feat: 打通全栈标注闭环、异步拆帧与模型状态

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。
This commit is contained in:
2026-05-01 13:29:14 +08:00
parent 4d65c37c73
commit f020ff3b4f
78 changed files with 7089 additions and 456 deletions

View File

@@ -1,10 +1,6 @@
"""Media upload and parsing endpoints."""
import logging
import os
import shutil
import subprocess
import tempfile
from pathlib import Path
from typing import List, Optional
@@ -12,13 +8,12 @@ from fastapi import APIRouter, Depends, File, Form, HTTPException, UploadFile, s
from sqlalchemy.orm import Session
from database import get_db
from minio_client import upload_file, get_presigned_url, download_file
from models import Project, Frame
from schemas import FrameOut
from services.frame_parser import (
parse_video, parse_dicom, upload_frames_to_minio,
extract_thumbnail, get_video_fps,
)
from minio_client import upload_file, get_presigned_url
from models import ProcessingTask, Project
from progress_events import publish_task_progress_event
from schemas import ProcessingTaskOut
from statuses import PROJECT_STATUS_PARSING, PROJECT_STATUS_PENDING, TASK_STATUS_QUEUED
from worker_tasks import parse_project_media
logger = logging.getLogger(__name__)
router = APIRouter(prefix="/api/media", tags=["Media"])
@@ -79,7 +74,7 @@ async def upload_media(
project = Project(
name=file.filename,
description="Auto-created from upload",
status="pending",
status=PROJECT_STATUS_PENDING,
video_path=object_name,
source_type="video",
)
@@ -135,7 +130,7 @@ async def upload_dicom_batch(
project = Project(
name=first_name,
description=f"DICOM series with {len(files)} files",
status="pending",
status=PROJECT_STATUS_PENDING,
source_type="dicom",
)
db.add(project)
@@ -168,19 +163,18 @@ async def upload_dicom_batch(
@router.post(
"/parse",
status_code=status.HTTP_202_ACCEPTED,
response_model=ProcessingTaskOut,
summary="Trigger frame extraction",
)
def parse_media(
project_id: int,
source_type: Optional[str] = None,
db: Session = Depends(get_db),
) -> dict:
"""Trigger frame extraction for a project's uploaded media.
) -> ProcessingTask:
"""Create a background task for media frame extraction.
* video: uses FFmpeg or OpenCV fallback, extracts thumbnail.
* dicom: uses pydicom to read DCM frames.
Extracted frames are uploaded to MinIO and registered in the database.
The Celery worker performs the heavy FFmpeg/OpenCV/pydicom work and
updates the persisted task record as it progresses.
"""
project = db.query(Project).filter(Project.id == project_id).first()
if not project:
@@ -190,100 +184,24 @@ def parse_media(
raise HTTPException(status_code=400, detail="Project has no media uploaded")
effective_source = source_type or project.source_type or "video"
parse_fps = project.parse_fps or 30.0
tmp_dir = tempfile.mkdtemp(prefix=f"seg_parse_{project_id}_")
output_dir = os.path.join(tmp_dir, "frames")
os.makedirs(output_dir, exist_ok=True)
try:
if effective_source == "dicom":
# Download all dicom files from MinIO
dcm_dir = os.path.join(tmp_dir, "dcm")
os.makedirs(dcm_dir, exist_ok=True)
from minio_client import get_minio_client, BUCKET_NAME
client = get_minio_client()
prefix = project.video_path
objects = list(client.list_objects(BUCKET_NAME, prefix=prefix, recursive=True))
for obj in objects:
if obj.object_name.lower().endswith(".dcm"):
data = download_file(obj.object_name)
local_dcm = os.path.join(dcm_dir, os.path.basename(obj.object_name))
with open(local_dcm, "wb") as f:
f.write(data)
frame_files = parse_dicom(dcm_dir, output_dir)
else:
# Video: download and parse
media_bytes = download_file(project.video_path)
local_path = os.path.join(tmp_dir, Path(project.video_path).name)
with open(local_path, "wb") as f:
f.write(media_bytes)
frame_files, original_fps = parse_video(local_path, output_dir, fps=int(parse_fps))
project.original_fps = original_fps
# Extract thumbnail from first frame
thumbnail_path = os.path.join(tmp_dir, "thumbnail.jpg")
try:
extract_thumbnail(local_path, thumbnail_path)
with open(thumbnail_path, "rb") as f:
thumb_data = f.read()
thumb_object = f"projects/{project_id}/thumbnail.jpg"
upload_file(thumb_object, thumb_data, content_type="image/jpeg", length=len(thumb_data))
project.thumbnail_url = thumb_object
logger.info("Uploaded thumbnail for project_id=%s", project_id)
except Exception as exc: # noqa: BLE001
logger.warning("Thumbnail extraction failed: %s", exc)
except Exception as exc: # noqa: BLE001
logger.error("Frame extraction failed: %s", exc)
shutil.rmtree(tmp_dir, ignore_errors=True)
raise HTTPException(status_code=500, detail="Frame extraction failed") from exc
# Upload frames to MinIO
try:
object_names = upload_frames_to_minio(frame_files, project_id)
except Exception as exc: # noqa: BLE001
logger.error("Frame upload failed: %s", exc)
shutil.rmtree(tmp_dir, ignore_errors=True)
raise HTTPException(status_code=500, detail="Frame upload to storage failed") from exc
# Register frames in DB
frames_out = []
for idx, obj_name in enumerate(object_names):
local_frame = frame_files[idx]
try:
import cv2
img = cv2.imread(local_frame)
h, w = img.shape[:2] if img is not None else (None, None)
except Exception: # noqa: BLE001
h, w = None, None
frame = Frame(
project_id=project_id,
frame_index=idx,
image_url=obj_name,
width=w,
height=h,
)
db.add(frame)
frames_out.append(frame)
task = ProcessingTask(
task_type=f"parse_{effective_source}",
status=TASK_STATUS_QUEUED,
progress=0,
message="解析任务已入队",
project_id=project_id,
payload={"source_type": effective_source},
)
project.status = PROJECT_STATUS_PARSING
db.add(task)
db.commit()
for f in frames_out:
db.refresh(f)
db.refresh(task)
publish_task_progress_event(task)
# Cleanup temp files
shutil.rmtree(tmp_dir, ignore_errors=True)
project.status = "ready"
async_result = parse_project_media.delay(task.id)
task.celery_task_id = async_result.id
db.commit()
db.refresh(task)
logger.info("Parsed %d frames for project_id=%s", len(frames_out), project_id)
return {
"project_id": project_id,
"frames_extracted": len(frames_out),
"status": "ready",
"message": "Frame extraction completed successfully.",
}
logger.info("Queued parse task id=%s project_id=%s celery_id=%s", task.id, project_id, async_result.id)
return task