Add deep training acceptance checks

2026-06-30 13:42:30 +08:00
parent 7d6e1692b1
commit cf920e97c3
8 changed files with 179 additions and 11 deletions
--- a/.env.example
+++ b/.env.example
@@ -7,4 +7,5 @@ SEG_MMSEG_CONDA_ENV=seg_mmcv
 SEG_BACKEND_CONDA_ENV=seg_smp
 SEG_WEIGHT_MODE=copy
 SEG_ENABLE_SHELL_TASKS=1
+SEG_VALIDATE_DEEP=1
 VITE_API_BASE=http://localhost:8010
--- a/README.md
+++ b/README.md
@@ -64,6 +64,12 @@ weight discovery. MMSeg full-model readiness is validated in
 `SEG_MMSEG_CONDA_ENV` by importing `mmcv._ext` and building a local MMSeg
 `EncoderDecoder` from the existing config tree.

+For stronger runtime proof, `POST /api/acceptance/deep` runs minimal training
+loops for the three model families: one SegModel optimizer step, one YOLO
+segmentation epoch on a synthetic 64x64 dataset, and one MMSeg optimizer step
+through the full `mmcv._ext` runtime. The latest report is available from
+`GET /api/acceptance/deep/latest` and is surfaced in the coverage panel.
+
 Current `seg_smp` uses `mmcv-lite` because no `torch 2.6/cu124` full `mmcv`
 wheel is available on this machine and `nvcc` is not installed for source
 builds. A dedicated `seg_mmcv` environment is used for MMSeg tasks and has
@@ -138,4 +144,5 @@ The validation agent checks catalog coverage, the `seg_smp` task env, the
 `seg_mmcv` MMSeg env, GPU visibility, no-weight Git safety, backend tests,
 frontend build, and live backend/frontend endpoints when the services are
 running. With live validation enabled it also runs the lightweight acceptance
-smoke above.
+smoke above. By default it also runs the deep training acceptance; set
+`SEG_VALIDATE_DEEP=0` when a quick non-training validation pass is needed.
--- a/backend/app/acceptance.py
+++ b/backend/app/acceptance.py
@@ -62,6 +62,66 @@ MMSEG_FULL_BUILD_SNIPPET = (
 )


+SEGMODEL_TRAIN_STEP_SNIPPET = (
+    "import torch, segmentation_models_pytorch as smp; "
+    "torch.manual_seed(7); "
+    "model=smp.Unet(encoder_name='resnet18', encoder_weights=None, classes=2).train(); "
+    "inputs=torch.randn(2,3,64,64); "
+    "targets=torch.randint(0,2,(2,64,64)); "
+    "optimizer=torch.optim.SGD(model.parameters(), lr=1e-3); "
+    "outputs=model(inputs); "
+    "loss=torch.nn.functional.cross_entropy(outputs, targets); "
+    "loss.backward(); optimizer.step(); "
+    "print('loss', round(float(loss.detach()), 6), 'shape', tuple(outputs.shape))"
+)
+
+
+def _yolo_tiny_train_snippet(root: Path, weight: Path) -> str:
+    return (
+        "import shutil, cv2, numpy as np; "
+        "from pathlib import Path; "
+        "from ultralytics import YOLO; "
+        f"root=Path({str(root)!r}); weight={str(weight)!r}; "
+        "shutil.rmtree(root, ignore_errors=True); "
+        "[ (root / item).mkdir(parents=True, exist_ok=True) for item in ['images/train','images/val','labels/train','labels/val','runs'] ]; "
+        "image=np.zeros((64,64,3), dtype=np.uint8); "
+        "cv2.rectangle(image, (16,16), (48,48), (255,255,255), -1); "
+        "label='0 0.25 0.25 0.75 0.25 0.75 0.75 0.25 0.75\\n'; "
+        "\nfor split in ['train','val']:\n"
+        "    cv2.imwrite(str(root / 'images' / split / 'sample.jpg'), image)\n"
+        "    (root / 'labels' / split / 'sample.txt').write_text(label, encoding='utf-8')\n"
+        "(root / 'data.yaml').write_text('path: '+str(root)+'\\ntrain: images/train\\nval: images/val\\nnc: 1\\nnames:\\n  0: object\\n', encoding='utf-8'); "
+        "model=YOLO(weight); "
+        "model.train(data=str(root/'data.yaml'), epochs=1, imgsz=64, batch=1, workers=0, device='cpu', project=str(root/'runs'), name='tiny', exist_ok=True, verbose=False, plots=False, val=False); "
+        "results=root/'runs'/'tiny'/'results.csv'; best=root/'runs'/'tiny'/'weights'/'best.pt'; "
+        "assert results.exists() and results.stat().st_size > 0; "
+        "assert best.exists() and best.stat().st_size > 0; "
+        "print('results', results, results.stat().st_size, 'best', best.stat().st_size)"
+    )
+
+
+def _mmseg_train_step_snippet(config_path: Path) -> str:
+    return (
+        "import torch; "
+        "from mmengine.config import Config; "
+        "from mmengine.structures import PixelData; "
+        "from mmseg.registry import MODELS; "
+        "from mmseg.structures import SegDataSample; "
+        "from mmseg.utils import register_all_modules; "
+        "register_all_modules(init_default_scope=True); "
+        f"cfg=Config.fromfile({str(config_path)!r}); "
+        "cfg.model.backbone.init_cfg=None; cfg.model.pretrained=None; "
+        "model=MODELS.build(cfg.model).train(); "
+        "sample=SegDataSample(); "
+        "sample.gt_sem_seg=PixelData(data=torch.randint(0,19,(1,64,64), dtype=torch.long)); "
+        "losses=model(torch.randn(1,3,64,64), [sample], mode='loss'); "
+        "loss=sum(value if torch.is_tensor(value) else sum(value) for value in losses.values()); "
+        "optimizer=torch.optim.SGD(model.parameters(), lr=1e-4); "
+        "loss.backward(); optimizer.step(); "
+        "print('loss', round(float(loss.detach()), 6), sorted(losses.keys()))"
+    )
+
+
 def _request_json(method: str, url: str, payload: dict[str, Any] | None = None, timeout: int = 10) -> dict[str, Any]:
    data = None
    headers = {"Accept": "application/json"}
@@ -252,6 +312,56 @@ def latest_acceptance_report() -> dict[str, Any]:
    return json.loads(path.read_text(encoding="utf-8"))


+def latest_deep_acceptance_report() -> dict[str, Any]:
+    path = settings.project_root / "var" / "acceptance" / "deep_latest.json"
+    if not path.exists():
+        return {"available": False, "path": str(path)}
+    return json.loads(path.read_text(encoding="utf-8"))
+
+
+def run_deep_acceptance() -> dict[str, Any]:
+    """Run minimal training loops for each model family without full datasets."""
+    acceptance_root = settings.project_root / "var" / "acceptance"
+    run_id = uuid.uuid4().hex[:8]
+    fixture_root = acceptance_root / f"deep_{run_id}"
+    fixture_root.mkdir(parents=True, exist_ok=True)
+
+    yolo_weight = settings.source_root / "Seg_All_In_One_YoloModel" / "yolo11n-seg.pt"
+    mmseg_config = settings.source_root / "Seg_All_In_One_MMSeg" / "configs" / "fcn" / "fcn_r18-d8_4xb2-80k_cityscapes-512x1024.py"
+
+    checks = [
+        {
+            "name": "segmodel_tiny_train_step",
+            "passed": False,
+            "detail": _run_snippet(SEGMODEL_TRAIN_STEP_SNIPPET, timeout=90),
+        },
+        {
+            "name": "yolo_tiny_segment_train_epoch",
+            "passed": False,
+            "detail": _run_snippet(_yolo_tiny_train_snippet(fixture_root / "yolo_tiny", yolo_weight), timeout=180),
+        },
+        {
+            "name": "mmseg_tiny_train_step",
+            "passed": False,
+            "detail": _run_conda_snippet(settings.mmseg_conda_env, _mmseg_train_step_snippet(mmseg_config), timeout=120),
+        },
+    ]
+    for check in checks:
+        check["passed"] = bool(check["detail"].get("passed"))
+
+    report = {
+        "available": True,
+        "run_id": run_id,
+        "fixture_root": str(fixture_root),
+        "passed": all(item["passed"] for item in checks),
+        "checks": checks,
+        "created_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
+    }
+    latest = acceptance_root / "deep_latest.json"
+    latest.write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
+    return report
+
+
 def run_live_acceptance(base_url: str = "http://127.0.0.1:8010") -> dict[str, Any]:
    """Run a lightweight end-to-end smoke against the live API and job runner."""
    acceptance_root = settings.project_root / "var" / "acceptance"
--- a/backend/app/agents/evaluation_agent.py
+++ b/backend/app/agents/evaluation_agent.py
@@ -42,6 +42,8 @@ def evaluate_project() -> dict:
        "loss_result_ui": "loss" in frontend_text.lower() and "heatmap" in frontend_text.lower() and "CurvePanel" in frontend_text,
        "dataset_api": "/api/datasets" in backend_text and "api_upload_dataset_files" in backend_text,
        "curve_api": "/api/results/curves" in backend_text,
+        "deep_acceptance_api": "/api/acceptance/deep" in backend_text,
+        "deep_acceptance_ui": "runDeepAcceptance" in frontend_text and "深度训练" in frontend_text,
        "coverage_api": "/api/coverage" in backend_text and coverage["task_build_passed"],
        "visual_tools": "visual.yolo11_heatmap_v2" in catalog["task_types"] and "visual.fps" in catalog["task_types"],
        "yolo_dataset_tools": "dataset.yolo_txt_sort" in catalog["task_types"] and "dataset.yolo_resize" in catalog["task_types"],
--- a/backend/app/agents/validation_agent.py
+++ b/backend/app/agents/validation_agent.py
@@ -8,7 +8,7 @@ import urllib.error
 import urllib.request
 from pathlib import Path

-from ..acceptance import run_live_acceptance
+from ..acceptance import run_deep_acceptance, run_live_acceptance
 from ..catalog import get_catalog
 from ..config import settings
 from ..coverage import get_coverage_report
@@ -108,6 +108,9 @@ def validate_project(run_build: bool = False) -> dict:
        if os.getenv("SEG_VALIDATE_ACCEPTANCE", "1") == "1":
            acceptance = run_live_acceptance(backend_url)
            checks.append({"name": "live_acceptance_smoke", "passed": acceptance["passed"], "detail": acceptance})
+        if os.getenv("SEG_VALIDATE_DEEP", "1") == "1":
+            deep_acceptance = run_deep_acceptance()
+            checks.append({"name": "deep_training_acceptance", "passed": deep_acceptance["passed"], "detail": deep_acceptance})

    if run_build:
        tests = _run(["conda", "run", "-n", settings.backend_conda_env, "python", "-m", "pytest", "-q"], cwd=settings.project_root, timeout=120)
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -9,7 +9,7 @@ from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, StreamingResponse

 from . import db
-from .acceptance import latest_acceptance_report, run_live_acceptance
+from .acceptance import latest_acceptance_report, latest_deep_acceptance_report, run_deep_acceptance, run_live_acceptance
 from .catalog import get_catalog
 from .config import settings
 from .coverage import get_coverage_report
@@ -81,6 +81,16 @@ def api_acceptance_smoke(base_url: str = "http://127.0.0.1:8010") -> dict:
    return run_live_acceptance(base_url)


+@app.get("/api/acceptance/deep/latest")
+def api_deep_acceptance_latest() -> dict:
+    return latest_deep_acceptance_report()
+
+
+@app.post("/api/acceptance/deep")
+def api_deep_acceptance() -> dict:
+    return run_deep_acceptance()
+
+
@app.get("/api/datasets")
 def api_datasets() -> list[dict]:
    return list_uploaded_datasets()
--- a/frontend/src/main.tsx
+++ b/frontend/src/main.tsx
@@ -109,6 +109,14 @@ type AcceptancePayload = {
  };
 };

+type DeepAcceptancePayload = {
+  available?: boolean;
+  passed?: boolean;
+  run_id?: string;
+  created_at?: string;
+  checks?: Array<{ name: string; passed: boolean }>;
+};
+
 type GpuPayload = {
  available: boolean;
  gpus: Array<{
@@ -191,11 +199,12 @@ function useData() {
  const [datasets, setDatasets] = useState<UploadedDataset[]>([]);
  const [coverage, setCoverage] = useState<CoveragePayload | null>(null);
  const [acceptance, setAcceptance] = useState<AcceptancePayload | null>(null);
+  const [deepAcceptance, setDeepAcceptance] = useState<DeepAcceptancePayload | null>(null);
  const [error, setError] = useState<string>("");

  async function refresh() {
    try {
-      const [catalogNext, gpusNext, jobsNext, resultsNext, curvesNext, datasetsNext, coverageNext, acceptanceNext] = await Promise.all([
+      const [catalogNext, gpusNext, jobsNext, resultsNext, curvesNext, datasetsNext, coverageNext, acceptanceNext, deepAcceptanceNext] = await Promise.all([
        api<Catalog>("/api/catalog"),
        api<GpuPayload>("/api/system/gpus"),
        api<Job[]>("/api/jobs"),
@@ -203,7 +212,8 @@ function useData() {
        api<TrainingCurve[]>("/api/results/curves"),
        api<UploadedDataset[]>("/api/datasets"),
        api<CoveragePayload>("/api/coverage"),
-        api<AcceptancePayload>("/api/acceptance/latest")
+        api<AcceptancePayload>("/api/acceptance/latest"),
+        api<DeepAcceptancePayload>("/api/acceptance/deep/latest")
      ]);
      setCatalog(catalogNext);
      setGpus(gpusNext);
@@ -213,6 +223,7 @@ function useData() {
      setDatasets(datasetsNext);
      setCoverage(coverageNext);
      setAcceptance(acceptanceNext);
+      setDeepAcceptance(deepAcceptanceNext);
      setError("");
    } catch (err) {
      setError(String(err));
@@ -225,7 +236,7 @@ function useData() {
    return () => window.clearInterval(timer);
  }, []);

-  return { catalog, gpus, jobs, results, curves, datasets, coverage, acceptance, error, refresh };
+  return { catalog, gpus, jobs, results, curves, datasets, coverage, acceptance, deepAcceptance, error, refresh };
 }

 function StatusPill({ status }: { status: string }) {
@@ -233,7 +244,7 @@ function StatusPill({ status }: { status: string }) {
 }

 function App() {
-  const { catalog, gpus, jobs, results, curves, datasets, coverage, acceptance, error, refresh } = useData();
+  const { catalog, gpus, jobs, results, curves, datasets, coverage, acceptance, deepAcceptance, error, refresh } = useData();
  const [taskType, setTaskType] = useState("mock.echo");
  const [params, setParams] = useState(JSON.stringify(defaultParams["mock.echo"], null, 2));
  const [selectedJob, setSelectedJob] = useState<Job | null>(null);
@@ -333,6 +344,16 @@ function App() {
    }
  }

+  async function runDeepAcceptance() {
+    setBusy(true);
+    try {
+      await api("/api/acceptance/deep", { method: "POST" });
+      await refresh();
+    } finally {
+      setBusy(false);
+    }
+  }
+
  async function createDataset() {
    setBusy(true);
    try {
@@ -591,9 +612,14 @@ function App() {
                <p className="eyebrow">Coverage</p>
                <h2>Seg 功能覆盖</h2>
              </div>
-              <button className="iconButton" disabled={busy} onClick={runAcceptanceSmoke} title="运行轻量验收">
-                <ClipboardCheck size={18} />
-              </button>
+              <div className="buttonRow compactButtons">
+                <button className="iconButton" disabled={busy} onClick={runAcceptanceSmoke} title="运行轻量验收">
+                  <ClipboardCheck size={18} />
+                </button>
+                <button className="iconButton" disabled={busy} onClick={runDeepAcceptance} title="运行深度训练验收">
+                  <Activity size={18} />
+                </button>
+              </div>
            </div>
            <div className="coverageGrid">
              <div>
@@ -616,12 +642,17 @@ function App() {
                <span>模型族</span>
                <strong>{acceptance?.model_family_readiness?.passed ? "OK" : "Check"}</strong>
              </div>
+              <div>
+                <span>深度训练</span>
+                <strong>{deepAcceptance?.available === false ? "New" : deepAcceptance?.passed ? "OK" : "Check"}</strong>
+              </div>
            </div>
            <div className="coverageStatus">
              {(coverage?.unmapped_user_scripts.length ?? 0) === 0 ? (
                <>
                  <span>当前用户侧脚本已全部映射到网页任务。</span>
                  <span>最近验收：{acceptance?.created_at ?? "尚未运行"} {acceptance?.run_id ? `#${acceptance.run_id}` : ""}</span>
+                  <span>深度验收：{deepAcceptance?.created_at ?? "尚未运行"} {deepAcceptance?.run_id ? `#${deepAcceptance.run_id}` : ""}，通过 {deepAcceptance?.checks?.filter((item) => item.passed).length ?? 0}/{deepAcceptance?.checks?.length ?? 0}</span>
                  <span>模型族 readiness：{acceptance?.model_family_readiness?.checks?.filter((item) => item.passed).length ?? 0}/{acceptance?.model_family_readiness?.checks?.length ?? 0}，warnings {acceptance?.model_family_readiness?.warnings?.length ?? 0}</span>
                </>
              ) : (
--- a/frontend/src/styles.css
+++ b/frontend/src/styles.css
@@ -358,6 +358,10 @@ textarea {
  gap: 10px;
 }

+.buttonRow.compactButtons {
+  gap: 8px;
+}
+
 .opGrid {
  display: grid;
  grid-template-columns: repeat(4, minmax(0, 1fr));
@@ -399,7 +403,7 @@ textarea {

 .coverageGrid {
  display: grid;
-  grid-template-columns: repeat(5, minmax(0, 1fr));
+  grid-template-columns: repeat(6, minmax(0, 1fr));
  gap: 10px;
  margin-bottom: 14px;
 }