  scheduler: cosine
+  amp: true
+  ddp: true
+  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: ${WORK}/sat-change-detection/results

levircd_segformer/metrics_card.md

@@ -0,0 +1,18 @@
+# Model card — levircd_segformer
+
+- **Architecture:** `siamese_segformer` (encoder `mit_b2`, fusion `diff`)
+- **Dataset:** LEVIR-CD (binary building change, 0.5 m RGB aerial)
+- **Checkpoint:** `best.pt` (epoch 181)
+- **Intended use:** portfolio/demo only; trained weights inherit LEVIR-CD research/non-commercial terms.
+
+## Metrics (LEVIR-CD test; threshold selected on val, applied to test)
+
+| F1 | IoU | Precision | Recall | AP | trainable params |
+|---|---|---|---|---|---|
+| 0.911 | 0.836 | 0.917 | 0.905 | 0.943 | 24722369 |
+
+Per-scene F1 mean±std: 0.761 ± 0.314 (n=128). Overall pixel accuracy is intentionally NOT reported (change is a tiny pixel fraction).
+
+## Export parity (PyTorch ↔ ONNXRuntime)
+- opset 17, tol 1e-03, input 256px, dynamic_hw=True
+- max |logit diff| = 2.29e-05 (post-sigmoid 4.58e-16) → **PASS**

levircd_segformer/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19fc35dd977842bc2f3940bc19df85dbbc7824ce250c53067050821ca4e6bc60
+size 99786543

levircd_segformer/parity.json

@@ -0,0 +1,34 @@
+{
+  "passed": true,
+  "tol": 0.001,
+  "opset": 17,
+  "input_size": 256,
+  "dynamic_hw": true,
+  "primary": {
+    "input_shape": [
+      1,
+      2,
+      3,
+      256,
+      256
+    ],
+    "max_abs": 2.288818359375e-05,
+    "mean_abs": 4.0007580537348986e-06,
+    "max_prob_abs": 4.579669976578771e-16
+  },
+  "secondary": {
+    "desc": "dynamic H/W at 288px",
+    "input_shape": [
+      1,
+      2,
+      3,
+      288,
+      288
+    ],
+    "max_abs": 2.47955322265625e-05,
+    "mean_abs": 4.17278488384909e-06,
+    "max_prob_abs": 4.40619762898109e-16
+  },
+  "torch_version": "2.5.1",
+  "onnxruntime_version": "1.27.0"
+}

levircd_segformer/preprocessing.json

@@ -0,0 +1,43 @@
+{
+  "input_name": "input",
+  "output_name": "logits",
+  "input_shape": [
+    "batch",
+    2,
+    3,
+    256,
+    256
+  ],
+  "input_layout": "(batch, 2 dates, 3 RGB channels, H, W)",
+  "band_order": [
+    "R",
+    "G",
+    "B"
+  ],
+  "value_range": "float32; divide 8-bit RGB by 255 BEFORE normalization",
+  "normalization": {
+    "mean": [
+      0.485,
+      0.456,
+      0.406
+    ],
+    "std": [
+      0.229,
+      0.224,
+      0.225
+    ]
+  },
+  "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": 0.48046875,
+    "threshold_source": "val-selected (max-F1)"
+  }
+}