Add SegFormer-B0 (IGN) checkpoint, config and metrics (#1)

- Add SegFormer-B0 (IGN) checkpoint, config and metrics (d195281d32403ca5c61ba5b69327b029de379c15)


Co-authored-by: OJAS KASANA <masteroshi006@users.noreply.huggingface.co>

segformer-b0/README.md

@@ -0,0 +1,77 @@
+# SegFormer-B0 — BDAPPV PV segmentation (IGN)
+
+Binary photovoltaic segmentation checkpoint for
+[BDAPPV](https://huggingface.co/datasets/gabrielkasmi/bdappv), trained as part
+of [gabrielkasmi/deeppvmapper#20](https://github.com/gabrielkasmi/deeppvmapper/pull/20)
+(issue #11, candidate backbones).
+
+## Files
+
+| File | Description |
+| --- | --- |
+| `segformer_ign_best.pth` | Raw PyTorch `state_dict` of the best checkpoint (epoch 47). |
+| `config.yaml` | Exact training config used for the run. |
+| `metrics.json` | Per-epoch train/val history and the best-epoch summary. |
+| `evaluate.py` | Reference-protocol evaluation (see below). |
+
+## Model
+
+- Architecture: SegFormer-B0 (`segmentation-models-pytorch`, encoder `mit_b0`).
+- Parameters: 3.71 M.
+- Input: 416 × 416, ImageNet normalization (RGB), single-channel logits.
+- Output: PV mask at `sigmoid(logit) >= 0.5`.
+
+## Training
+
+| Setting | Value |
+| --- | --- |
+| Dataset | BDAPPV, provider `ign` |
+| Split | Canonical `annotations.csv` split (1357 val / 1239 test positives) |
+| Epochs | 50 (best epoch 47, selected by `val_pixelF1`) |
+| Batch | 4 × gradient accumulation 8 = effective 32 |
+| Precision | AMP fp16 (CUDA) |
+| Optimizer | AdamW, lr 1e-4, weight decay 0.01 |
+| Loss | 0.5 × BCEWithLogits + 0.5 × Dice |
+| Augmentation | horizontal/vertical flip, 90° rotations |
+| Seed | 42 |
+
+Best validation result (epoch 47): `val_pixelF1` 0.8261, `val_iouPV` 0.7037.
+
+## Results — canonical test split (1239 positives, IGN)
+
+| Protocol | IoU (PV) | Pixel F1 | Instance F1 |
+| --- | --- | --- | --- |
+| Reference (mean of per-batch metrics, batch 8) | **0.7356** | **0.8437** | — |
+| Pixel-aggregated over the split | 0.7445 | 0.8535 | 0.8549 |
+
+Pixel precision 0.8705, recall 0.8372 (aggregated). The two rows differ only in
+how per-image scores are combined: `train/evaluate.py` averages per-batch IoU/F1,
+whereas the training scaffold aggregates TP/FP/FN over the whole split. For
+comparison with the published DeepLabV3-ResNet101 baseline (0.735 IoU /
+0.844 F1), use the first row, noting the resolution caveat below.
+
+The published baseline is evaluated at 400 × 400, while SegFormer's encoder
+requires spatial dimensions divisible by 32, so this model was trained and
+evaluated at 416 × 416. The comparison is therefore not resolution-aligned, as
+already flagged on the GitHub PR.
+
+## Evaluate
+
+```bash
+pip install segmentation-models-pytorch datasets
+python evaluate.py --provider ign --checkpoint segformer_ign_best.pth
+```
+
+`evaluate.py` mirrors `train/evaluate.py`: same `SegmentationDataset` transform,
+same per-batch metric aggregation, same CLI shape. The only changes are the
+model (SegFormer-B0 instead of DeepLabV3-ResNet101) and the default input size
+(416 instead of 400).
+
+The reference `train/evaluate.py` builds `deeplabv3_resnet101` and loads a raw
+DeepLab state dict, so it cannot load this checkpoint directly; SegFormer also
+raises a shape error at the reference's 400 × 400 default because 400 is not
+divisible by 32.
+
+## License
+
+Code MIT; weights CC-BY-4.0, matching the repository.

segformer-b0/config.yaml

@@ -0,0 +1,38 @@
+run_name: segformer-full
+seed: 42
+device: cuda
+data_dir: data/bdappv
+annotations_file: annotations.csv
+data_provider: ign
+split_mode: annotations
+val_fraction: 0.2
+image_size: 416
+subset: 1.0
+model: segformer
+models_config: configs/model.yaml
+pretrained: true
+epochs: 50
+batch_size: 4
+grad_accum_steps: 8
+amp: true
+num_workers: 4
+optimizer: adamw
+lr: 0.0001
+weight_decay: 0.01
+sgd_momentum: 0.9
+augmentations:
+- hflip
+- vflip
+- rot90
+loss: bce_dice
+threshold: 0.5
+selection_metric: val_pixelF1
+num_val_previews: 8
+fps_batches: 50
+dry_run: false
+model_spec:
+  arch: segformer
+  encoder: mit_b0
+  encoder_weights: imagenet
+  classes: 1
+  pretrained_hf_id: null

segformer-b0/evaluate.py

@@ -0,0 +1,140 @@
+"""
+evaluate.py — evaluate the SegFormer-B0 checkpoint on the BDAPPV test split.
+
+Mirrors train/evaluate.py: same SegmentationDataset transform, same per-batch
+metric aggregation, same CLI shape. The model is swapped for
+segmentation-models-pytorch's SegFormer-B0, and the input size defaults to 416
+because SegFormer requires dimensions divisible by 32 (the reference 400
+default raises a shape error).
+
+Usage:
+    pip install segmentation-models-pytorch datasets
+    python evaluate.py --provider ign --checkpoint segformer_ign_best.pth
+"""
+
+import argparse
+
+import torch
+import torch.nn as nn
+from torch.utils.data import DataLoader, Dataset
+from torchvision import transforms as T
+
+from datasets import load_dataset
+import segmentation_models_pytorch as smp
+
+
+IMG_MEAN = [0.485, 0.456, 0.406]
+IMG_STD = [0.229, 0.224, 0.225]
+
+
+class SegmentationDataset(Dataset):
+    """Positives only (has_mask=True) -> (image [3,H,W], mask [1,H,W])."""
+
+    def __init__(self, hf_split, img_size: int = 416, augment: bool = False):
+        self.data = hf_split.filter(lambda x: x["has_mask"])
+        self.augment = augment
+        self.img_size = img_size
+
+        self.img_tf = T.Compose([
+            T.Resize((img_size, img_size)),
+            T.ToTensor(),
+            T.Normalize(mean=IMG_MEAN, std=IMG_STD),
+        ])
+        self.mask_tf = T.Compose([
+            T.Resize((img_size, img_size), interpolation=T.InterpolationMode.NEAREST),
+            T.ToTensor(),
+        ])
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        ex = self.data[idx]
+        image = ex["image"].convert("RGB")
+        mask = ex["mask"].convert("L")
+        image = self.img_tf(image)
+        mask = (self.mask_tf(mask) > 0.5).float()
+        return image, mask
+
+
+def compute_seg_metrics(pred_logits, target, threshold=0.5):
+    pred = (torch.sigmoid(pred_logits) > threshold).float()
+    tp = (pred * target).sum().item()
+    fp = (pred * (1 - target)).sum().item()
+    fn = ((1 - pred) * target).sum().item()
+    iou = tp / max(tp + fp + fn, 1e-6)
+    prec = tp / max(tp + fp, 1e-6)
+    rec = tp / max(tp + fn, 1e-6)
+    f1 = 2 * prec * rec / max(prec + rec, 1e-6)
+    return iou, f1, tp, fp, fn
+
+
+def eval_seg(args):
+    device = (
+        torch.device("mps") if torch.backends.mps.is_available() else
+        torch.device("cuda") if torch.cuda.is_available() else
+        torch.device("cpu")
+    )
+
+    ds = load_dataset("gabrielkasmi/bdappv", args.provider)
+    test_ds = SegmentationDataset(ds["test"], img_size=args.img_size, augment=False)
+    loader = DataLoader(test_ds, batch_size=args.batch_size, shuffle=False,
+                        num_workers=args.num_workers)
+
+    model = smp.Segformer(
+        encoder_name=args.encoder,
+        encoder_weights=None,
+        in_channels=3,
+        classes=1,
+    )
+    state = torch.load(args.checkpoint, map_location="cpu", weights_only=False)
+    if isinstance(state, dict) and "state_dict" in state:
+        state = state["state_dict"]
+    model.load_state_dict(state)
+    model = model.to(device)
+    model.eval()
+
+    total_iou, total_f1 = 0.0, 0.0
+    tp_all = fp_all = fn_all = 0.0
+    with torch.no_grad():
+        for images, masks in loader:
+            images, masks = images.to(device), masks.to(device)
+            logits = model(images)
+            if logits.shape[-2:] != masks.shape[-2:]:
+                logits = nn.functional.interpolate(
+                    logits, size=masks.shape[-2:], mode="bilinear", align_corners=False)
+            iou, f1, tp, fp, fn = compute_seg_metrics(
+                logits, masks, threshold=args.threshold)
+            total_iou += iou
+            total_f1 += f1
+            tp_all += tp
+            fp_all += fp
+            fn_all += fn
+
+    n = len(loader)
+    global_iou = tp_all / max(tp_all + fp_all + fn_all, 1e-6)
+    global_f1 = 2 * tp_all / max(2 * tp_all + fp_all + fn_all, 1e-6)
+
+    # Keep the CLI output portable on Windows consoles using the cp1252 code page.
+    print("\n== Segmentation results ==")
+    print(f"  Provider   : {args.provider}")
+    print(f"  Checkpoint : {args.checkpoint}")
+    print(f"  Test images: {len(test_ds):,}")
+    print(f"  IoU  : {total_iou / n:.4f}")
+    print(f"  F1   : {total_f1 / n:.4f}")
+    print(f"\n  (reference protocol: mean of per-batch IoU/F1, "
+          f"batch_size={args.batch_size})")
+    print(f"  Pixel-aggregated over the split: "
+          f"IoU {global_iou:.4f}, F1 {global_f1:.4f}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--provider", required=True, choices=["google", "ign"])
+    parser.add_argument("--checkpoint", required=True)
+    parser.add_argument("--encoder", default="mit_b0")
+    parser.add_argument("--batch_size", type=int, default=8)
+    parser.add_argument("--img_size", type=int, default=416)
+    parser.add_argument("--num_workers", type=int, default=0)
+    parser.add_argument("--threshold", type=float, default=0.5)
+    eval_seg(parser.parse_args())

segformer-b0/metrics.json

@@ -0,0 +1,834 @@
+{
+  "run_name": "segformer-full",
+  "model": "segformer",
+  "params_M": 3.71,
+  "image_size": 416,
+  "subset": 1.0,
+  "seed": 42,
+  "device": "cuda",
+  "batch_size": 4,
+  "grad_accum_steps": 8,
+  "effective_batch_size": 32,
+  "amp": true,
+  "dry_run": false,
+  "history": [
+    {
+      "epoch": 1,
+      "train_loss": 0.593149,
+      "epoch_time_s": 1100.46,
+      "val_mIoU": 0.6738252127681424,
+      "val_iouPV": 0.3550285065441205,
+      "val_iouBG": 0.9926219189921643,
+      "val_pixelPrecision": 0.4079796484346806,
+      "val_pixelRecall": 0.732293559451075,
+      "val_pixelF1": 0.5240162916566073,
+      "val_instanceF1": 0.20053715308503964,
+      "val_matchedInstances": 560,
+      "val_gtInstances": 1748,
+      "val_predInstances": 3837,
+      "val_images": 1357
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.431503,
+      "epoch_time_s": 399.13,
+      "val_mIoU": 0.7205020839977585,
+      "val_iouPV": 0.44603630855919163,
+      "val_iouBG": 0.9949678594363255,
+      "val_pixelPrecision": 0.533849987127753,
+      "val_pixelRecall": 0.730575164920334,
+      "val_pixelF1": 0.616908864485727,
+      "val_instanceF1": 0.34677241682426146,
+      "val_matchedInstances": 787,
+      "val_gtInstances": 1748,
+      "val_predInstances": 2791,
+      "val_images": 1357
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.318786,
+      "epoch_time_s": 211.39,
+      "val_mIoU": 0.7542999246703384,
+      "val_iouPV": 0.512509915388656,
+      "val_iouBG": 0.9960899339520206,
+      "val_pixelPrecision": 0.6242333459295994,
+      "val_pixelRecall": 0.7411707748039157,
+      "val_pixelF1": 0.6776946189565454,
+      "val_instanceF1": 0.5182308522002539,
+      "val_matchedInstances": 1201,
+      "val_gtInstances": 1748,
+      "val_predInstances": 2887,
+      "val_images": 1357
+    },
+    {
+      "epoch": 4,
+      "train_loss": 0.21663,
+      "epoch_time_s": 29584.91,
+      "val_mIoU": 0.7970613566066507,
+      "val_iouPV": 0.5966902762176065,
+      "val_iouBG": 0.9974324369956948,
+      "val_pixelPrecision": 0.8221626022276559,
+      "val_pixelRecall": 0.6851159588664082,
+      "val_pixelF1": 0.7474089184423466,
+      "val_instanceF1": 0.6716778523309617,
+      "val_matchedInstances": 1251,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1977,
+      "val_images": 1357
+    },
+    {
+      "epoch": 5,
+      "train_loss": 0.168087,
+      "epoch_time_s": 207.28,
+      "val_mIoU": 0.8048857081484486,
+      "val_iouPV": 0.6124470965260008,
+      "val_iouBG": 0.9973243197708964,
+      "val_pixelPrecision": 0.7570294376602414,
+      "val_pixelRecall": 0.7622872147283324,
+      "val_pixelF1": 0.7596492286109866,
+      "val_instanceF1": 0.7219193020522923,
+      "val_matchedInstances": 1324,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1920,
+      "val_images": 1357
+    },
+    {
+      "epoch": 6,
+      "train_loss": 0.143904,
+      "epoch_time_s": 205.05,
+      "val_mIoU": 0.8085076600525211,
+      "val_iouPV": 0.6194444890938778,
+      "val_iouBG": 0.9975708310111643,
+      "val_pixelPrecision": 0.8251618607314468,
+      "val_pixelRecall": 0.7130296552186145,
+      "val_pixelF1": 0.7650086103790732,
+      "val_instanceF1": 0.736150770966986,
+      "val_matchedInstances": 1289,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1754,
+      "val_images": 1357
+    },
+    {
+      "epoch": 7,
+      "train_loss": 0.131723,
+      "epoch_time_s": 201.73,
+      "val_mIoU": 0.8144340639502909,
+      "val_iouPV": 0.6312749696875195,
+      "val_iouBG": 0.9975931582130623,
+      "val_pixelPrecision": 0.80768795598255,
+      "val_pixelRecall": 0.7429454469482699,
+      "val_pixelF1": 0.773965127177111,
+      "val_instanceF1": 0.7424581005379202,
+      "val_matchedInstances": 1329,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1832,
+      "val_images": 1357
+    },
+    {
+      "epoch": 8,
+      "train_loss": 0.123723,
+      "epoch_time_s": 202.41,
+      "val_mIoU": 0.8286441859312119,
+      "val_iouPV": 0.6595244793131402,
+      "val_iouBG": 0.9977638925492835,
+      "val_pixelPrecision": 0.8093845673126608,
+      "val_pixelRecall": 0.7808002368284728,
+      "val_pixelF1": 0.7948354935820053,
+      "val_instanceF1": 0.7475955289641905,
+      "val_matchedInstances": 1438,
+      "val_gtInstances": 1748,
+      "val_predInstances": 2099,
+      "val_images": 1357
+    },
+    {
+      "epoch": 9,
+      "train_loss": 0.114458,
+      "epoch_time_s": 206.72,
+      "val_mIoU": 0.8227074172902802,
+      "val_iouPV": 0.647753180682846,
+      "val_iouBG": 0.9976616538977142,
+      "val_pixelPrecision": 0.7976205019505921,
+      "val_pixelRecall": 0.77515243138558,
+      "val_pixelF1": 0.7862259812654834,
+      "val_instanceF1": 0.7708506511285439,
+      "val_matchedInstances": 1391,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1861,
+      "val_images": 1357
+    },
+    {
+      "epoch": 10,
+      "train_loss": 0.11371,
+      "epoch_time_s": 202.72,
+      "val_mIoU": 0.8226504239802666,
+      "val_iouPV": 0.6476455201578021,
+      "val_iouBG": 0.9976553278027309,
+      "val_pixelPrecision": 0.7956418479865086,
+      "val_pixelRecall": 0.7768754514736042,
+      "val_pixelF1": 0.7861466707909044,
+      "val_instanceF1": 0.7705320600062611,
+      "val_matchedInstances": 1412,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1917,
+      "val_images": 1357
+    },
+    {
+      "epoch": 11,
+      "train_loss": 0.109594,
+      "epoch_time_s": 204.48,
+      "val_mIoU": 0.8317378468911223,
+      "val_iouPV": 0.6657099119785261,
+      "val_iouBG": 0.9977657818037183,
+      "val_pixelPrecision": 0.7967049463333474,
+      "val_pixelRecall": 0.801933637129595,
+      "val_pixelF1": 0.7993107409534442,
+      "val_instanceF1": 0.7875477880724319,
+      "val_matchedInstances": 1442,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1914,
+      "val_images": 1357
+    },
+    {
+      "epoch": 12,
+      "train_loss": 0.106425,
+      "epoch_time_s": 203.79,
+      "val_mIoU": 0.8352092518586216,
+      "val_iouPV": 0.6725981551477012,
+      "val_iouBG": 0.9978203485695419,
+      "val_pixelPrecision": 0.8014800682306861,
+      "val_pixelRecall": 0.8070502744111236,
+      "val_pixelF1": 0.8042555267416355,
+      "val_instanceF1": 0.787533875317411,
+      "val_matchedInstances": 1453,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1942,
+      "val_images": 1357
+    },
+    {
+      "epoch": 13,
+      "train_loss": 0.106637,
+      "epoch_time_s": 202.45,
+      "val_mIoU": 0.8324540768541852,
+      "val_iouPV": 0.6670970587201767,
+      "val_iouBG": 0.9978110949881936,
+      "val_pixelPrecision": 0.8102227552251562,
+      "val_pixelRecall": 0.7906364843913816,
+      "val_pixelF1": 0.8003098022766643,
+      "val_instanceF1": 0.7827745979617668,
+      "val_matchedInstances": 1436,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1921,
+      "val_images": 1357
+    },
+    {
+      "epoch": 14,
+      "train_loss": 0.102226,
+      "epoch_time_s": 10648.35,
+      "val_mIoU": 0.8327790759506015,
+      "val_iouPV": 0.6677675564042089,
+      "val_iouBG": 0.9977905954969942,
+      "val_pixelPrecision": 0.8011727675322264,
+      "val_pixelRecall": 0.8004118287833937,
+      "val_pixelF1": 0.8007921173906866,
+      "val_instanceF1": 0.7503149407722269,
+      "val_matchedInstances": 1489,
+      "val_gtInstances": 1748,
+      "val_predInstances": 2221,
+      "val_images": 1357
+    },
+    {
+      "epoch": 15,
+      "train_loss": 0.100515,
+      "epoch_time_s": 201.48,
+      "val_mIoU": 0.8317349804664309,
+      "val_iouPV": 0.6656130939273999,
+      "val_iouBG": 0.9978568670054621,
+      "val_pixelPrecision": 0.8319039637069516,
+      "val_pixelRecall": 0.7690459248454279,
+      "val_pixelF1": 0.799240947797703,
+      "val_instanceF1": 0.7761442440838795,
+      "val_matchedInstances": 1399,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1857,
+      "val_images": 1357
+    },
+    {
+      "epoch": 16,
+      "train_loss": 0.099064,
+      "epoch_time_s": 430.21,
+      "val_mIoU": 0.8388071740550058,
+      "val_iouPV": 0.6797362286035546,
+      "val_iouBG": 0.997878119506457,
+      "val_pixelPrecision": 0.8070136306740522,
+      "val_pixelRecall": 0.8116735189157686,
+      "val_pixelF1": 0.8093368673350005,
+      "val_instanceF1": 0.7861485593236546,
+      "val_matchedInstances": 1487,
+      "val_gtInstances": 1748,
+      "val_predInstances": 2035,
+      "val_images": 1357
+    },
+    {
+      "epoch": 17,
+      "train_loss": 0.098674,
+      "epoch_time_s": 299.1,
+      "val_mIoU": 0.8345996383798491,
+      "val_iouPV": 0.6712863804989567,
+      "val_iouBG": 0.9979128962607415,
+      "val_pixelPrecision": 0.841602927740337,
+      "val_pixelRecall": 0.7683628842199293,
+      "val_pixelF1": 0.8033169998053193,
+      "val_instanceF1": 0.7892252151960782,
+      "val_matchedInstances": 1421,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1853,
+      "val_images": 1357
+    },
+    {
+      "epoch": 18,
+      "train_loss": 0.097218,
+      "epoch_time_s": 568.08,
+      "val_mIoU": 0.8371065410274661,
+      "val_iouPV": 0.676367258414287,
+      "val_iouBG": 0.997845823640645,
+      "val_pixelPrecision": 0.8025278003134739,
+      "val_pixelRecall": 0.8114090912244073,
+      "val_pixelF1": 0.8069440094577817,
+      "val_instanceF1": 0.7974068071097405,
+      "val_matchedInstances": 1476,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1954,
+      "val_images": 1357
+    },
+    {
+      "epoch": 19,
+      "train_loss": 0.095671,
+      "epoch_time_s": 961.17,
+      "val_mIoU": 0.8421912475412392,
+      "val_iouPV": 0.6864076351702193,
+      "val_iouBG": 0.9979748599122592,
+      "val_pixelPrecision": 0.8297048251126157,
+      "val_pixelRecall": 0.7989694258372222,
+      "val_pixelF1": 0.8140471151281712,
+      "val_instanceF1": 0.8010928961529756,
+      "val_matchedInstances": 1466,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1912,
+      "val_images": 1357
+    },
+    {
+      "epoch": 20,
+      "train_loss": 0.093194,
+      "epoch_time_s": 909.18,
+      "val_mIoU": 0.8279437182954349,
+      "val_iouPV": 0.6582717669202494,
+      "val_iouBG": 0.9976156696706204,
+      "val_pixelPrecision": 0.762787112358299,
+      "val_pixelRecall": 0.8277134097217784,
+      "val_pixelF1": 0.7939250731414127,
+      "val_instanceF1": 0.7935991320641823,
+      "val_matchedInstances": 1463,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1939,
+      "val_images": 1357
+    },
+    {
+      "epoch": 21,
+      "train_loss": 0.092722,
+      "epoch_time_s": 7165.49,
+      "val_mIoU": 0.8386570221453022,
+      "val_iouPV": 0.6793784569561795,
+      "val_iouBG": 0.9979355873344251,
+      "val_pixelPrecision": 0.830774704173666,
+      "val_pixelRecall": 0.7884956222954338,
+      "val_pixelF1": 0.8090832106868054,
+      "val_instanceF1": 0.8048368953654433,
+      "val_matchedInstances": 1431,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1808,
+      "val_images": 1357
+    },
+    {
+      "epoch": 22,
+      "train_loss": 0.090862,
+      "epoch_time_s": 199.33,
+      "val_mIoU": 0.8406898687158819,
+      "val_iouPV": 0.683430215857697,
+      "val_iouBG": 0.9979495215740668,
+      "val_pixelPrecision": 0.8265175007006968,
+      "val_pixelRecall": 0.7978862745067192,
+      "val_pixelF1": 0.8119495651436833,
+      "val_instanceF1": 0.8096693525754468,
+      "val_matchedInstances": 1457,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1851,
+      "val_images": 1357
+    },
+    {
+      "epoch": 23,
+      "train_loss": 0.090107,
+      "epoch_time_s": 250.94,
+      "val_mIoU": 0.8312692891724652,
+      "val_iouPV": 0.6648312684321094,
+      "val_iouBG": 0.9977073099128211,
+      "val_pixelPrecision": 0.7787956082466407,
+      "val_pixelRecall": 0.8196001822468939,
+      "val_pixelF1": 0.7986770563941035,
+      "val_instanceF1": 0.7834344499925235,
+      "val_matchedInstances": 1485,
+      "val_gtInstances": 1748,
+      "val_predInstances": 2043,
+      "val_images": 1357
+    },
+    {
+      "epoch": 24,
+      "train_loss": 0.089943,
+      "epoch_time_s": 206.21,
+      "val_mIoU": 0.8432043442539974,
+      "val_iouPV": 0.6884068893577852,
+      "val_iouBG": 0.9980017991502096,
+      "val_pixelPrecision": 0.8361811821623424,
+      "val_pixelRecall": 0.7957245974029967,
+      "val_pixelF1": 0.8154514100799868,
+      "val_instanceF1": 0.8061196105478131,
+      "val_matchedInstances": 1449,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1847,
+      "val_images": 1357
+    },
+    {
+      "epoch": 25,
+      "train_loss": 0.089125,
+      "epoch_time_s": 205.48,
+      "val_mIoU": 0.8429536380725373,
+      "val_iouPV": 0.6879148176763237,
+      "val_iouBG": 0.9979924584687511,
+      "val_pixelPrecision": 0.8333925388198025,
+      "val_pixelRecall": 0.7976048864386527,
+      "val_pixelF1": 0.8151060829282191,
+      "val_instanceF1": 0.7983935742758128,
+      "val_matchedInstances": 1491,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1987,
+      "val_images": 1357
+    },
+    {
+      "epoch": 26,
+      "train_loss": 0.088248,
+      "epoch_time_s": 206.59,
+      "val_mIoU": 0.8397173523323961,
+      "val_iouPV": 0.6815594716367182,
+      "val_iouBG": 0.997875233028074,
+      "val_pixelPrecision": 0.8018616459598616,
+      "val_pixelRecall": 0.8195878474274717,
+      "val_pixelF1": 0.8106278524580917,
+      "val_instanceF1": 0.7880741337419266,
+      "val_matchedInstances": 1467,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1975,
+      "val_images": 1357
+    },
+    {
+      "epoch": 27,
+      "train_loss": 0.085908,
+      "epoch_time_s": 202.62,
+      "val_mIoU": 0.8458093475433397,
+      "val_iouPV": 0.6935751398077196,
+      "val_iouBG": 0.9980435552789598,
+      "val_pixelPrecision": 0.8410818486603752,
+      "val_pixelRecall": 0.7981738299844969,
+      "val_pixelF1": 0.8190662740673729,
+      "val_instanceF1": 0.8034886889941814,
+      "val_matchedInstances": 1474,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1921,
+      "val_images": 1357
+    },
+    {
+      "epoch": 28,
+      "train_loss": 0.08649,
+      "epoch_time_s": 208.36,
+      "val_mIoU": 0.83951497443859,
+      "val_iouPV": 0.6811227632033853,
+      "val_iouBG": 0.9979071856737948,
+      "val_pixelPrecision": 0.8149882794405313,
+      "val_pixelRecall": 0.8057026953892595,
+      "val_pixelF1": 0.8103188870103732,
+      "val_instanceF1": 0.7984994640729234,
+      "val_matchedInstances": 1490,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1984,
+      "val_images": 1357
+    },
+    {
+      "epoch": 29,
+      "train_loss": 0.085521,
+      "epoch_time_s": 213.64,
+      "val_mIoU": 0.8436663557889087,
+      "val_iouPV": 0.6893834746640788,
+      "val_iouBG": 0.9979492369137386,
+      "val_pixelPrecision": 0.8120398933368533,
+      "val_pixelRecall": 0.8202739717578259,
+      "val_pixelF1": 0.8161361644681145,
+      "val_instanceF1": 0.7978552278606473,
+      "val_matchedInstances": 1488,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1982,
+      "val_images": 1357
+    },
+    {
+      "epoch": 30,
+      "train_loss": 0.085817,
+      "epoch_time_s": 215.2,
+      "val_mIoU": 0.8436078357260084,
+      "val_iouPV": 0.6892536195868806,
+      "val_iouBG": 0.9979620518651363,
+      "val_pixelPrecision": 0.8174155456493378,
+      "val_pixelRecall": 0.8146793602236909,
+      "val_pixelF1": 0.8160451593472893,
+      "val_instanceF1": 0.8094713656164794,
+      "val_matchedInstances": 1470,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1884,
+      "val_images": 1357
+    },
+    {
+      "epoch": 31,
+      "train_loss": 0.085633,
+      "epoch_time_s": 218.84,
+      "val_mIoU": 0.8450143205205473,
+      "val_iouPV": 0.6919885139213121,
+      "val_iouBG": 0.9980401271197824,
+      "val_pixelPrecision": 0.8437995367417472,
+      "val_pixelRecall": 0.793653889592512,
+      "val_pixelF1": 0.8179588788313651,
+      "val_instanceF1": 0.8063263040841753,
+      "val_matchedInstances": 1453,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1856,
+      "val_images": 1357
+    },
+    {
+      "epoch": 32,
+      "train_loss": 0.085422,
+      "epoch_time_s": 2930.77,
+      "val_mIoU": 0.8473283011040855,
+      "val_iouPV": 0.6966063745952851,
+      "val_iouBG": 0.998050227612886,
+      "val_pixelPrecision": 0.83599731298952,
+      "val_pixelRecall": 0.8068714195295033,
+      "val_pixelF1": 0.8211761844422589,
+      "val_instanceF1": 0.8046040010741408,
+      "val_matchedInstances": 1468,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1901,
+      "val_images": 1357
+    },
+    {
+      "epoch": 33,
+      "train_loss": 0.084356,
+      "epoch_time_s": 205.22,
+      "val_mIoU": 0.8444034326205179,
+      "val_iouPV": 0.6908059603542308,
+      "val_iouBG": 0.9980009048868049,
+      "val_pixelPrecision": 0.829622908789513,
+      "val_pixelRecall": 0.8050119455016221,
+      "val_pixelF1": 0.8171321565598267,
+      "val_instanceF1": 0.8143413006999906,
+      "val_matchedInstances": 1465,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1850,
+      "val_images": 1357
+    },
+    {
+      "epoch": 34,
+      "train_loss": 0.08246,
+      "epoch_time_s": 403.49,
+      "val_mIoU": 0.8377969423270138,
+      "val_iouPV": 0.677743752308217,
+      "val_iouBG": 0.9978501323458107,
+      "val_pixelPrecision": 0.8010792007547736,
+      "val_pixelRecall": 0.8148844265965832,
+      "val_pixelF1": 0.8079228444460441,
+      "val_instanceF1": 0.8034330010851763,
+      "val_matchedInstances": 1451,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1864,
+      "val_images": 1357
+    },
+    {
+      "epoch": 35,
+      "train_loss": 0.083514,
+      "epoch_time_s": 204.74,
+      "val_mIoU": 0.8455978690792829,
+      "val_iouPV": 0.6931767969891778,
+      "val_iouBG": 0.9980189411693879,
+      "val_pixelPrecision": 0.8312820561709257,
+      "val_pixelRecall": 0.8066648113041831,
+      "val_pixelF1": 0.8187884433826296,
+      "val_instanceF1": 0.8155872667174647,
+      "val_matchedInstances": 1486,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1896,
+      "val_images": 1357
+    },
+    {
+      "epoch": 36,
+      "train_loss": 0.083226,
+      "epoch_time_s": 202.85,
+      "val_mIoU": 0.8458458897700466,
+      "val_iouPV": 0.6936996724220734,
+      "val_iouBG": 0.9979921071180198,
+      "val_pixelPrecision": 0.818760368642236,
+      "val_pixelRecall": 0.8195462174119221,
+      "val_pixelF1": 0.8191531045525313,
+      "val_instanceF1": 0.8129496402652754,
+      "val_matchedInstances": 1469,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1866,
+      "val_images": 1357
+    },
+    {
+      "epoch": 37,
+      "train_loss": 0.081685,
+      "epoch_time_s": 204.12,
+      "val_mIoU": 0.8431581663938573,
+      "val_iouPV": 0.6882680335471802,
+      "val_iouBG": 0.9980482992405344,
+      "val_pixelPrecision": 0.8579518845801389,
+      "val_pixelRecall": 0.776786024032794,
+      "val_pixelF1": 0.81535398393,
+      "val_instanceF1": 0.7989864864639925,
+      "val_matchedInstances": 1419,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1804,
+      "val_images": 1357
+    },
+    {
+      "epoch": 38,
+      "train_loss": 0.082452,
+      "epoch_time_s": 212.66,
+      "val_mIoU": 0.8407923324134924,
+      "val_iouPV": 0.6836536222822299,
+      "val_iouBG": 0.9979310425447548,
+      "val_pixelPrecision": 0.8184385535527137,
+      "val_pixelRecall": 0.8058730700825273,
+      "val_pixelF1": 0.8121072092673339,
+      "val_instanceF1": 0.798493408641418,
+      "val_matchedInstances": 1484,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1969,
+      "val_images": 1357
+    },
+    {
+      "epoch": 39,
+      "train_loss": 0.082606,
+      "epoch_time_s": 212.19,
+      "val_mIoU": 0.8418933727363175,
+      "val_iouPV": 0.6858057559702427,
+      "val_iouBG": 0.9979809895023922,
+      "val_pixelPrecision": 0.8338576056441077,
+      "val_pixelRecall": 0.794348494111219,
+      "val_pixelF1": 0.8136236972041141,
+      "val_instanceF1": 0.8106312292134377,
+      "val_matchedInstances": 1464,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1864,
+      "val_images": 1357
+    },
+    {
+      "epoch": 40,
+      "train_loss": 0.080357,
+      "epoch_time_s": 213.65,
+      "val_mIoU": 0.8457843333599484,
+      "val_iouPV": 0.6934946168495064,
+      "val_iouBG": 0.9980740498703903,
+      "val_pixelPrecision": 0.8555180043073596,
+      "val_pixelRecall": 0.7854905519137253,
+      "val_pixelF1": 0.8190101225590541,
+      "val_instanceF1": 0.8232965789984932,
+      "val_matchedInstances": 1456,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1789,
+      "val_images": 1357
+    },
+    {
+      "epoch": 41,
+      "train_loss": 0.080324,
+      "epoch_time_s": 1217.16,
+      "val_mIoU": 0.846830429220599,
+      "val_iouPV": 0.6956627769048629,
+      "val_iouBG": 0.9979980815363352,
+      "val_pixelPrecision": 0.816400061054658,
+      "val_pixelRecall": 0.8246821278487979,
+      "val_pixelF1": 0.8205201958548316,
+      "val_instanceF1": 0.8088314485511893,
+      "val_matchedInstances": 1502,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1966,
+      "val_images": 1357
+    },
+    {
+      "epoch": 42,
+      "train_loss": 0.084833,
+      "epoch_time_s": 300.53,
+      "val_mIoU": 0.8439916695827654,
+      "val_iouPV": 0.6899710404812384,
+      "val_iouBG": 0.9980122986842925,
+      "val_pixelPrecision": 0.8367066928671356,
+      "val_pixelRecall": 0.797337375042436,
+      "val_pixelF1": 0.816547767924782,
+      "val_instanceF1": 0.7922573609380519,
+      "val_matchedInstances": 1453,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1920,
+      "val_images": 1357
+    },
+    {
+      "epoch": 43,
+      "train_loss": 0.08315,
+      "epoch_time_s": 750.64,
+      "val_mIoU": 0.8442857281510686,
+      "val_iouPV": 0.6905020310919432,
+      "val_iouBG": 0.998069425210194,
+      "val_pixelPrecision": 0.8618549131725465,
+      "val_pixelRecall": 0.7764375653841197,
+      "val_pixelF1": 0.8169194930170279,
+      "val_instanceF1": 0.8113260442721804,
+      "val_matchedInstances": 1447,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1819,
+      "val_images": 1357
+    },
+    {
+      "epoch": 44,
+      "train_loss": 0.080176,
+      "epoch_time_s": 15830.14,
+      "val_mIoU": 0.8490489042698492,
+      "val_iouPV": 0.6999753275927932,
+      "val_iouBG": 0.9981224809469051,
+      "val_pixelPrecision": 0.8605027486583657,
+      "val_pixelRecall": 0.7895710643637979,
+      "val_pixelF1": 0.8235123371866525,
+      "val_instanceF1": 0.8175787727800559,
+      "val_matchedInstances": 1479,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1870,
+      "val_images": 1357
+    },
+    {
+      "epoch": 45,
+      "train_loss": 0.084558,
+      "epoch_time_s": 1108.71,
+      "val_mIoU": 0.8410883528058986,
+      "val_iouPV": 0.6842262036169512,
+      "val_iouBG": 0.9979505019948459,
+      "val_pixelPrecision": 0.8249571144691431,
+      "val_pixelRecall": 0.8004349565698101,
+      "val_pixelF1": 0.8125110536192166,
+      "val_instanceF1": 0.8005420053983593,
+      "val_matchedInstances": 1477,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1942,
+      "val_images": 1357
+    },
+    {
+      "epoch": 46,
+      "train_loss": 0.082218,
+      "epoch_time_s": 994.9,
+      "val_mIoU": 0.8359112374427674,
+      "val_iouPV": 0.6739746814446353,
+      "val_iouBG": 0.9978477934408995,
+      "val_pixelPrecision": 0.8085916973199583,
+      "val_pixelRecall": 0.801913593048034,
+      "val_pixelF1": 0.805238799505614,
+      "val_instanceF1": 0.7659247889289863,
+      "val_matchedInstances": 1497,
+      "val_gtInstances": 1748,
+      "val_predInstances": 2161,
+      "val_images": 1357
+    },
+    {
+      "epoch": 47,
+      "train_loss": 0.081081,
+      "epoch_time_s": 208.93,
+      "val_mIoU": 0.8509209091354384,
+      "val_iouPV": 0.7037282852574396,
+      "val_iouBG": 0.9981135330134371,
+      "val_pixelPrecision": 0.8454646777443242,
+      "val_pixelRecall": 0.8076099668424014,
+      "val_pixelF1": 0.8261038938507775,
+      "val_instanceF1": 0.8132825258352401,
+      "val_matchedInstances": 1494,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1926,
+      "val_images": 1357
+    },
+    {
+      "epoch": 48,
+      "train_loss": 0.078675,
+      "epoch_time_s": 205.6,
+      "val_mIoU": 0.8467959333510855,
+      "val_iouPV": 0.6955549791808134,
+      "val_iouBG": 0.9980368875213577,
+      "val_pixelPrecision": 0.832899628252722,
+      "val_pixelRecall": 0.808357765269866,
+      "val_pixelF1": 0.8204452084672149,
+      "val_instanceF1": 0.8164609053273948,
+      "val_matchedInstances": 1488,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1897,
+      "val_images": 1357
+    },
+    {
+      "epoch": 49,
+      "train_loss": 0.077831,
+      "epoch_time_s": 1448.22,
+      "val_mIoU": 0.850042356023367,
+      "val_iouPV": 0.7020280299596217,
+      "val_iouBG": 0.9980566820871124,
+      "val_pixelPrecision": 0.8247402025979876,
+      "val_pixelRecall": 0.8251223267169241,
+      "val_pixelF1": 0.8249312204056666,
+      "val_instanceF1": 0.8114532684815923,
+      "val_matchedInstances": 1502,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1954,
+      "val_images": 1357
+    },
+    {
+      "epoch": 50,
+      "train_loss": 0.079206,
+      "epoch_time_s": 217.0,
+      "val_mIoU": 0.8439486548989446,
+      "val_iouPV": 0.6898645199875201,
+      "val_iouBG": 0.9980327898103691,
+      "val_pixelPrecision": 0.8462019432763722,
+      "val_pixelRecall": 0.7887623627654367,
+      "val_pixelF1": 0.8164731690947802,
+      "val_instanceF1": 0.8010047446050516,
+      "val_matchedInstances": 1435,
+      "val_gtInstances": 1748,
+      "val_predInstances": 1835,
+      "val_images": 1357
+    }
+  ],
+  "best_epoch": 47,
+  "best": {
+    "epoch": 47,
+    "train_loss": 0.081081,
+    "epoch_time_s": 208.93,
+    "val_mIoU": 0.8509209091354384,
+    "val_iouPV": 0.7037282852574396,
+    "val_iouBG": 0.9981135330134371,
+    "val_pixelPrecision": 0.8454646777443242,
+    "val_pixelRecall": 0.8076099668424014,
+    "val_pixelF1": 0.8261038938507775,
+    "val_instanceF1": 0.8132825258352401,
+    "val_matchedInstances": 1494,
+    "val_gtInstances": 1748,
+    "val_predInstances": 1926,
+    "val_images": 1357
+  },
+  "avg_epoch_time_s": 1670.29
+}

segformer-b0/segformer_ign_best.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91ad5ede3d421a479b4cbb2888c65349c39ada31be6dd1a7d6abe4f9b603cb05
+size 14914662