bugfixes

backend/ml/train_detector_model.py

@@ -385,6 +385,7 @@ def main():
         "ok": True,
         "model": str(final_model),
         "sourceModel": str(best),
+        "baseModel": str(base_used),
         "runs": str(result_path),
         "trainedAt": datetime.now(timezone.utc).isoformat(),
         "trainSamples": train_count,
@@ -392,6 +393,7 @@ def main():
         "epochs": epochs,
         "imgsz": imgsz,
         "batchSize": batch_size,
+        "patience": patience,
         "threads": threads,
         "workers": workers,
         "device": str(train_device),

backend/ml/train_pose_model.py

@@ -371,6 +371,7 @@ def main():
         "ok": True,
         "model": str(final_model),
         "sourceModel": str(best),
+        "baseModel": str(base_model),
         "runs": str(result_path),
         "trainedAt": datetime.now(timezone.utc).isoformat(),
         "trainSamples": train_count,
@@ -378,6 +379,7 @@ def main():
         "epochs": epochs,
         "imgsz": imgsz,
         "batchSize": batch_size,
+        "patience": patience,
         "threads": threads,
         "workers": workers,
         "device": str(train_device),

backend/ml/train_videomae_model.py

@@ -200,6 +200,7 @@ def main():
     parser.add_argument("--workers", default="0")
     parser.add_argument("--threads", default="0")
     parser.add_argument("--num-frames", default="16")
+    parser.add_argument("--patience", default="3")
     parser.add_argument("--freeze-backbone", action="store_true")
     args = parser.parse_args()
 
@@ -213,6 +214,7 @@ def main():
     workers = max(0, safe_int(args.workers, 0))
     threads = max(0, safe_int(args.threads, 0))
     num_frames = max(2, safe_int(args.num_frames, 16))
+    patience = max(0, safe_int(args.patience, 3))
     lr = max(1e-7, safe_float(args.lr, 5e-5))
 
     if threads > 0:
@@ -249,18 +251,23 @@ def main():
     if not val_entries:
         raise SystemExit("no VideoMAE val clips found")
 
+    base_model = str(args.base or DEFAULT_BASE_MODEL).strip() or DEFAULT_BASE_MODEL
+    base_path = Path(base_model).expanduser()
+    if base_path.exists():
+        base_model = str(base_path.resolve())
+
     emit_progress(
         "videomae",
         0.03,
         "VideoMAE-Basismodell wird geladen...",
-        base=args.base,
+        base=base_model,
         labels=len(labels),
         device=str(device),
     )
 
-    image_processor = AutoImageProcessor.from_pretrained(args.base)
+    image_processor = AutoImageProcessor.from_pretrained(base_model)
     model = VideoMAEForVideoClassification.from_pretrained(
-        args.base,
+        base_model,
         num_labels=len(labels),
         label2id=label_to_id,
         id2label=id_to_label,
@@ -302,8 +309,11 @@ def main():
     best_accuracy = -1.0
     best_loss = math.inf
     best_epoch = 0
+    completed_epochs = 0
+    epochs_without_improvement = 0
 
     for epoch in range(1, epochs + 1):
+        completed_epochs = epoch
         model.train()
         running_loss = 0.0
         seen = 0
@@ -346,8 +356,11 @@ def main():
             best_accuracy = val_accuracy
             best_loss = val_loss
             best_epoch = epoch
+            epochs_without_improvement = 0
             model.save_pretrained(tmp_dir)
             image_processor.save_pretrained(tmp_dir)
+        else:
+            epochs_without_improvement += 1
 
         emit_progress(
             "videomae",
@@ -360,16 +373,36 @@ def main():
             device=str(device),
             accuracy=val_accuracy,
             loss=val_loss,
+            patience=patience,
+            epochsWithoutImprovement=epochs_without_improvement,
         )
 
+        if patience > 0 and epochs_without_improvement >= patience:
+            emit_progress(
+                "videomae",
+                0.96,
+                "VideoMAE Early-Stopping: keine weitere Verbesserung.",
+                epoch=epoch,
+                epochs=epochs,
+                bestEpoch=best_epoch,
+                patience=patience,
+                trainSamples=len(train_entries),
+                valSamples=len(val_entries),
+                device=str(device),
+                accuracy=best_accuracy if best_accuracy >= 0 else 0.0,
+                loss=best_loss if math.isfinite(best_loss) else 0.0,
+            )
+            break
+
     if best_epoch <= 0:
         model.save_pretrained(tmp_dir)
         image_processor.save_pretrained(tmp_dir)
-        best_epoch = epochs
+        best_epoch = completed_epochs or epochs
 
     status = {
         "trainedAt": datetime.now(timezone.utc).isoformat(),
         "epochs": epochs,
+        "completedEpochs": completed_epochs,
         "bestEpoch": best_epoch,
         "trainSamples": len(train_entries),
         "valSamples": len(val_entries),
@@ -377,7 +410,8 @@ def main():
         "workers": workers,
         "threads": threads,
         "numFrames": num_frames,
-        "baseModel": args.base,
+        "patience": patience,
+        "baseModel": base_model,
         "device": str(device),
         "accuracy": best_accuracy if best_accuracy >= 0 else 0.0,
         "loss": best_loss if math.isfinite(best_loss) else 0.0,

backend/training.go

@@ -118,6 +118,103 @@ type TrainingSkipRequest struct {
 	SampleID string `json:"sampleId"`
 }
 
+type TrainingTrainRequest struct {
+	Scope   string   `json:"scope,omitempty"`
+	Targets []string `json:"targets,omitempty"`
+}
+
+type trainingTrainTargets struct {
+	Detector bool
+	Pose     bool
+	VideoMAE bool
+}
+
+func trainingAllTrainTargets() trainingTrainTargets {
+	return trainingTrainTargets{
+		Detector: true,
+		Pose:     true,
+		VideoMAE: true,
+	}
+}
+
+func (t trainingTrainTargets) empty() bool {
+	return !t.Detector && !t.Pose && !t.VideoMAE
+}
+
+func (t trainingTrainTargets) list() []string {
+	out := make([]string, 0, 3)
+	if t.Detector {
+		out = append(out, "detector")
+	}
+	if t.Pose {
+		out = append(out, "pose")
+	}
+	if t.VideoMAE {
+		out = append(out, "videomae")
+	}
+	return out
+}
+
+func trainingReadTrainRequest(r *http.Request) (TrainingTrainRequest, error) {
+	var req TrainingTrainRequest
+	if r.Body == nil {
+		return req, nil
+	}
+
+	body, err := io.ReadAll(io.LimitReader(r.Body, 1<<20))
+	if err != nil {
+		return req, err
+	}
+
+	if strings.TrimSpace(string(body)) == "" {
+		return req, nil
+	}
+
+	if err := json.Unmarshal(body, &req); err != nil {
+		return req, err
+	}
+
+	return req, nil
+}
+
+func trainingNormalizeTrainTargets(req TrainingTrainRequest) (trainingTrainTargets, bool, error) {
+	scope := strings.ToLower(strings.TrimSpace(req.Scope))
+	if scope == "" || scope == "full" || scope == "all" || scope == "complete" {
+		return trainingAllTrainTargets(), false, nil
+	}
+
+	if scope != "custom" && scope != "selected" && scope != "partial" {
+		return trainingTrainTargets{}, false, fmt.Errorf("unbekannter Trainingsumfang: %s", req.Scope)
+	}
+
+	var targets trainingTrainTargets
+	for _, raw := range req.Targets {
+		key := strings.ToLower(strings.TrimSpace(raw))
+		key = strings.ReplaceAll(key, "_", "")
+		key = strings.ReplaceAll(key, "-", "")
+		key = strings.ReplaceAll(key, " ", "")
+
+		switch key {
+		case "detector", "yolo", "yolo26", "yolo26detector", "box", "boxes", "boxdetection":
+			targets.Detector = true
+		case "pose", "yolo26pose", "posedetection":
+			targets.Pose = true
+		case "videomae", "video", "clip", "scene", "clipanalysis", "clipanalyse":
+			targets.VideoMAE = true
+		case "":
+			continue
+		default:
+			return trainingTrainTargets{}, true, fmt.Errorf("unbekanntes Training: %s", raw)
+		}
+	}
+
+	if targets.empty() {
+		return trainingTrainTargets{}, true, errors.New("kein Training ausgewählt")
+	}
+
+	return targets, true, nil
+}
+
 type TrainingAnnotation struct {
 	SampleID        string              `json:"sampleId"`
 	FrameURL        string              `json:"frameUrl"`
@@ -1313,6 +1410,21 @@ func trainingRuntimeOptionsFromRecorderSettings(s RecorderSettings) trainingRunt
 	return opts
 }
 
+func trainingYoloEarlyStoppingPatience(epochs int) int {
+	epochs = clampTrainingInt(epochs, 1, 300)
+	patience := epochs / 4
+	if patience < 5 {
+		patience = 5
+	}
+	if patience > 20 {
+		patience = 20
+	}
+	if patience > epochs {
+		patience = epochs
+	}
+	return patience
+}
+
 func trainingCommandEnv(opts trainingRuntimeOptions) []string {
 	env := os.Environ()
 	if opts.CPUThreads <= 0 {
@@ -3212,6 +3324,18 @@ func trainingTrainHandler(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	req, err := trainingReadTrainRequest(r)
+	if err != nil {
+		trainingWriteError(w, http.StatusBadRequest, "invalid json")
+		return
+	}
+
+	targets, customTargets, err := trainingNormalizeTrainTargets(req)
+	if err != nil {
+		trainingWriteError(w, http.StatusBadRequest, err.Error())
+		return
+	}
+
 	current := trainingGetJobStatus()
 	if current.Running {
 		trainingWriteJSON(w, http.StatusOK, map[string]any{
@@ -3301,7 +3425,70 @@ func trainingTrainHandler(w http.ResponseWriter, r *http.Request) {
 	videoMAEDataReady := videoMAEEligibleCount >= minVideoMAETrainCount ||
 		(videoMAETrainCount >= minVideoMAETrainCount && videoMAEValCount >= minVideoMAEValCount)
 
-	if detectorDataReady || poseDataReady || videoMAEDataReady {
+	runtimeOpts := trainingRuntimeOptionsFromSettings()
+	selectedDataReady :=
+		(targets.Detector && detectorDataReady) ||
+			(targets.Pose && poseDataReady) ||
+			(targets.VideoMAE && videoMAEDataReady)
+
+	if customTargets {
+		if targets.Detector && !detectorDataReady {
+			trainingWriteError(
+				w,
+				http.StatusBadRequest,
+				fmt.Sprintf(
+					"YOLO26 Detector ist noch nicht trainingsbereit. Train=%d (%d positiv), Val=%d (%d positiv). Benoetigt: mindestens %d Train, %d Val und je ein positives Beispiel.",
+					trainCount,
+					positiveTrainCount,
+					valCount,
+					positiveValCount,
+					minDetectorTrainCount,
+					minDetectorValCount,
+				),
+			)
+			return
+		}
+
+		if targets.Pose && !poseDataReady {
+			trainingWriteError(
+				w,
+				http.StatusBadRequest,
+				fmt.Sprintf(
+					"YOLO26 Pose ist noch nicht trainingsbereit. Train=%d, Val=%d. Benoetigt: mindestens %d Train und %d Val.",
+					poseTrainCount,
+					poseValCount,
+					minPoseTrainCount,
+					minPoseValCount,
+				),
+			)
+			return
+		}
+
+		if targets.VideoMAE && !runtimeOpts.VideoMAEEnabled {
+			trainingWriteError(w, http.StatusBadRequest, "VideoMAE ist in den Training-Settings deaktiviert.")
+			return
+		}
+
+		if targets.VideoMAE && !videoMAEDataReady {
+			trainingWriteError(
+				w,
+				http.StatusBadRequest,
+				fmt.Sprintf(
+					"VideoMAE ist noch nicht trainingsbereit. Eligible=%d, Train=%d, Val=%d. Benoetigt: mindestens %d Train und %d Val.",
+					videoMAEEligibleCount,
+					videoMAETrainCount,
+					videoMAEValCount,
+					minVideoMAETrainCount,
+					minVideoMAEValCount,
+				),
+			)
+			return
+		}
+
+		goto startTraining
+	}
+
+	if selectedDataReady {
 		goto startTraining
 	}
 
@@ -3348,12 +3535,13 @@ startTraining:
 
 	trainingStartJob(cancel)
 
-	go trainingRunJob(ctx, root, feedbackCount)
+	go trainingRunJob(ctx, root, feedbackCount, targets)
 
 	trainingWriteJSON(w, http.StatusAccepted, map[string]any{
 		"ok":       true,
 		"message":  "Training gestartet.",
 		"training": trainingGetJobStatus(),
+		"targets":  targets.list(),
 		"detector": map[string]any{
 			"trainCount":         trainCount,
 			"valCount":           valCount,
@@ -3382,6 +3570,7 @@ startTraining:
 			"requiredTrain": minVideoMAETrainCount,
 			"requiredVal":   minVideoMAEValCount,
 			"manifest":      videoMAEManifest,
+			"dataReady":     videoMAEDataReady,
 			"source":        "videomae_clip",
 		},
 	})
@@ -3424,7 +3613,7 @@ func trainingCancelHandler(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
-func trainingRunJob(ctx context.Context, root string, count int) {
+func trainingRunJob(ctx context.Context, root string, count int, targets trainingTrainTargets) {
 	if err := ensureMLPythonSetup(ctx); err != nil {
 		if errors.Is(err, context.Canceled) {
 			trainingFinishCancelled(root)
@@ -3459,7 +3648,7 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 	appLogln("ML-Python für Training:", python)
 	runtimeOpts := trainingRuntimeOptionsFromSettings()
 	appLogf(
-		"Training-Laufzeit: mode=%s cpuCores=%d schonmodus=%v threads=%d workers=%d yoloBatch=%d lowPriority=%v videoMAE=%v",
+		"Training-Laufzeit: mode=%s cpuCores=%d schonmodus=%v threads=%d workers=%d yoloBatch=%d lowPriority=%v videoMAE=%v targets=%s",
 		runtimeOpts.PerformanceMode,
 		runtimeOpts.CPUCoreCount,
 		runtimeOpts.PowerSaveMode,
@@ -3468,6 +3657,7 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 		runtimeOpts.YoloBatchSize,
 		runtimeOpts.LowPriority,
 		runtimeOpts.VideoMAEEnabled,
+		strings.Join(targets.list(), ","),
 	)
 
 	cleanOutput := func(text string) string {
@@ -3485,10 +3675,13 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 
 	detectorOutput := ""
 	detectorStatus := "skipped"
+	detectorDurationMs := int64(0)
 	poseOutput := ""
 	poseStatus := "skipped"
+	poseDurationMs := int64(0)
 	videoMAEOutput := ""
 	videoMAEStatus := "skipped"
+	videoMAEDurationMs := int64(0)
 
 	detectorDatasetYAML := filepath.Join(root, "detector", "dataset", "dataset.yaml")
 	detectorTrainImages := filepath.Join(root, "detector", "dataset", "images", "train")
@@ -3514,29 +3707,50 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 		fileExistsNonEmpty(detectorDatasetYAML),
 	)
 
-	if fileExistsNonEmpty(detectorDatasetYAML) &&
+	if !targets.Detector {
+		trainingSetJobStatus(func(s *TrainingJobStatus) {
+			if s.Progress < 58 {
+				s.Progress = 58
+			}
+			s.Step = "YOLO26 Detector wurde nicht ausgewählt."
+		})
+		detectorStatus = "skipped_unselected"
+		detectorOutput = "YOLO26 Detector übersprungen: nicht ausgewählt."
+		appLogln(detectorOutput)
+	} else if fileExistsNonEmpty(detectorDatasetYAML) &&
 		trainCount >= minDetectorTrainCount &&
 		valCount >= minDetectorValCount &&
 		positiveTrainCount > 0 &&
 		positiveValCount > 0 {
+		detectorStartedAt := time.Now()
 		trainingSetJobStatus(func(s *TrainingJobStatus) {
 			s.Progress = 15
 			s.Step = "YOLO26 Detector wird trainiert…"
 		})
 
-		detectorBasePath := "yolo26n.pt"
-		if p, err := trainingMLCacheFilePath("yolo26n.pt"); err == nil {
-			detectorBasePath = p
+		detectorModel := trainingResolveDetectorModel(root)
+		detectorBasePath := detectorModel.BestPath
+		if !detectorModel.TrainedExists {
+			detectorBasePath = "yolo26n.pt"
+			if p, err := trainingMLCacheFilePath("yolo26n.pt"); err == nil {
+				detectorBasePath = p
+			}
+		}
+		detectorEpochs := trainingDetectorEpochs()
+		detectorPatience := trainingYoloEarlyStoppingPatience(detectorEpochs)
+		if detectorModel.TrainedExists {
+			appLogln("YOLO26 Detector Fine-Tuning startet von:", detectorBasePath)
 		}
 
 		detectorScript := trainingScriptPath("train_detector_model.py")
 		detectorArgs := []string{
 			"--root", root,
 			"--base", detectorBasePath,
-			"--epochs", strconv.Itoa(trainingDetectorEpochs()),
+			"--epochs", strconv.Itoa(detectorEpochs),
 			"--imgsz", "640",
 			"--workers", strconv.Itoa(runtimeOpts.Workers),
 			"--threads", strconv.Itoa(runtimeOpts.CPUThreads),
+			"--patience", strconv.Itoa(detectorPatience),
 		}
 		if runtimeOpts.YoloBatchSize > 0 {
 			detectorArgs = append(detectorArgs, "--batch", strconv.Itoa(runtimeOpts.YoloBatchSize))
@@ -3556,6 +3770,7 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 			},
 			detectorArgs...,
 		)
+		detectorDurationMs = time.Since(detectorStartedAt).Milliseconds()
 
 		if errors.Is(detectorErr, errTrainingCancelled) {
 			appLogln("⛔ YOLO26 detector training cancelled")
@@ -3601,24 +3816,38 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 	poseTrainLabels := filepath.Join(root, "pose", "dataset", "labels", "train")
 	poseValImages := filepath.Join(root, "pose", "dataset", "images", "val")
 	poseValLabels := filepath.Join(root, "pose", "dataset", "labels", "val")
+	poseStartedAt := time.Time{}
 
-	trainingSetJobStatus(func(s *TrainingJobStatus) {
-		if s.Progress < 60 {
-			s.Progress = 60
-		}
-		s.Step = "YOLO26 Pose-Daten werden aufgebaut..."
-	})
-
-	if written, err := trainingSyncPoseDataset(root); err != nil {
-		poseStatus = "failed"
-		poseOutput = "YOLO26 Pose-Dataset konnte nicht aufgebaut werden: " + err.Error()
+	if !targets.Pose {
+		trainingSetJobStatus(func(s *TrainingJobStatus) {
+			if s.Progress < 82 {
+				s.Progress = 82
+			}
+			s.Step = "YOLO26 Pose wurde nicht ausgewählt."
+		})
+		poseStatus = "skipped_unselected"
+		poseOutput = "YOLO26 Pose übersprungen: nicht ausgewählt."
 		appLogln(poseOutput)
 	} else {
-		appLogln("pose samples synced:", written)
-	}
+		poseStartedAt = time.Now()
+		trainingSetJobStatus(func(s *TrainingJobStatus) {
+			if s.Progress < 60 {
+				s.Progress = 60
+			}
+			s.Step = "YOLO26 Pose-Daten werden aufgebaut..."
+		})
 
-	if err := trainingEnsurePoseValidationSample(root); err != nil {
-		appLogln("pose val sample ensure failed:", err)
+		if written, err := trainingSyncPoseDataset(root); err != nil {
+			poseStatus = "failed"
+			poseOutput = "YOLO26 Pose-Dataset konnte nicht aufgebaut werden: " + err.Error()
+			appLogln(poseOutput)
+		} else {
+			appLogln("pose samples synced:", written)
+		}
+
+		if err := trainingEnsurePoseValidationSample(root); err != nil {
+			appLogln("pose val sample ensure failed:", err)
+		}
 	}
 
 	poseTrainCount := trainingCountDetectorSamples(poseTrainImages, poseTrainLabels)
@@ -3631,7 +3860,8 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 		fileExistsNonEmpty(poseDatasetYAML),
 	)
 
-	if poseStatus != "failed" &&
+	if targets.Pose &&
+		poseStatus != "failed" &&
 		fileExistsNonEmpty(poseDatasetYAML) &&
 		poseTrainCount >= minPoseTrainCount &&
 		poseValCount >= minPoseValCount {
@@ -3642,19 +3872,26 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 			s.Step = "YOLO26 Pose wird trainiert..."
 		})
 
+		poseModel := trainingResolvePoseModel(root)
 		poseBasePath := "yolo26n-pose.pt"
-		if p, err := embeddedPoseModelPath(); err == nil && fileExistsNonEmpty(p) {
-			poseBasePath = p
+		if poseModel.EffectiveExists {
+			poseBasePath = poseModel.EffectivePath
+		}
+		poseEpochs := trainingDetectorEpochs()
+		posePatience := trainingYoloEarlyStoppingPatience(poseEpochs)
+		if poseModel.TrainedExists {
+			appLogln("YOLO26 Pose Fine-Tuning startet von:", poseBasePath)
 		}
 
 		poseScript := trainingScriptPath("train_pose_model.py")
 		poseArgs := []string{
 			"--root", root,
 			"--base", poseBasePath,
-			"--epochs", strconv.Itoa(trainingDetectorEpochs()),
+			"--epochs", strconv.Itoa(poseEpochs),
 			"--imgsz", "640",
 			"--workers", strconv.Itoa(runtimeOpts.Workers),
 			"--threads", strconv.Itoa(runtimeOpts.CPUThreads),
+			"--patience", strconv.Itoa(posePatience),
 		}
 		if runtimeOpts.YoloBatchSize > 0 {
 			poseArgs = append(poseArgs, "--batch", strconv.Itoa(runtimeOpts.YoloBatchSize))
@@ -3674,6 +3911,7 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 			},
 			poseArgs...,
 		)
+		poseDurationMs = time.Since(poseStartedAt).Milliseconds()
 
 		if errors.Is(poseErr, errTrainingCancelled) {
 			appLogln("YOLO26 pose training cancelled")
@@ -3698,7 +3936,7 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 				appLogln("YOLO26 pose training:", poseOutputClean)
 			}
 		}
-	} else if poseStatus != "failed" {
+	} else if targets.Pose && poseStatus != "failed" {
 		poseStatus = "skipped_no_pose_data"
 		poseOutput = fmt.Sprintf(
 			"YOLO26 Pose übersprungen: zu wenige Skeleton-Beispiele. Train=%d, Val=%d. Benoetigt: mindestens %d Train und %d Val.",
@@ -3710,9 +3948,26 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 		appLogln(poseOutput)
 	}
 
-	poseOutputClean := cleanOutput(poseOutput)
+	if poseStatus == "trained" && poseDurationMs <= 0 && !poseStartedAt.IsZero() {
+		poseDurationMs = time.Since(poseStartedAt).Milliseconds()
+	}
 
-	if !runtimeOpts.VideoMAEEnabled {
+	poseOutputClean := cleanOutput(poseOutput)
+	videoMAEStartedAt := time.Time{}
+	videoMAETrainCount := 0
+	videoMAEValCount := 0
+
+	if !targets.VideoMAE {
+		trainingSetJobStatus(func(s *TrainingJobStatus) {
+			if s.Progress < 98 {
+				s.Progress = 98
+			}
+			s.Step = "VideoMAE wurde nicht ausgewählt."
+		})
+		videoMAEStatus = "skipped_unselected"
+		videoMAEOutput = "VideoMAE übersprungen: nicht ausgewählt."
+		appLogln(videoMAEOutput)
+	} else if !runtimeOpts.VideoMAEEnabled {
 		trainingSetJobStatus(func(s *TrainingJobStatus) {
 			if s.Progress < 84 {
 				s.Progress = 84
@@ -3723,6 +3978,7 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 		videoMAEOutput = "VideoMAE übersprungen: in den Training-Settings deaktiviert."
 		appLogln(videoMAEOutput)
 	} else {
+		videoMAEStartedAt = time.Now()
 		trainingSetJobStatus(func(s *TrainingJobStatus) {
 			if s.Progress < 84 {
 				s.Progress = 84
@@ -3730,7 +3986,9 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 			s.Step = "VideoMAE Clip-Daten werden aufgebaut..."
 		})
 
-		videoMAETrainCount, videoMAEValCount, videoMAEWritten, videoMAESyncErr :=
+		var videoMAEWritten int
+		var videoMAESyncErr error
+		videoMAETrainCount, videoMAEValCount, videoMAEWritten, videoMAESyncErr =
 			trainingSyncVideoMAEDataset(ctx, root)
 		if errors.Is(videoMAESyncErr, context.Canceled) || errors.Is(videoMAESyncErr, errTrainingCancelled) {
 			appLogln("VideoMAE dataset sync cancelled")
@@ -3761,6 +4019,13 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 			})
 
 			videoMAEScript := trainingScriptPath("train_videomae_model.py")
+			videoMAEBase := strings.TrimSpace(os.Getenv("VIDEOMAE_BASE_MODEL"))
+			if videoMAEBase == "" {
+				if model := trainingResolveVideoMAEModel(root); model.TrainedExists {
+					videoMAEBase = model.EffectivePath
+					appLogln("VideoMAE Fine-Tuning startet von:", videoMAEBase)
+				}
+			}
 			videoMAEArgs := []string{
 				"--root", root,
 				"--epochs", strconv.Itoa(trainingVideoMAEEpochs()),
@@ -3768,9 +4033,10 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 				"--num-frames", strconv.Itoa(trainingVideoMAENumFrames),
 				"--workers", strconv.Itoa(runtimeOpts.Workers),
 				"--threads", strconv.Itoa(runtimeOpts.CPUThreads),
+				"--patience", strconv.Itoa(trainingVideoMAEEarlyStoppingPatience()),
 			}
-			if base := strings.TrimSpace(os.Getenv("VIDEOMAE_BASE_MODEL")); base != "" {
-				videoMAEArgs = append(videoMAEArgs, "--base", base)
+			if videoMAEBase != "" {
+				videoMAEArgs = append(videoMAEArgs, "--base", videoMAEBase)
 			}
 
 			videoMAEOut, videoMAEErr := trainingRunCommandStreaming(
@@ -3787,6 +4053,7 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 				},
 				videoMAEArgs...,
 			)
+			videoMAEDurationMs = time.Since(videoMAEStartedAt).Milliseconds()
 
 			if errors.Is(videoMAEErr, errTrainingCancelled) {
 				appLogln("VideoMAE training cancelled")
@@ -3822,6 +4089,10 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 		}
 	}
 
+	if videoMAEStatus == "trained" && videoMAEDurationMs <= 0 && !videoMAEStartedAt.IsZero() {
+		videoMAEDurationMs = time.Since(videoMAEStartedAt).Milliseconds()
+	}
+
 	videoMAEOutputClean := cleanOutput(videoMAEOutput)
 
 	message := "Training abgeschlossen."
@@ -3831,6 +4102,9 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 	case "trained":
 		message = "Training abgeschlossen. YOLO26 Detector wurde trainiert."
 
+	case "skipped_unselected":
+		message = "Training abgeschlossen."
+
 	case "skipped_no_detector_data":
 		message = detectorOutput
 
@@ -3887,8 +4161,31 @@ func trainingRunJob(ctx context.Context, root string, count int) {
 	}
 
 	if detectorStatus == "trained" || poseStatus == "trained" || videoMAEStatus == "trained" {
-		// Verlaufseintrag schreiben, solange die Job-Startzeit für die Dauer noch verfügbar ist.
-		trainingAppendRunHistory(root)
+		// Verlauf pro erfolgreich trainiertem Ziel schreiben.
+		if detectorStatus == "trained" {
+			trainingAppendTargetHistory(root, "detector", detectorStatus, detectorDurationMs, runtimeOpts, TrainingHistoryEntry{
+				Epochs:       trainingDetectorEpochs(),
+				TrainSamples: trainCount,
+				ValSamples:   valCount,
+				Imgsz:        640,
+			})
+		}
+		if poseStatus == "trained" {
+			trainingAppendTargetHistory(root, "pose", poseStatus, poseDurationMs, runtimeOpts, TrainingHistoryEntry{
+				Epochs:       trainingDetectorEpochs(),
+				TrainSamples: poseTrainCount,
+				ValSamples:   poseValCount,
+				Imgsz:        640,
+			})
+		}
+		if videoMAEStatus == "trained" {
+			trainingAppendTargetHistory(root, "videomae", videoMAEStatus, videoMAEDurationMs, runtimeOpts, TrainingHistoryEntry{
+				Epochs:       trainingVideoMAEEpochs(),
+				TrainSamples: videoMAETrainCount,
+				ValSamples:   videoMAEValCount,
+				Imgsz:        trainingVideoMAEFrameSize,
+			})
+		}
 	}
 
 	trainingSetJobStatus(func(s *TrainingJobStatus) {
@@ -4338,20 +4635,22 @@ func trainingStatusHandler(w http.ResponseWriter, r *http.Request) {
 			"predictsClothing":    false,
 			"predictsBoxes":       false,
 
-			"feedbackCount": feedbackCount,
-			"eligibleCount": videoMAEEligibleCount,
-			"trainCount":    videoMAETrainCount,
-			"valCount":      videoMAEValCount,
-			"requiredTrain": minVideoMAETrainCount,
-			"requiredVal":   minVideoMAEValCount,
-			"requiredCount": minVideoMAETrainCount,
-			"datasetReady":  videoMAEDatasetReady,
-			"manifest":      videoMAEManifest,
-			"dataReady":     videoMAEDataReady,
-			"modelReady":    videoMAEModel.EffectiveExists,
-			"modelExists":   videoMAEModel.EffectiveExists,
-			"modelPath":     videoMAEModel.EffectivePath,
-			"modelSource":   videoMAEModel.Source,
+			"feedbackCount":      feedbackCount,
+			"eligibleCount":      videoMAEEligibleCount,
+			"trainCount":         videoMAETrainCount,
+			"valCount":           videoMAEValCount,
+			"requiredTrain":      minVideoMAETrainCount,
+			"requiredVal":        minVideoMAEValCount,
+			"requiredCount":      minVideoMAETrainCount,
+			"datasetReady":       videoMAEDatasetReady,
+			"manifest":           videoMAEManifest,
+			"dataReady":          videoMAEDataReady,
+			"modelReady":         videoMAEModel.EffectiveExists,
+			"modelExists":        videoMAEModel.EffectiveExists,
+			"modelPath":          videoMAEModel.EffectivePath,
+			"modelSource":        videoMAEModel.Source,
+			"trainedModelExists": videoMAEModel.TrainedExists,
+			"trainedModelPath":   videoMAEModel.BestPath,
 		},
 
 		"pipeline": map[string]any{
@@ -4464,16 +4763,24 @@ func trainingReadModelInfoFor(root string, kind string) *TrainingModelInfo {
 }
 
 type TrainingHistoryEntry struct {
-	TrainedAt    string  `json:"trainedAt,omitempty"`
-	TrainedAtMs  int64   `json:"trainedAtMs,omitempty"`
-	DurationMs   int64   `json:"durationMs,omitempty"`
-	Epochs       int     `json:"epochs,omitempty"`
-	TrainSamples int     `json:"trainSamples,omitempty"`
-	ValSamples   int     `json:"valSamples,omitempty"`
-	Imgsz        int     `json:"imgsz,omitempty"`
-	Device       string  `json:"device,omitempty"`
-	MAP50        float64 `json:"map50,omitempty"`
-	MAP5095      float64 `json:"map5095,omitempty"`
+	TrainedAt       string  `json:"trainedAt,omitempty"`
+	TrainedAtMs     int64   `json:"trainedAtMs,omitempty"`
+	Target          string  `json:"target,omitempty"`
+	Status          string  `json:"status,omitempty"`
+	DurationMs      int64   `json:"durationMs,omitempty"`
+	Epochs          int     `json:"epochs,omitempty"`
+	TrainSamples    int     `json:"trainSamples,omitempty"`
+	ValSamples      int     `json:"valSamples,omitempty"`
+	Imgsz           int     `json:"imgsz,omitempty"`
+	Device          string  `json:"device,omitempty"`
+	MAP50           float64 `json:"map50,omitempty"`
+	MAP5095         float64 `json:"map5095,omitempty"`
+	PerformanceMode string  `json:"performanceMode,omitempty"`
+	CPUCoreCount    int     `json:"cpuCoreCount,omitempty"`
+	CPUThreads      int     `json:"cpuThreads,omitempty"`
+	Workers         int     `json:"workers,omitempty"`
+	YoloBatchSize   int     `json:"yoloBatchSize,omitempty"`
+	LowPriority     bool    `json:"lowPriority,omitempty"`
 }
 
 type TrainingHistoryResponse struct {
@@ -4485,31 +4792,81 @@ func trainingHistoryPath(root string) string {
 	return filepath.Join(root, "detector", "training_history.jsonl")
 }
 
-// trainingAppendRunHistory hängt nach einem erfolgreichen Trainingslauf einen
-// Verlaufseintrag an (Datum, mAP, Samples, Epochen, Dauer).
-func trainingAppendRunHistory(root string) {
-	info := trainingReadModelInfo(root)
-	if info == nil {
-		return
+// trainingAppendTargetHistory haengt nach einem erfolgreichen Trainingsziel einen
+// Verlaufseintrag an. Alte History-Zeilen ohne Target bleiben weiterhin lesbar.
+func trainingHistoryKindForTarget(target string) string {
+	switch strings.ToLower(strings.TrimSpace(target)) {
+	case "pose":
+		return "pose"
+	case "videomae", "video_mae", "scene":
+		return "videomae"
+	default:
+		return "detector"
 	}
+}
+
+func trainingAppendTargetHistory(root string, target string, status string, durationMs int64, runtimeOpts trainingRuntimeOptions, fallback TrainingHistoryEntry) {
+	kind := trainingHistoryKindForTarget(target)
+	info := trainingReadModelInfoFor(root, kind)
+	now := time.Now().UTC()
 
 	entry := TrainingHistoryEntry{
-		TrainedAt:    info.TrainedAt,
-		TrainedAtMs:  info.TrainedAtMs,
-		Epochs:       info.Epochs,
-		TrainSamples: info.TrainSamples,
-		ValSamples:   info.ValSamples,
-		Imgsz:        info.Imgsz,
-		Device:       info.Device,
-		MAP50:        info.MAP50,
-		MAP5095:      info.MAP5095,
+		TrainedAt:       now.Format(time.RFC3339),
+		TrainedAtMs:     now.UnixMilli(),
+		Target:          kind,
+		Status:          strings.TrimSpace(status),
+		DurationMs:      durationMs,
+		Epochs:          fallback.Epochs,
+		TrainSamples:    fallback.TrainSamples,
+		ValSamples:      fallback.ValSamples,
+		Imgsz:           fallback.Imgsz,
+		Device:          strings.TrimSpace(fallback.Device),
+		MAP50:           fallback.MAP50,
+		MAP5095:         fallback.MAP5095,
+		PerformanceMode: runtimeOpts.PerformanceMode,
+		CPUCoreCount:    runtimeOpts.CPUCoreCount,
+		CPUThreads:      runtimeOpts.CPUThreads,
+		Workers:         runtimeOpts.Workers,
+		YoloBatchSize:   runtimeOpts.YoloBatchSize,
+		LowPriority:     runtimeOpts.LowPriority,
 	}
 
-	// Dauer aus der Startzeit des aktuellen Jobs ableiten.
-	job := trainingGetJobStatus()
-	if startedAt, err := time.Parse(time.RFC3339, strings.TrimSpace(job.StartedAt)); err == nil {
-		if ms := time.Now().UTC().Sub(startedAt).Milliseconds(); ms > 0 {
-			entry.DurationMs = ms
+	if info != nil {
+		if strings.TrimSpace(info.TrainedAt) != "" {
+			entry.TrainedAt = info.TrainedAt
+		}
+		if info.TrainedAtMs > 0 {
+			entry.TrainedAtMs = info.TrainedAtMs
+		}
+		if info.Epochs > 0 {
+			entry.Epochs = info.Epochs
+		}
+		if info.TrainSamples > 0 {
+			entry.TrainSamples = info.TrainSamples
+		}
+		if info.ValSamples > 0 {
+			entry.ValSamples = info.ValSamples
+		}
+		if info.Imgsz > 0 {
+			entry.Imgsz = info.Imgsz
+		}
+		if strings.TrimSpace(info.Device) != "" {
+			entry.Device = strings.TrimSpace(info.Device)
+		}
+		if info.MAP50 > 0 {
+			entry.MAP50 = info.MAP50
+		}
+		if info.MAP5095 > 0 {
+			entry.MAP5095 = info.MAP5095
+		}
+	}
+
+	if entry.DurationMs <= 0 {
+		job := trainingGetJobStatus()
+		if startedAt, err := time.Parse(time.RFC3339, strings.TrimSpace(job.StartedAt)); err == nil {
+			if ms := now.Sub(startedAt).Milliseconds(); ms > 0 {
+				entry.DurationMs = ms
+			}
 		}
 	}
 

backend/training_videomae.go

@@ -55,6 +55,21 @@ func trainingVideoMAEBatchSize() int {
 	return n
 }
 
+func trainingVideoMAEEarlyStoppingPatience() int {
+	epochs := trainingVideoMAEEpochs()
+	patience := epochs / 3
+	if patience < 2 {
+		patience = 2
+	}
+	if patience > 8 {
+		patience = 8
+	}
+	if patience > epochs {
+		patience = epochs
+	}
+	return patience
+}
+
 type trainingVideoMAEManifestEntry struct {
 	SampleID   string  `json:"sampleId"`
 	Split      string  `json:"split"`