phase cv 0

gpu/server.py

@@ -52,74 +52,25 @@ def _gpu_log(job_id: str, log_level: str, stage: str, level: str, msg: str):
         emit_log(job_id, stage, level, msg, log_level=log_level)
 
 
-# --- Request/Response models ---
+# --- Request/Response models (generated from core/schema/models/inference.py) ---
 
-class DetectRequest(BaseModel):
-    image: str
-    model: str | None = None
-    confidence: float | None = None
-    target_classes: list[str] | None = None
-
-
-class BBox(BaseModel):
-    x: int
-    y: int
-    w: int
-    h: int
-    confidence: float
-    label: str
-
-
-class DetectResponse(BaseModel):
-    detections: list[BBox]
-
-
-class OCRRequest(BaseModel):
-    image: str
-    languages: list[str] | None = None
-
-
-class OCRTextResult(BaseModel):
-    text: str
-    confidence: float
-    bbox: list[int]
-
-
-class OCRResponse(BaseModel):
-    results: list[OCRTextResult]
-
-
-class PreprocessRequest(BaseModel):
-    image: str
-    binarize: bool = False
-    deskew: bool = False
-    contrast: bool = True
-
-
-class PreprocessResponse(BaseModel):
-    image: str  # base64 JPEG of processed image
-
-
-class VLMRequest(BaseModel):
-    image: str
-    prompt: str
-    model: str | None = None
-
-
-class VLMResponse(BaseModel):
-    brand: str
-    confidence: float
-    reasoning: str
-
-
-class ConfigUpdate(BaseModel):
-    device: str | None = None
-    yolo_model: str | None = None
-    yolo_confidence: float | None = None
-    vram_budget_mb: int | None = None
-    strategy: str | None = None
-    ocr_languages: list[str] | None = None
-    ocr_min_confidence: float | None = None
+from models.inference_contract import (
+    AnalyzeRegionsDebugResponse,
+    AnalyzeRegionsRequest,
+    AnalyzeRegionsResponse,
+    BBox,
+    ConfigUpdate,
+    DetectRequest,
+    DetectResponse,
+    OCRRequest,
+    OCRResponse,
+    OCRTextResult,
+    PreprocessRequest,
+    PreprocessResponse,
+    RegionBox,
+    VLMRequest,
+    VLMResponse,
+)
 
 
 # --- App ---
@@ -281,6 +232,84 @@ def vlm(req: VLMRequest, request: Request):
     return VLMResponse(**result)
 
 
+@app.post("/detect_edges", response_model=AnalyzeRegionsResponse)
+def detect_edges_endpoint(req: AnalyzeRegionsRequest, request: Request):
+    job_id, log_level = _job_ctx(request)
+
+    try:
+        image = _decode_image(req.image)
+        h, w = image.shape[:2]
+    except Exception as e:
+        raise HTTPException(status_code=400, detail=f"Bad image: {e}")
+
+    try:
+        t0 = time.monotonic()
+        from models.cv.edges import detect_edges
+
+        edge_regions = detect_edges(
+            image,
+            canny_low=req.edge_canny_low,
+            canny_high=req.edge_canny_high,
+            hough_threshold=req.edge_hough_threshold,
+            hough_min_length=req.edge_hough_min_length,
+            hough_max_gap=req.edge_hough_max_gap,
+            pair_max_distance=req.edge_pair_max_distance,
+            pair_min_distance=req.edge_pair_min_distance,
+        )
+        infer_ms = (time.monotonic() - t0) * 1000
+
+        _gpu_log(job_id, log_level, "GPU:CV", "DEBUG",
+                 f"Edge analysis {w}x{h}: {infer_ms:.0f}ms → {len(edge_regions)} regions")
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Region analysis failed: {e}")
+
+    boxes = [RegionBox(**r) for r in edge_regions]
+    return AnalyzeRegionsResponse(regions=boxes)
+
+
+@app.post("/detect_edges/debug", response_model=AnalyzeRegionsDebugResponse)
+def detect_edges_debug_endpoint(req: AnalyzeRegionsRequest, request: Request):
+    job_id, log_level = _job_ctx(request)
+
+    try:
+        image = _decode_image(req.image)
+        h, w = image.shape[:2]
+    except Exception as e:
+        raise HTTPException(status_code=400, detail=f"Bad image: {e}")
+
+    try:
+        t0 = time.monotonic()
+        from models.cv.edges import detect_edges_debug
+
+        result = detect_edges_debug(
+            image,
+            canny_low=req.edge_canny_low,
+            canny_high=req.edge_canny_high,
+            hough_threshold=req.edge_hough_threshold,
+            hough_min_length=req.edge_hough_min_length,
+            hough_max_gap=req.edge_hough_max_gap,
+            pair_max_distance=req.edge_pair_max_distance,
+            pair_min_distance=req.edge_pair_min_distance,
+        )
+        infer_ms = (time.monotonic() - t0) * 1000
+
+        _gpu_log(job_id, log_level, "GPU:CV", "DEBUG",
+                 f"Edge debug {w}x{h}: {infer_ms:.0f}ms → {len(result['regions'])} regions, "
+                 f"{result['horizontal_count']} horizontals, {result['pair_count']} pairs")
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Region debug analysis failed: {e}")
+
+    boxes = [RegionBox(**r) for r in result["regions"]]
+    response = AnalyzeRegionsDebugResponse(
+        regions=boxes,
+        edge_overlay_b64=result["edge_overlay_b64"],
+        lines_overlay_b64=result["lines_overlay_b64"],
+        horizontal_count=result["horizontal_count"],
+        pair_count=result["pair_count"],
+    )
+    return response
+
+
 if __name__ == "__main__":
     import uvicorn