admin
f020ff3b4f
后端能力： - 新增 Celery app、worker task、ProcessingTask 模型、/api/tasks 查询接口和 media_task_runner，将 /api/media/parse 改为创建后台任务并由 worker 执行 FFmpeg/OpenCV/pydicom 拆帧。 - 新增 Redis 进度事件模块和 FastAPI Redis pub/sub 订阅，将 worker 任务进度广播到 /ws/progress；Dashboard 后端概览接口改为聚合 projects/frames/annotations/templates/processing_tasks。 - 统一项目状态为 pending/parsing/ready/error，新增共享 status 常量，并让前端兼容归一化旧状态值。 - 扩展 AI 后端：新增 SAM registry、SAM2 真实运行状态、SAM3 状态检测与文本语义推理适配入口，以及 /api/ai/models/status GPU/模型状态接口。 - 补齐标注保存/更新/删除、COCO/PNG mask 导出相关后端契约和模板 mapping_rules 打包/解包行为。 前端能力： - 新增运行时 API/WS 地址推导配置，前端 API 封装对齐 FastAPI 路由、字段映射、任务轮询、标注归档、导出下载和 AI 预测响应转换。 - Dashboard 改为读取 /api/dashboard/overview，并订阅 WebSocket progress/complete/error/status 更新解析队列和实时流转记录。 - 项目库导入视频/DICOM 后创建项目、上传媒体、触发异步解析并刷新真实项目列表。 - 工作区加载真实帧、无帧时触发解析任务、回显已保存标注、保存未归档 mask、更新 dirty mask、清空当前帧后端标注、导出 COCO JSON。 - Canvas 支持当前帧点/框提示调用后端 AI、渲染推理/已保存 mask、应用模板分类并维护保存状态计数；时间轴按项目 fps 播放。 - AI 页面新增 SAM2/SAM3 模型选择，预测请求携带 model；侧边栏和工作区新增真实 GPU/SAM 状态徽标。 - 模板库和本体面板接入真实模板 CRUD、分类编辑、拖拽排序、JSON 导入、默认腹腔镜分类和本地自定义分类选择。 测试与文档： - 新增 Vitest 配置、前端测试 setup、API/config/websocket/store/组件测试，覆盖登录、项目库、Dashboard、Canvas、工作区、模型状态、时间轴、本体和模板库。 - 新增 pytest 后端测试夹具和 auth/projects/templates/media/AI/export/dashboard/tasks/progress 测试，使用 SQLite、fake MinIO、fake SAM registry 和 Redis monkeypatch 隔离外部服务。 - 新增 doc/ 文档结构，冻结当前需求、设计、接口契约、测试计划、前端逐元素审计、实现地图和后续实施计划，并同步更新 README 与 AGENTS。 验证： - conda run -n seg_server pytest backend/tests：27 passed。 - npm run test:run：54 passed。 - npm run lint、npm run build、compileall、git diff --check 均通过；Vite 仅提示大 chunk 警告。
298 lines
11 KiB
Python
298 lines
11 KiB
Python
"""SAM 2 engine wrapper with lazy loading and explicit runtime status."""
|
|
|
|
import logging
|
|
import os
|
|
from typing import Optional
|
|
|
|
import numpy as np
|
|
|
|
from config import settings
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# Attempt to import PyTorch and SAM 2; fall back to stubs if unavailable.
|
|
# ---------------------------------------------------------------------------
|
|
try:
|
|
import torch
|
|
|
|
TORCH_AVAILABLE = True
|
|
except Exception as exc: # noqa: BLE001
|
|
TORCH_AVAILABLE = False
|
|
torch = None # type: ignore[assignment]
|
|
logger.warning("PyTorch import failed (%s). SAM2 will be unavailable.", exc)
|
|
|
|
try:
|
|
from sam2.build_sam import build_sam2
|
|
from sam2.sam2_image_predictor import SAM2ImagePredictor
|
|
|
|
SAM2_AVAILABLE = True
|
|
logger.info("SAM2 library imported successfully.")
|
|
except Exception as exc: # noqa: BLE001
|
|
SAM2_AVAILABLE = False
|
|
logger.warning("SAM2 import failed (%s). Using stub engine.", exc)
|
|
|
|
|
|
class SAM2Engine:
|
|
"""Lazy-loaded SAM 2 inference engine."""
|
|
|
|
def __init__(self) -> None:
|
|
self._predictor: Optional[SAM2ImagePredictor] = None
|
|
self._model_loaded = False
|
|
self._loaded_device: str | None = None
|
|
self._last_error: str | None = None
|
|
|
|
# -----------------------------------------------------------------------
|
|
# Internal helpers
|
|
# -----------------------------------------------------------------------
|
|
def _load_model(self) -> None:
|
|
"""Load the SAM 2 model and predictor on first use."""
|
|
if self._model_loaded:
|
|
return
|
|
|
|
if not TORCH_AVAILABLE:
|
|
self._last_error = "PyTorch is not installed."
|
|
logger.warning("PyTorch not available; skipping SAM2 model load.")
|
|
self._model_loaded = True
|
|
return
|
|
|
|
if not SAM2_AVAILABLE:
|
|
self._last_error = "sam2 package is not installed."
|
|
logger.warning("SAM2 not available; skipping model load.")
|
|
self._model_loaded = True
|
|
return
|
|
|
|
if not os.path.isfile(settings.sam_model_path):
|
|
self._last_error = f"SAM2 checkpoint not found: {settings.sam_model_path}"
|
|
logger.error("SAM checkpoint not found at %s", settings.sam_model_path)
|
|
self._model_loaded = True
|
|
return
|
|
|
|
try:
|
|
device = self._best_device()
|
|
model = build_sam2(
|
|
settings.sam_model_config,
|
|
settings.sam_model_path,
|
|
device=device,
|
|
)
|
|
self._predictor = SAM2ImagePredictor(model)
|
|
self._model_loaded = True
|
|
self._loaded_device = device
|
|
self._last_error = None
|
|
logger.info("SAM 2 model loaded from %s on %s", settings.sam_model_path, device)
|
|
except Exception as exc: # noqa: BLE001
|
|
self._last_error = str(exc)
|
|
logger.error("Failed to load SAM 2 model: %s", exc)
|
|
self._model_loaded = True # Prevent repeated load attempts
|
|
|
|
def _best_device(self) -> str:
|
|
if TORCH_AVAILABLE and torch is not None and torch.cuda.is_available():
|
|
return "cuda"
|
|
return "cpu"
|
|
|
|
def _ensure_ready(self) -> bool:
|
|
"""Ensure the model is loaded; return whether it is usable."""
|
|
self._load_model()
|
|
return SAM2_AVAILABLE and self._predictor is not None
|
|
|
|
def status(self) -> dict:
|
|
"""Return lightweight, real runtime status without forcing model load."""
|
|
checkpoint_exists = os.path.isfile(settings.sam_model_path)
|
|
device = self._loaded_device or self._best_device()
|
|
available = bool(TORCH_AVAILABLE and SAM2_AVAILABLE and checkpoint_exists)
|
|
if self._predictor is not None:
|
|
message = "SAM 2 model loaded and ready."
|
|
elif available:
|
|
message = "SAM 2 dependencies and checkpoint are present; model will load on first inference."
|
|
else:
|
|
missing = []
|
|
if not TORCH_AVAILABLE:
|
|
missing.append("PyTorch")
|
|
if not SAM2_AVAILABLE:
|
|
missing.append("sam2 package")
|
|
if not checkpoint_exists:
|
|
missing.append("checkpoint")
|
|
message = f"SAM 2 unavailable: missing {', '.join(missing)}."
|
|
if self._last_error and not self._predictor:
|
|
message = self._last_error
|
|
return {
|
|
"id": "sam2",
|
|
"label": "SAM 2",
|
|
"available": available,
|
|
"loaded": self._predictor is not None,
|
|
"device": device,
|
|
"supports": ["point", "box", "auto"],
|
|
"message": message,
|
|
"package_available": SAM2_AVAILABLE,
|
|
"checkpoint_exists": checkpoint_exists,
|
|
"checkpoint_path": settings.sam_model_path,
|
|
"python_ok": True,
|
|
"torch_ok": TORCH_AVAILABLE,
|
|
"cuda_required": False,
|
|
}
|
|
|
|
# -----------------------------------------------------------------------
|
|
# Public API
|
|
# -----------------------------------------------------------------------
|
|
def predict_points(
|
|
self,
|
|
image: np.ndarray,
|
|
points: list[list[float]],
|
|
labels: list[int],
|
|
) -> tuple[list[list[list[float]]], list[float]]:
|
|
"""Run point-prompt segmentation.
|
|
|
|
Args:
|
|
image: HWC numpy array (uint8).
|
|
points: List of [x, y] normalized coordinates (0-1).
|
|
labels: 1 for foreground, 0 for background.
|
|
|
|
Returns:
|
|
Tuple of (polygons, scores).
|
|
"""
|
|
if not self._ensure_ready():
|
|
logger.warning("SAM2 not ready; returning dummy masks.")
|
|
return self._dummy_polygons(image.shape[1], image.shape[0]), [0.5]
|
|
|
|
try:
|
|
h, w = image.shape[:2]
|
|
pts = np.array([[p[0] * w, p[1] * h] for p in points], dtype=np.float32)
|
|
lbls = np.array(labels, dtype=np.int32)
|
|
|
|
with torch.inference_mode(): # type: ignore[name-defined]
|
|
self._predictor.set_image(image)
|
|
masks, scores, _ = self._predictor.predict(
|
|
point_coords=pts,
|
|
point_labels=lbls,
|
|
multimask_output=True,
|
|
)
|
|
|
|
polygons = []
|
|
for m in masks:
|
|
poly = self._mask_to_polygon(m)
|
|
if poly:
|
|
polygons.append(poly)
|
|
|
|
return polygons, scores.tolist()
|
|
except Exception as exc: # noqa: BLE001
|
|
logger.error("SAM2 point prediction failed: %s", exc)
|
|
return self._dummy_polygons(image.shape[1], image.shape[0]), [0.5]
|
|
|
|
def predict_box(
|
|
self,
|
|
image: np.ndarray,
|
|
box: list[float],
|
|
) -> tuple[list[list[list[float]]], list[float]]:
|
|
"""Run box-prompt segmentation.
|
|
|
|
Args:
|
|
image: HWC numpy array (uint8).
|
|
box: [x1, y1, x2, y2] normalized coordinates.
|
|
|
|
Returns:
|
|
Tuple of (polygons, scores).
|
|
"""
|
|
if not self._ensure_ready():
|
|
logger.warning("SAM2 not ready; returning dummy masks.")
|
|
return self._dummy_polygons(image.shape[1], image.shape[0]), [0.5]
|
|
|
|
try:
|
|
h, w = image.shape[:2]
|
|
bbox = np.array(
|
|
[box[0] * w, box[1] * h, box[2] * w, box[3] * h],
|
|
dtype=np.float32,
|
|
)
|
|
|
|
with torch.inference_mode(): # type: ignore[name-defined]
|
|
self._predictor.set_image(image)
|
|
masks, scores, _ = self._predictor.predict(
|
|
box=bbox[None, :],
|
|
multimask_output=False,
|
|
)
|
|
|
|
polygons = []
|
|
for m in masks:
|
|
poly = self._mask_to_polygon(m)
|
|
if poly:
|
|
polygons.append(poly)
|
|
|
|
return polygons, scores.tolist()
|
|
except Exception as exc: # noqa: BLE001
|
|
logger.error("SAM2 box prediction failed: %s", exc)
|
|
return self._dummy_polygons(image.shape[1], image.shape[0]), [0.5]
|
|
|
|
def predict_auto(self, image: np.ndarray) -> tuple[list[list[list[float]]], list[float]]:
|
|
"""Run automatic mask generation (grid of points).
|
|
|
|
Args:
|
|
image: HWC numpy array (uint8).
|
|
|
|
Returns:
|
|
Tuple of (polygons, scores).
|
|
"""
|
|
if not self._ensure_ready():
|
|
logger.warning("SAM2 not ready; returning dummy masks.")
|
|
return self._dummy_polygons(image.shape[1], image.shape[0]), [0.5]
|
|
|
|
try:
|
|
with torch.inference_mode(): # type: ignore[name-defined]
|
|
self._predictor.set_image(image)
|
|
# Generate a uniform 16x16 grid of point prompts
|
|
h, w = image.shape[:2]
|
|
grid = np.mgrid[0:1:17j, 0:1:17j].reshape(2, -1).T
|
|
pts = grid * np.array([w, h])
|
|
lbls = np.ones(pts.shape[0], dtype=np.int32)
|
|
|
|
masks, scores, _ = self._predictor.predict(
|
|
point_coords=pts,
|
|
point_labels=lbls,
|
|
multimask_output=True,
|
|
)
|
|
|
|
polygons = []
|
|
for m in masks[:3]: # Limit to top 3 masks
|
|
poly = self._mask_to_polygon(m)
|
|
if poly:
|
|
polygons.append(poly)
|
|
|
|
return polygons, scores[:3].tolist()
|
|
except Exception as exc: # noqa: BLE001
|
|
logger.error("SAM2 auto prediction failed: %s", exc)
|
|
return self._dummy_polygons(image.shape[1], image.shape[0]), [0.5]
|
|
|
|
# -----------------------------------------------------------------------
|
|
# Helpers
|
|
# -----------------------------------------------------------------------
|
|
@staticmethod
|
|
def _mask_to_polygon(mask: np.ndarray) -> list[list[float]]:
|
|
"""Convert a binary mask to a normalized polygon."""
|
|
import cv2
|
|
|
|
if mask.dtype != np.uint8:
|
|
mask = (mask > 0).astype(np.uint8)
|
|
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
|
h, w = mask.shape[:2]
|
|
largest = []
|
|
for cnt in contours:
|
|
if len(cnt) > len(largest):
|
|
largest = cnt
|
|
if len(largest) < 3:
|
|
return []
|
|
return [[float(pt[0][0]) / w, float(pt[0][1]) / h] for pt in largest]
|
|
|
|
@staticmethod
|
|
def _dummy_polygons(w: int, h: int) -> list[list[list[float]]]:
|
|
"""Return a dummy rectangle polygon for fallback mode."""
|
|
return [
|
|
[
|
|
[0.25, 0.25],
|
|
[0.75, 0.25],
|
|
[0.75, 0.75],
|
|
[0.25, 0.75],
|
|
]
|
|
]
|
|
|
|
|
|
# Singleton instance
|
|
sam_engine = SAM2Engine()
|