2026-04-08 19:32:23 +08:00
|
|
|
|
# FishServer 核心仓库（已瘦身）
|
|
|
|
|
|
|
|
|
|
|
|
本目录面向 **FishMeasure（SVO2 称重/点云）**、**FishAction（MP4 行为 X3D）** 与 **fish_api（FastAPI 网关）** 的部署与运行，已去掉训练数据、历史推理产物、旧版 Django 前端和 SlowFast 训练栈等大体积非运行时内容。
|
|
|
|
|
|
|
|
|
|
|
|
## 目录结构
|
|
|
|
|
|
|
|
|
|
|
|
| 路径 | 说明 |
|
|
|
|
|
|
|------|------|
|
|
|
|
|
|
| `fish_api/` | FastAPI：`uv sync` 后 `uv run uvicorn app.main:app`，见其中 `README.md` |
|
|
|
|
|
|
| `FishMeasure/` | 双目链路：`predict_weigth_from_svo2.py`、`fish_video_weight_evaluation.py`、`weight_estimator/`、`pointcloud_classifier/` 等 |
|
|
|
|
|
|
| `FishAction/` | 行为推断：`predict_video_x3d_3class.py`、`train_pytorchvideo_x3d.py`、`checkpoints/`（X3D） |
|
|
|
|
|
|
| `packaging/` | **单一 Conda 环境**：网关 + 两条算法依赖定义，见 [`packaging/README.md`](packaging/README.md) |
|
2026-04-10 10:30:01 +08:00
|
|
|
|
| `scripts/start_fresh.sh` | 在已激活的 `fishserver` 环境中清空缓存后启动 uvicorn |
|
|
|
|
|
|
| `scripts/start_no_fresh.sh` | 保留 SQLite 与推理缓存启动 uvicorn |
|
2026-04-08 19:32:23 +08:00
|
|
|
|
|
|
|
|
|
|
### 一键打包成「单环境」运行（推荐服务器）
|
|
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
|
bash packaging/bootstrap_fishserver.sh
|
|
|
|
|
|
conda activate fishserver
|
|
|
|
|
|
bash packaging/patch_cuda_torch.sh # Linux + NVIDIA 时建议
|
|
|
|
|
|
# 再按 packaging/README.md 安装 ZED SDK 与 pyzed
|
2026-04-10 10:30:01 +08:00
|
|
|
|
PORT=8001 bash scripts/start_fresh.sh
|
2026-04-08 19:32:23 +08:00
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
|
|
多 Conda 环境、分别设置 `PYTHON_FISH_MEASURE` / `PYTHON_FISH_ACTION` 的方式仍支持，见 `fish_api/README.md`。
|
|
|
|
|
|
|
|
|
|
|
|
## 已删除内容（需训练/旧功能时可从备份找回）
|
|
|
|
|
|
|
|
|
|
|
|
- `FishMeasure/output_weight_estimator/`、`output-yolo-sam/`：推理输出
|
|
|
|
|
|
- `FishMeasure/datasets/`：训练集
|
|
|
|
|
|
- `FishMeasure/project_jiuzhou01/`：九州 Django + 前端工程
|
|
|
|
|
|
- `FishMeasure/measure/`、`detect_refbox/dataset`、`detect_refbox/runs`：独立实验数据与跑次
|
|
|
|
|
|
- `FishMeasure/runs/predict`、`runs/segment`：旧预测/分割输出
|
|
|
|
|
|
- `FishMeasure/utils/data/`：工具附带大数据
|
|
|
|
|
|
- `FishMeasure/weight_estimator/runs/` 中除 `dgcnn_20260312_171043/` 外的历史训练目录
|
|
|
|
|
|
- `FishAction/slowfast/`：SlowFast 训练代码（当前网关仅走 PyTorchVideo X3D）
|
|
|
|
|
|
- 根目录重复 `yolo*.pt`、`fish_video_weight_evaluation__v1.py` 等
|
|
|
|
|
|
|
|
|
|
|
|
## 仍占空间的大文件（运行时一般需要）
|
|
|
|
|
|
|
|
|
|
|
|
- **`FishMeasure/sam_vit_h_4b8939.pth`**（约 2.4GB）：SAM `vit_h`。若未放置，可从 [Segment Anything 官方权重](https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth) 下载到 `FishMeasure/` 下同名文件。
|
|
|
|
|
|
- **`FishMeasure/runs/train/fish_detection_20251127_104658/weights/best.pt`**：YOLO 检测，与 `fish_api` 默认环境变量一致。
|
|
|
|
|
|
- **`FishMeasure/weight_estimator/runs/dgcnn_20260312_171043/best.pt`**：DGCNN 体重估计。
|
|
|
|
|
|
- **`FishAction/checkpoints/ptv_x3d_m/checkpoint_best.pt`**：行为分类（已删除其它实验版 checkpoint 目录以省空间；若需恢复请从备份拿回）。
|
|
|
|
|
|
|
|
|
|
|
|
## Git 与体积
|
|
|
|
|
|
|
|
|
|
|
|
`FishMeasure/`、`FishAction/` 内嵌各自 `.git`，其中 FishMeasure 历史对象可能很大（含 LFS）。**仅部署运行时**可用 rsync 排除版本库：
|
|
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
|
rsync -avz --exclude '.git' --exclude 'fish_api/.venv' ...
|
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
|
|
## 接口说明
|
|
|
|
|
|
|
|
|
|
|
|
- 业务 API 契约可参考仓库根目录 `接口文档.docx`（若仍保留）。
|
|
|
|
|
|
- HTTP 网关路径以 `fish_api/app/routers/` 为准。
|
|
|
|
|
|
|
|
|
|
|
|
## 同步到服务器示例
|
|
|
|
|
|
|
|
|
|
|
|
```bash
|
|
|
|
|
|
rsync -avz --delete \
|
|
|
|
|
|
--exclude '.venv' --exclude 'fish_api/.venv' --exclude '__pycache__' \
|
|
|
|
|
|
./ ubuntu@192.168.10.93:/home/ubuntu/projects/FishServer/
|
|
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
|
|
（按实际 IP、用户与路径修改；`--delete` 慎用，会删远端多余文件。）
|
