feat: train fixed-size classifiers and export class probabilities

README.md

@@ -801,36 +801,64 @@ ultralytics 自己还会在 `runs/detect/film/` 下写：
 配置支持 `_base_` 继承与命令行 `--set a.b.c=value` 覆盖，见 [utils/config.py](utils/config.py)。
 新增骨架 / 数据集只需用注册表装饰器注册，再在 yaml 里按 `name` 引用。
 
-## 分类模型复用 YOLO OpenVINO 接口
+## 分类模型导出 OpenVINO
 
-生产端只使用类别时，可以将分类器导出成与既有 YOLO IR 相同的输入输出接口：
+分类器直接输出按训练类别顺序排列的 Softmax 概率，不适配 YOLO 检测输出：
 
 ```bash
 uv pip install openvino
 uv run python scripts/export_classifier_openvino.py \
     --checkpoint runs/classify/best.pth \
-    --template runs/detect/models/film_128_20261013_0902_openvino_model/film_128_20261013_0902.xml \
+    --fp32-weights --out runs/classify/models/resnet18_128_probabilities
-    --out runs/classify/models/resnet18_128_yolo_adapter
 ```
 
-`--template` 请指定实际部署的 YOLO `.xml`，其同目录需要有 `metadata.yaml`。
 脚本读取分类检查点的配置和同目录 `eval_report.json` 中的类别顺序；也可用
 `--classes classes.txt` 或 `--classes eval_report.json` 明确指定分类器的类别顺序。
-两者类别集合不一致、训练预处理不受支持或输入尺寸不匹配时会拒绝导出。
+指定顺序不能与检查点记录矛盾。输入尺寸从检查点读取，不依赖 YOLO 模型。
 
-产出的 XML/BIN 使用模板文件名，输入为 RGB float32 BCHW、值域 `[0,1]`。
+产出的 XML/BIN 按骨架命名，如 `resnet18.xml/.bin`、`mobilenet_v3_small.xml/.bin`。
-逐图标准化按训练配置嵌入模型，分类概率按 YOLO 类别顺序排列。
+输入为 RGB float32 BCHW、值域 `[0,1]`，输入名为 `images`。
-输出保持 `[1,4+nc,anchors]`：仅第一个候选携带全图伪框及 Softmax 概率，其他候选为零。
+逐图标准化仅在训练配置使用时嵌入模型。脚本不进行 resize，图片必须已切为目标尺寸。
-伪框是像素 `cx,cy,w,h`，不表示缺陷位置。脚本不进行 resize，图片必须已切为目标尺寸。
+输出名为 `probabilities`，形状 `[1,C]`：九类模型是 `[1,9]`，索引直接对应 `classes.txt`。
+每类概率在 `[0,1]` 内，概率和为 1；normal 也返回真实概率，不清零、不做阈值过滤。
 
-部署时成对替换 XML/BIN；使用元数据的程序也需同步 `metadata.yaml`。
+部署端直接对 C 个概率取 argmax；normal 是否表示正常、是否应用阈值由业务层决定。
-仓库的 `deploy/cpp` 解码器可直接读取适配模型，生产代码若有额外硬编码则需核对。
+现有 `deploy/cpp` 是 YOLO 框解码器，不能用于此分类输出；需要按 `[1,C]` 读取，
-Softmax 分数不同于 YOLO 置信度，需在实际数据上确认过滤阈值。
+不要跳过前四个通道或遍历候选框。模型附带 `metadata.yaml`、`classes.txt` 和 `export.json`。
-默认 top1 为 `normal` 时，所有候选类别分数清零，C++ 在正数置信度阈值下返回空列表，
-与 YOLO 的背景行为一致。类别编号保持不变，`normal` 仍占其原通道。
-用 `--background-class` 指定其他背景类别，或 `--keep-background-box` 恢复普通分类输出。
-原图被切成多片时，每片独立分类，仅前景片返回伪框。
 输出目录必须是新目录，原有模型不会被覆盖。
+默认压缩权重为 FP16；需要优先核对转换前后判别一致性时，加 `--fp32-weights`
+保留 FP32 权重，减少分类边界处的舍入差异。输入输出接口仍是 float32。
+
+### 在固定 128×128 YOLO 裁片上监督训练分类器
+
+```bash
+uv run python classify.py --config configs/classify/film_tiles128_resnet18.yaml \
+    --work-dir runs/classify_tiles128_resnet18
+uv run python classify.py --config configs/classify/film_tiles128_mobilenet_v3_small.yaml \
+    --work-dir runs/classify_tiles128_mobilenet_v3_small
+```
+
+数据从 `data/film_yolo_128/data.yaml` 读取，保留现有 train/val 划分和类别编号。
+监督标签来自 YOLO 标注文件，空标签表示 normal；多类别裁片取面积最大的框对应类别。
+缺失标签文件会报错，不会自动当背景。图片必须已经是 128×128 PNG，禁止 resize。
+这两份配置使用 RGB / 255，与检测部署输入一致；不进行逐图标准化。
+最佳权重按验证 macro-F1 保存，可用上述导出脚本生成 OpenVINO 分类概率模型。
+分类准确率、转换前后一致率和与 YOLO 的一致率是不同指标，应分别报告。
+
+导出后，在同一固定裁片验证集上检查这三个指标（示例）：
+
+```bash
+uv run python scripts/compare_openvino_classes.py \
+    --teacher runs/detect/models/film_128_20261013_0902_openvino_model/film_128_20261013_0902.xml \
+    --student runs/classify/models/resnet18_tiles128/resnet18.xml \
+    --checkpoint runs/classify_tiles128_resnet18/best.pth \
+    --conf 0.25 --report runs/classify_tiles128_resnet18/comparison.json
+```
+
+`--student` 指向实际导出的 XML。这里 `--teacher` 只用于比较，不参与训练。
+YOLO 有有效框时取最高分候选的类别，无有效框记为 normal。比较工具在业务层对分类概率
+使用同一置信度阈值，低于阈值时记为 normal；模型本身仍返回全部真实概率。
+报告同时记录各类支持数、总体/前景一致率和逐图结果。
 
 ## 开发
 

classify.py

@@ -16,6 +16,7 @@ import json
 from pathlib import Path
 
 from datasets import FolderClassificationDataset, classification_collate
+from datasets.tile_classification import YoloTileClassificationDataset
 from engine import ClassificationTrainer
 from models import build_model
 from utils import load_config, seed_everything, setup_logging
@@ -40,6 +41,13 @@ def build_loaders(cfg, split_cfg, classes=None):
 
     datasets = {}
     for split, transforms_key in (("train", "train_transforms"), ("val", "val_transforms")):
+        if split_cfg.get("format", "folder") == "yolo_tiles":
+            datasets[split] = YoloTileClassificationDataset(
+                yaml_path=split_cfg["yaml"], split=split,
+                image_size=split_cfg["image_size"], transforms=split_cfg[transforms_key],
+                classes=classes or split_cfg.get("classes"),
+            )
+            continue
         datasets[split] = FolderClassificationDataset(
             root=split_cfg["root"],
             split=split,

configs/classify/film_tiles128_mobilenet_v3_small.yaml

@@ -0,0 +1,8 @@
+_base_: film_tiles128_resnet18.yaml
+work_dir: runs/classify_tiles128_mobilenet_v3_small
+model:
+  backbone:
+    arch: mobilenet_v3_small
+train:
+  swanlab:
+    tags: [classify, mobilenet_v3_small, tiles128]

configs/classify/film_tiles128_resnet18.yaml

@@ -0,0 +1,20 @@
+# Train on human YOLO labels, preserving the existing fixed-tile train/val split.
+_base_: film.yaml
+work_dir: runs/classify_tiles128_resnet18
+data:
+  root: data/film_yolo_128
+  format: yolo_tiles
+  yaml: data/film_yolo_128/data.yaml
+  image_size: [128, 128]
+  classes: [black, blackdot, grease, mosquito, normal, obstacle, ripple, uneven, wrinkle]
+  train_transforms:
+    - {name: RandomHorizontalFlip, p: 0.5}
+    - {name: RandomVerticalFlip, p: 0.5}
+    - {name: ToTensor}
+  val_transforms:
+    - {name: ToTensor}
+train:
+  # Strong inverse-frequency weights plus uniform smoothing bias scores toward rare classes.
+  label_smoothing: 0.0
+  swanlab:
+    tags: [classify, resnet18, tiles128]

datasets/tile_classification.py

@@ -0,0 +1,81 @@
+"""Supervised classification from fixed-size YOLO tiles and human box labels."""
+from __future__ import annotations
+
+import math
+from pathlib import Path
+
+import yaml
+from PIL import Image
+
+from datasets.classification import FolderClassificationDataset
+from datasets.transforms import build_transforms
+
+
+def tile_label(text: str, num_classes: int, background: int) -> int:
+    boxes = []
+    for line in text.splitlines():
+        if not line.strip():
+            continue
+        parts = line.split()
+        if len(parts) != 5:
+            raise ValueError(f"Malformed YOLO label: {line}")
+        cls, cx, cy, w, h = map(float, parts)
+        if (not all(math.isfinite(v) for v in (cls, cx, cy, w, h))
+                or not cls.is_integer() or not 0 <= cls < num_classes
+                or not 0 <= cx <= 1 or not 0 <= cy <= 1 or not 0 < w <= 1 or not 0 < h <= 1):
+            raise ValueError(f"Invalid YOLO label: {line}")
+        boxes.append((int(cls), w * h))
+    return max(boxes, key=lambda b: b[1])[0] if boxes else background
+
+
+class YoloTileClassificationDataset(FolderClassificationDataset):
+    """Keep the existing YOLO split. Empty labels are normal; largest box wins.
+
+    No dataset transform may resample the image. The size is checked on every read.
+    """
+
+    def __init__(self, yaml_path, split, image_size=(128, 128), transforms=(), classes=None):
+        path = Path(yaml_path).resolve()
+        cfg = yaml.safe_load(path.read_text())
+        names = cfg["names"]
+        self.classes = ([names.get(i, names.get(str(i))) for i in range(len(names))]
+                        if isinstance(names, dict) else list(names))
+        if not self.classes or any(not isinstance(n, str) for n in self.classes):
+            raise ValueError("Invalid YOLO names mapping")
+        if classes is not None and list(classes) != self.classes:
+            raise ValueError("Classifier class order must match YOLO names")
+        background = self.classes.index("normal")
+        self.class_to_idx = dict(zip(self.classes, range(len(self.classes)), strict=True))
+        self.root = Path(cfg.get("path", path.parent))
+        if not self.root.is_absolute():
+            self.root = path.parent / self.root
+        folder = self.root / cfg[split]
+        if not folder.is_dir():
+            raise FileNotFoundError(folder)
+        relative = folder.relative_to(self.root)
+        if relative.parts[0] != "images":
+            raise ValueError("YOLO split must be beneath images/")
+        labels = self.root / "labels" / Path(*relative.parts[1:])
+        allowed = {"ToTensor", "RandomHorizontalFlip", "RandomVerticalFlip", "PerImageStandardize"}
+        if any(t["name"] not in allowed for t in transforms):
+            raise ValueError("Fixed tiles forbid resampling transforms")
+        self.transforms = build_transforms(transforms)
+        self.image_size = tuple(image_size)
+        self.samples = []
+        for image in sorted(folder.rglob("*.png")):
+            label = labels / image.relative_to(folder).with_suffix(".txt")
+            if not label.is_file():
+                raise FileNotFoundError(f"Missing label (not an implicit background): {label}")
+            self.samples.append((image, tile_label(label.read_text(), len(self.classes), background)))
+        if not self.samples:
+            raise ValueError(f"No PNG tiles in {folder}")
+
+    def __getitem__(self, index):
+        path, label = self.samples[index]
+        with Image.open(path) as raw:
+            if (raw.height, raw.width) != self.image_size:
+                raise ValueError(f"{path}: size {raw.size}, expected {self.image_size}; resize is forbidden")
+            image = raw.convert("RGB")
+        sample = self.transforms({"image": image})
+        return {"image": sample["image"], "label": label,
+                "meta": {"path": str(path), "index": index, "class": self.classes[label]}}

scripts/compare_openvino_classes.py

@@ -0,0 +1,134 @@
+#!/usr/bin/env python
+"""Compare supervised classification accuracy, export parity and YOLO agreement.
+
+Uses fixed-size human-labelled validation tiles, never training data. No resize.
+Empty detector output is normal; the strongest valid candidate supplies top1.
+"""
+from __future__ import annotations
+
+import argparse
+import copy
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+import yaml
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+from datasets.tile_classification import YoloTileClassificationDataset  # noqa: E402
+from models import build_model  # noqa: E402
+
+
+def output_class(output: np.ndarray, conf: float, background: int) -> int:
+    scores = output[0, 4:, :]
+    index = int(scores.max(axis=0).argmax())
+    return int(scores[:, index].argmax()) if scores[:, index].max() >= conf else background
+
+
+def probability_class(output: np.ndarray, conf: float, background: int) -> int:
+    if output.ndim != 2 or output.shape[0] != 1:
+        raise ValueError("Expected classifier probabilities [1,C]")
+    return int(output[0].argmax()) if output[0].max() >= conf else background
+
+
+def matrix_report(matrix, names, background):
+    counts = matrix.sum(1)
+    foreground = int(matrix.sum() - counts[background])
+    return {"agreement": float(np.trace(matrix) / matrix.sum()),
+            "foreground_agreement": float((np.trace(matrix) - matrix[background, background])
+                                          / foreground) if foreground else None,
+            "per_reference_class": {name: {"support": int(counts[i]),
+                                             "agreement": float(matrix[i, i] / counts[i])
+                                             if counts[i] else None}
+                                    for i, name in enumerate(names)},
+            "confusion_reference_rows_prediction_columns": matrix.tolist()}
+
+
+def main():
+    import openvino as ov
+
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--teacher", required=True, help="Reference YOLO .xml; used only for comparison")
+    parser.add_argument("--student", required=True, help="Classifier probability .xml, output [1,C]")
+    parser.add_argument("--checkpoint", required=True)
+    parser.add_argument("--report", required=True)
+    parser.add_argument("--conf", type=float, default=0.25)
+    parser.add_argument("--device", default="cuda:0")
+    args = parser.parse_args()
+    if not 0 < args.conf < 1:
+        raise ValueError("Positive confidence threshold required")
+    state = torch.load(args.checkpoint, map_location="cpu", weights_only=False)
+    cfg = state["config"]
+    names = cfg["data"]["classes"]
+    background = names.index("normal")
+    for xml in (args.teacher, args.student):
+        meta = yaml.safe_load((Path(xml).parent / "metadata.yaml").read_text())
+        if [meta["names"][i] for i in range(len(meta["names"]))] != names:
+            raise ValueError(f"Class order mismatch: {xml}")
+    if cfg["data"]["val_transforms"] != [{"name": "ToTensor"}]:
+        raise ValueError("This comparison expects the new raw RGB / 255 tile classifiers")
+    ds = YoloTileClassificationDataset(cfg["data"]["yaml"], "val", cfg["data"]["image_size"],
+                                      transforms=[{"name": "ToTensor"}], classes=names)
+    model_cfg = copy.deepcopy(cfg["model"])
+    model_cfg["backbone"]["pretrained"] = False
+    model_cfg["init_from"] = None
+    model_cfg["num_classes"] = len(names)
+    model = build_model(model_cfg).to(args.device).eval()
+    model.load_state_dict(state["model"])
+    core = ov.Core()
+    settings = {"PERFORMANCE_HINT": "LATENCY", "INFERENCE_NUM_THREADS": 4}
+    teacher = core.compile_model(args.teacher, "CPU", settings)
+    student = core.compile_model(args.student, "CPU", settings)
+    for compiled in (teacher, student):
+        if list(compiled.input().shape) != [1, 3, *cfg["data"]["image_size"]]:
+            raise ValueError("Unexpected model interface")
+    if (list(teacher.output().shape)[:2] != [1, 4 + len(names)]
+            or list(student.output().shape) != [1, len(names)]):
+        raise ValueError("Expected raw YOLO output and plain classifier probabilities")
+    matrices = {key: np.zeros((len(names), len(names)), dtype=int)
+                for key in ["pt_accuracy", "ov_accuracy", "ov_deployment_accuracy",
+                            "yolo_deployment_accuracy",
+                            "export_top1_parity", "export_deployment_parity", "yolo_student_agreement"]}
+    records = []
+    with torch.inference_mode():
+        for start in range(0, len(ds), 64):
+            batch = [ds[i] for i in range(start, min(start + 64, len(ds)))]
+            images = torch.stack([b["image"] for b in batch])
+            probabilities = model(images.to(args.device)).softmax(1).cpu().numpy()
+            for i, item in enumerate(batch):
+                x = images[i:i+1].numpy()
+                t = output_class(teacher(x)[teacher.output()], args.conf, background)
+                raw = student(x)[student.output()]
+                if (not np.isfinite(raw).all() or (raw < 0).any() or (raw > 1).any()
+                        or not np.allclose(raw.sum(1), 1, atol=1e-5)):
+                    raise ValueError("Invalid classifier probability output")
+                s = probability_class(raw, args.conf, background)
+                raw_s = int(raw[0].argmax())
+                p = int(probabilities[i].argmax())
+                deployed_p = p if probabilities[i].max() >= args.conf else background
+                y = item["label"]
+                for key, row, col in [("pt_accuracy", y, p), ("ov_accuracy", y, raw_s),
+                                      ("ov_deployment_accuracy", y, s),
+                                      ("yolo_deployment_accuracy", y, t),
+                                      ("export_top1_parity", p, raw_s),
+                                      ("export_deployment_parity", deployed_p, s),
+                                      ("yolo_student_agreement", t, s)]:
+                    matrices[key][row, col] += 1
+                records.append({"path": item["meta"]["path"], "truth": y, "pytorch": p,
+                                "pytorch_deployed": deployed_p, "openvino": s,
+                                "openvino_top1": raw_s, "yolo": t,
+                                "student_max_score": float(raw.max())})
+    report = {"classes": names, "images": len(ds), "conf": args.conf,
+              "teacher": args.teacher, "student": args.student, "checkpoint": args.checkpoint,
+              "metrics": {k: matrix_report(m, names, background) for k, m in matrices.items()}}
+    out = Path(args.report)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(json.dumps(report, indent=2))
+    out.with_suffix(".predictions.json").write_text(json.dumps(records, indent=2))
+    print(json.dumps(report, indent=2), flush=True)
+
+
+if __name__ == "__main__":
+    main()

scripts/export_classifier_openvino.py

@@ -1,10 +1,8 @@
 #!/usr/bin/env python
-"""Export a classifier with the input/output contract of a YOLO OpenVINO IR.
+"""Export RGB / 255 BCHW -> Softmax probabilities [1, C], in training class order.
 
-The sole active candidate is a full-image pseudo box, not a localization result.
+Training normalization is embedded. No resize, pseudo boxes or background gating.
-Input must already be RGB float32 BCHW in [0, 1], exactly as in deploy/cpp.
 """
-
 from __future__ import annotations
 
 import argparse
@@ -19,7 +17,6 @@ import yaml
 from torch import nn
 
 sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
-
 from models import build_model  # noqa: E402
 
 
@@ -36,63 +33,48 @@ def ordered_names(names) -> list[str]:
     return names
 
 
-def class_order(source: list[str], target: list[str]) -> list[int]:
+def input_size(config: dict) -> list[int]:
-    source, target = ordered_names(source), ordered_names(target)
+    size = config["data"]["image_size"]
-    if set(source) != set(target):
+    size = [size, size] if isinstance(size, int) else list(size)
-        raise ValueError(f"Class sets differ: classifier={source}, YOLO={target}")
+    if len(size) != 2 or any(not isinstance(v, int) or v < 1 for v in size):
-    return [source.index(name) for name in target]
+        raise ValueError(f"Invalid training image_size: {size}")
+    return size
 
 
-def preprocessing(config: dict, size: int) -> float | None:
+def preprocessing(config: dict, size: int | list[int]) -> float | None:
-    """Reject preprocessing that cannot be reproduced without resampling the input."""
+    shape = [size, size] if isinstance(size, int) else list(size)
     transforms = config["data"]["val_transforms"]
     sequence = [t["name"] for t in transforms]
     if sequence not in (["Resize", "ToTensor", "PerImageStandardize"],
                         ["ToTensor", "PerImageStandardize"], ["Resize", "ToTensor"], ["ToTensor"]):
         raise ValueError(f"Unsupported validation preprocessing: {sequence}")
+    eps = None
     for transform in transforms:
-        if transform["name"] == "Resize":
+        if transform["name"] == "Resize" and list(transform["size"]) != shape:
-            if transform["size"] != [size, size]:
+            raise ValueError("Training size differs from export input")
-                raise ValueError("Training size differs from template; retrain at the target size")
         if transform["name"] == "PerImageStandardize":
-            return float(transform.get("eps", 1e-5))
+            eps = float(transform.get("eps", 1e-5))
-    return None
+            if eps <= 0:
+                raise ValueError("Standardization eps must be positive")
+    return eps
 
 
-class YoloClassifierAdapter(nn.Module):
+class ClassifierProbabilities(nn.Module):
-    def __init__(self, classifier: nn.Module, size: int, anchors: int,
+    def __init__(self, classifier: nn.Module, eps: float | None = None):
-                 order: list[int], eps: float | None = 1e-5,
-                 background_index: int | None = None):
         super().__init__()
-        if size <= 0 or anchors <= 0:
+        self.classifier, self.eps = classifier, eps
-            raise ValueError("size and anchors must be positive")
-        self.classifier = classifier
-        self.eps = eps
-        if background_index is not None and not 0 <= background_index < len(order):
-            raise ValueError("background_index must be an output class index")
-        self.background_index = background_index
-        self.register_buffer("order", torch.tensor(order, dtype=torch.long))
-        self.register_buffer("box", torch.tensor([size / 2, size / 2, size, size])
-                             .reshape(1, 4, 1))
-        # Only candidate zero is active; do not duplicate predictions across anchors.
-        self.register_buffer("inactive", torch.zeros(1, 4 + len(order), anchors - 1))
 
     def normalize(self, x: torch.Tensor) -> torch.Tensor:
         if self.eps is None:
             return x
         dims = (1, 2, 3)
         mean = x.mean(dim=dims, keepdim=True)
-        # Match datasets.transforms.PerImageStandardize: torch.std correction=1.
+        # Match training torch.std(correction=1), separately for each image.
         variance = ((x - mean) ** 2).sum(dim=dims, keepdim=True) / (x[0].numel() - 1)
         return (x - mean) / (torch.sqrt(variance) + self.eps)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
-        scores = self.classifier(self.normalize(x)).softmax(dim=1)[:, self.order]
+        return self.classifier(self.normalize(x)).softmax(dim=1)
-        if self.background_index is not None:
-            foreground = (scores.argmax(dim=1) != self.background_index).unsqueeze(1)
-            scores = scores * foreground.to(scores.dtype)
-        active = torch.cat((self.box.expand(x.shape[0], -1, -1), scores.unsqueeze(2)), dim=1)
-        return torch.cat((active, self.inactive.expand(x.shape[0], -1, -1)), dim=2)
 
 
 def main() -> int:
@@ -100,97 +82,71 @@ def main() -> int:
 
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--checkpoint", default="runs/classify/best.pth")
-    parser.add_argument("--template", required=True, help="Production YOLO .xml (metadata.yaml beside it)")
+    parser.add_argument("--classes", help="Training class order: classes.txt or eval_report.json")
-    parser.add_argument("--classes", help="Classifier class order: classes.txt or eval_report.json")
+    parser.add_argument("--fp32-weights", action="store_true", help="Keep full-precision weights")
-    parser.add_argument("--background-class", default="normal", help="Top1 class to emit no boxes for")
-    parser.add_argument("--keep-background-box", action="store_true",
-                        help="Opt out of background suppression (legacy classification behavior)")
     parser.add_argument("--out", required=True, help="New directory; existing directories are rejected")
     args = parser.parse_args()
-    checkpoint, template, out = Path(args.checkpoint), Path(args.template), Path(args.out)
+    checkpoint, out = Path(args.checkpoint), Path(args.out)
     if out.exists():
         raise FileExistsError(f"Refusing to overwrite {out}")
     state = torch.load(checkpoint, map_location="cpu", weights_only=False)
     config = state["config"]
+    names = config["data"].get("classes")
     if args.classes:
         path = Path(args.classes)
-        source = (json.loads(path.read_text())["classes"] if path.suffix == ".json"
+        supplied = (json.loads(path.read_text())["classes"] if path.suffix == ".json"
                     else path.read_text().splitlines())
-    else:
+        if names and ordered_names(supplied) != ordered_names(names):
-        source = config["data"].get("classes")
+            raise ValueError("--classes contradicts checkpoint class order")
-        if not source:
+        names = supplied
+    if not names:
         report = checkpoint.parent / "eval_report.json"
         if not report.exists():
             raise ValueError("No recorded class order; provide --classes (do not guess folder order)")
-            source = json.loads(report.read_text())["classes"]
+        names = json.loads(report.read_text())["classes"]
-    source = ordered_names(source)
+    names = ordered_names(names)
-    metadata = yaml.safe_load((template.parent / "metadata.yaml").read_text())
+    size = input_size(config)
-    if metadata.get("end2end") or metadata.get("args", {}).get("nms"):
-        raise ValueError("Expected raw [1, 4+nc, anchors] output, not end-to-end/NMS output")
-    target = ordered_names(metadata["names"])
-    order = class_order(source, target)
-    if not args.keep_background_box and args.background_class not in target:
-        raise ValueError(f"Background class {args.background_class!r} is not in {target}")
-    background_index = None if args.keep_background_box else target.index(args.background_class)
-    core = ov.Core()
-    reference = core.read_model(str(template))
-    if len(reference.inputs) != 1 or len(reference.outputs) != 1:
-        raise ValueError("Expected one input and one output")
-    input_shape, output_shape = list(reference.input().shape), list(reference.output().shape)
-    if (len(input_shape) != 4 or input_shape[:2] != [1, 3]
-            or input_shape[2] != input_shape[3] or len(output_shape) != 3
-            or output_shape[:2] != [1, 4 + len(target)] or output_shape[2] < 1):
-        raise ValueError(f"Unsupported template shapes: {input_shape}, {output_shape}")
-    if (reference.input().element_type != ov.Type.f32
-            or reference.output().element_type != ov.Type.f32):
-        raise ValueError("Only float32 input/output are supported")
-    size = input_shape[2]
     eps = preprocessing(config, size)
     model_config = copy.deepcopy(config["model"])
     model_config["backbone"]["pretrained"] = False
     model_config["init_from"] = None
-    model_config["num_classes"] = len(source)
+    model_config["num_classes"] = len(names)
     model = build_model(model_config).eval()
     model.load_state_dict(state["model"], strict=True)
-    adapter = YoloClassifierAdapter(model, size, output_shape[2], order, eps,
+    wrapper = ClassifierProbabilities(model, eps).eval()
-                                    background_index=background_index).eval()
+    shape = [1, model_config["backbone"].get("in_channels", 3), *size]
     torch.manual_seed(42)
-    example = torch.rand(input_shape)
+    example = torch.rand(shape)
-    converted = ov.convert_model(adapter, example_input=example, input=input_shape)
+    converted = ov.convert_model(wrapper, example_input=example, input=shape)
-    converted.input().get_tensor().set_names(reference.input().get_names())
+    converted.input().get_tensor().set_names({"images"})
-    converted.output().get_tensor().set_names(reference.output().get_names())
+    converted.output().get_tensor().set_names({"probabilities"})
-    # Preserve any interface metadata used by OpenVINO consumers, then mark pseudo boxes.
+    converted.set_rt_info("classify", ["model_info", "task"])
-    for key, value in reference.get_rt_info().items():
+    converted.set_rt_info(json.dumps(dict(enumerate(names))), ["model_info", "names"])
-        if key != "conversion_parameters":
-            converted.set_rt_info(value, [key])
-    converted.set_rt_info("classifier_pseudo_box", ["adapter_kind"])
     out.mkdir(parents=True)
-    xml = out / template.name
+    arch = model_config["backbone"].get("arch", model_config["backbone"]["name"])
-    ov.save_model(converted, str(xml), compress_to_fp16=True)
+    xml = out / f"{arch}.xml"
-    compiled = core.compile_model(str(xml), "CPU", {"PERFORMANCE_HINT": "LATENCY"})
+    ov.save_model(converted, str(xml), compress_to_fp16=not args.fp32_weights)
+    compiled = ov.Core().compile_model(str(xml), "CPU", {"PERFORMANCE_HINT": "LATENCY"})
     errors = []
     with torch.inference_mode():
         for sample in (example, torch.zeros_like(example), torch.ones_like(example)):
-            expected = adapter(sample).numpy()
+            expected = wrapper(sample).numpy()
             actual = compiled(sample.numpy())[compiled.output()]
+            assert actual.shape == (1, len(names))
+            assert np.isfinite(actual).all() and (actual >= 0).all() and (actual <= 1).all()
+            np.testing.assert_allclose(actual.sum(axis=1), 1, atol=1e-5)
             np.testing.assert_allclose(actual, expected, atol=0.02, rtol=0.02)
-            assert np.argmax(actual[0, 4:, 0]) == np.argmax(expected[0, 4:, 0])
+            assert actual.argmax(axis=1).tolist() == expected.argmax(axis=1).tolist()
-            assert np.count_nonzero(actual[:, :, 1:]) == 0
             errors.append(float(np.max(np.abs(expected - actual))))
-    metadata.update({"names": dict(enumerate(target)), "adapter_kind": "classifier_pseudo_box",
+    metadata = {"task": "classify", "names": dict(enumerate(names)), "imgsz": size,
-                     "description": "Classifier with one full-image pseudo box; no localization",
+                "channels": shape[1], "layout": "BCHW", "input_range": [0, 1],
-                     "background_class": None if background_index is None else args.background_class,
+                "output": "softmax_probabilities", "input_shape": shape,
-                     "end2end": False, "args": {"nms": False}, "imgsz": [size, size]})
+                "output_shape": [1, len(names)], "standardize_eps": eps,
+                "weight_storage": "fp32" if args.fp32_weights else "fp16"}
     (out / "metadata.yaml").write_text(yaml.safe_dump(metadata, allow_unicode=True))
-    (out / "classes.txt").write_text("\n".join(target) + "\n")
+    (out / "classes.txt").write_text("\n".join(names) + "\n")
-    summary = {"checkpoint": str(checkpoint.resolve()), "template": str(template.resolve()),
+    summary = {"checkpoint": str(checkpoint.resolve()), **metadata, "max_abs_errors": errors}
-               "classifier_classes": source, "output_classes": target, "class_order": order,
+    (out / "export.json").write_text(json.dumps(summary, ensure_ascii=False, indent=2))
-               "input_shape": input_shape, "output_shape": output_shape,
-               "standardize_eps": eps, "max_abs_errors": errors,
-               "background_class": None if background_index is None else args.background_class,
-               "background_index": background_index,
-               "pseudo_box": [size / 2, size / 2, size, size]}
-    (out / "adapter.json").write_text(json.dumps(summary, ensure_ascii=False, indent=2))
     print(json.dumps({"model": str(xml), **summary}, ensure_ascii=False, indent=2))
     return 0
 

tests/test_classifier_adapter.py

@@ -1,83 +0,0 @@
-import pytest
-import torch
-from torch import nn
-
-from scripts.export_classifier_openvino import YoloClassifierAdapter, class_order, preprocessing
-
-
-class FixedClassifier(nn.Module):
-    def forward(self, x):
-        return torch.tensor([1.0, 3.0, -1.0]).expand(x.shape[0], -1)
-
-
-def test_yolo_contract_and_class_reordering():
-    adapter = YoloClassifierAdapter(FixedClassifier(), 128, 336, [2, 0, 1])
-    result = adapter(torch.rand(1, 3, 128, 128))
-    assert result.shape == (1, 7, 336)
-    torch.testing.assert_close(result[0, :4, 0], torch.tensor([64., 64., 128., 128.]))
-    expected = torch.tensor([1., 3., -1.]).softmax(0)[[2, 0, 1]]
-    torch.testing.assert_close(result[0, 4:, 0], expected)
-    assert result[0, 4:, 0].argmax().item() == 2
-    assert torch.count_nonzero(result[:, :, 1:]) == 0
-
-
-def test_normalization_matches_training_for_each_image():
-    adapter = YoloClassifierAdapter(FixedClassifier(), 128, 1, [0, 1, 2])
-    images = torch.stack((torch.rand(3, 128, 128), torch.ones(3, 128, 128)))
-    expected = torch.stack([(x - x.mean()) / (x.std() + 1e-5) for x in images])
-    torch.testing.assert_close(adapter.normalize(images), expected)
-    assert torch.isfinite(adapter(images)).all()
-
-
-def test_class_sets_must_match():
-    assert class_order(["normal", "mosquito"], ["mosquito", "normal"]) == [1, 0]
-    with pytest.raises(ValueError, match="Class sets differ"):
-        class_order(["normal", "mosquito"], ["normal", "blackdot"])
-    with pytest.raises(ValueError, match="Duplicate"):
-        class_order(["normal", "normal"], ["normal"])
-
-
-def test_background_suppression_uses_reordered_output_index():
-    # Source class 1 wins, but it is output class 2 after reordering.
-    adapter = YoloClassifierAdapter(FixedClassifier(), 128, 336, [2, 0, 1], background_index=2)
-    result = adapter(torch.rand(1, 3, 128, 128))
-    assert result.shape == (1, 7, 336)
-    assert torch.count_nonzero(result[:, 4:, :]) == 0
-    # The existing C++ decoder returns zero candidates at any positive threshold.
-    assert not (result[0, 4:, :].max(dim=0).values >= 0.25).any()
-
-
-def test_foreground_class_indices_and_scores_are_preserved():
-    foreground = YoloClassifierAdapter(FixedClassifier(), 128, 336, [0, 1, 2],
-                                      background_index=0)
-    legacy = YoloClassifierAdapter(FixedClassifier(), 128, 336, [0, 1, 2])
-    x = torch.rand(1, 3, 128, 128)
-    torch.testing.assert_close(foreground(x), legacy(x))
-    assert foreground(x)[0, 4:, 0].argmax() == 1
-
-
-def test_background_gate_is_per_image_in_batch():
-    class BatchClassifier(nn.Module):
-        def forward(self, x):
-            return x[:, :, 0, 0]
-
-    x = torch.zeros(2, 3, 4, 4)
-    x[0, 0, 0, 0] = 3  # normal
-    x[1, 1, 0, 0] = 3  # blackdot
-    adapter = YoloClassifierAdapter(BatchClassifier(), 4, 10, [0, 1, 2],
-                                    eps=None, background_index=0)
-    result = adapter(x)
-    assert not torch.count_nonzero(result[0, 4:, :])
-    assert result[1, 4:, 0].argmax() == 1
-    assert result[1, 5, 0] > 0.9
-
-
-def test_refuses_resampling_and_unknown_preprocessing():
-    config = {"data": {"val_transforms": [
-        {"name": "Resize", "size": [160, 160]}, {"name": "ToTensor"},
-        {"name": "PerImageStandardize"}]}}
-    with pytest.raises(ValueError, match="Training size differs"):
-        preprocessing(config, 128)
-    config["data"]["val_transforms"] = [{"name": "Normalize"}]
-    with pytest.raises(ValueError, match="Unsupported"):
-        preprocessing(config, 128)

tests/test_classifier_export.py

@@ -0,0 +1,52 @@
+import pytest
+import torch
+from torch import nn
+
+from scripts.export_classifier_openvino import ClassifierProbabilities, ordered_names, preprocessing
+
+
+class FixedClassifier(nn.Module):
+    def forward(self, x):
+        return torch.tensor([1.0, 3.0, -1.0]).expand(x.shape[0], -1)
+
+
+def test_probabilities_keep_training_class_order_and_sum_to_one():
+    result = ClassifierProbabilities(FixedClassifier())(torch.rand(2, 3, 128, 128))
+    assert result.shape == (2, 3)
+    torch.testing.assert_close(result[0], torch.tensor([1., 3., -1.]).softmax(0))
+    torch.testing.assert_close(result.sum(1), torch.ones(2))
+    assert result.argmax(1).tolist() == [1, 1]
+
+
+def test_normal_probability_is_not_suppressed():
+    # Suppose class 1 is normal: its probability is returned like any other class.
+    result = ClassifierProbabilities(FixedClassifier())(torch.zeros(1, 3, 128, 128))
+    assert result[0, 1] > .8
+    assert (result > 0).all()
+
+
+def test_standardization_matches_training_per_image():
+    exported = ClassifierProbabilities(FixedClassifier(), eps=1e-5)
+    images = torch.stack((torch.rand(3, 128, 128), torch.ones(3, 128, 128)))
+    expected = torch.stack([(x-x.mean())/(x.std()+1e-5) for x in images])
+    torch.testing.assert_close(exported.normalize(images), expected)
+    assert torch.isfinite(exported(images)).all()
+
+
+def test_classes_validate_index_order_and_duplicates():
+    assert ordered_names({1: "blackdot", 0: "normal"}) == ["normal", "blackdot"]
+    with pytest.raises(ValueError, match="Duplicate"):
+        ordered_names(["normal", "normal"])
+    with pytest.raises(ValueError, match="contiguous"):
+        ordered_names({1: "normal"})
+
+
+def test_refuses_resampling_and_unknown_preprocessing():
+    config = {"data": {"val_transforms": [
+        {"name": "Resize", "size": [160, 160]}, {"name": "ToTensor"},
+        {"name": "PerImageStandardize"}]}}
+    with pytest.raises(ValueError, match="Training size differs"):
+        preprocessing(config, 128)
+    config["data"]["val_transforms"] = [{"name": "Normalize"}]
+    with pytest.raises(ValueError, match="Unsupported"):
+        preprocessing(config, 128)

tests/test_compare_openvino_classes.py

@@ -0,0 +1,26 @@
+import numpy as np
+
+from scripts.compare_openvino_classes import matrix_report, output_class, probability_class
+
+
+def test_detector_output_channel_offset_and_background():
+    output = np.zeros((1, 13, 336), dtype=np.float32)
+    output[0, :4, 0] = [64, 64, 128, 128]
+    assert output_class(output, 0.25, 4) == 4
+    output[0, 5, 0] = 0.6
+    assert output_class(output, 0.25, 4) == 1  # blackdot, not black
+    output[0, 4, 50] = 0.9
+    assert output_class(output, 0.25, 4) == 0
+
+
+def test_report_excludes_background_from_foreground_agreement():
+    matrix = np.array([[1, 1], [0, 8]])
+    report = matrix_report(matrix, ["blackdot", "normal"], 1)
+    assert report["agreement"] == 0.9
+    assert report["foreground_agreement"] == 0.5
+
+
+def test_probability_output_has_no_coordinate_offset():
+    probabilities = np.array([[0.1, 0.8, 0.1]])
+    assert probability_class(probabilities, .25, 2) == 1
+    assert probability_class(np.array([[.3, .4, .3]]), .5, 2) == 2

tests/test_tile_classification.py

@@ -0,0 +1,40 @@
+import pytest
+from PIL import Image
+
+from datasets.tile_classification import YoloTileClassificationDataset, tile_label
+from datasets.yolo import write_data_yaml
+
+
+def test_label_uses_human_class_ids_and_empty_background():
+    assert tile_label("", 9, 4) == 4
+    assert tile_label("1 0.5 0.5 0.2 0.2", 9, 4) == 1
+    assert tile_label("1 0.5 0.5 0.2 0.2\n3 0.5 0.5 0.4 0.4", 9, 4) == 3
+    with pytest.raises(ValueError):
+        tile_label("1 0.5 0.5 nan 0.2", 9, 4)
+
+
+def test_existing_split_is_preserved_and_no_resize(tmp_path):
+    for split in ("train", "val"):
+        (tmp_path / "images" / split).mkdir(parents=True)
+        (tmp_path / "labels" / split).mkdir(parents=True)
+        Image.new("RGB", (16, 16)).save(tmp_path / "images" / split / "sample.png")
+        (tmp_path / "labels" / split / "sample.txt").write_text(
+            "1 0.5 0.5 0.2 0.2" if split == "train" else "")
+    yaml = write_data_yaml(tmp_path / "data.yaml", ["normal", "blackdot"], root=tmp_path)
+    kwargs = {"yaml_path": yaml, "image_size": [16, 16], "transforms": [{"name": "ToTensor"}]}
+    train = YoloTileClassificationDataset(split="train", **kwargs)
+    val = YoloTileClassificationDataset(split="val", **kwargs)
+    assert train.class_counts() == [0, 1]
+    assert val.class_counts() == [1, 0]
+    assert train[0]["image"].shape == (3, 16, 16)
+    Image.new("RGB", (12, 16)).save(train.samples[0][0])
+    with pytest.raises(ValueError, match="resize is forbidden"):
+        train[0]
+
+
+def test_missing_labels_are_not_silent_background(tmp_path):
+    (tmp_path / "images" / "train").mkdir(parents=True)
+    Image.new("RGB", (16, 16)).save(tmp_path / "images" / "train" / "sample.png")
+    yaml = write_data_yaml(tmp_path / "data.yaml", ["normal", "blackdot"], root=tmp_path)
+    with pytest.raises(FileNotFoundError, match="Missing label"):
+        YoloTileClassificationDataset(yaml, "train")