a FastAPI backend for catalog discovery, job orchestration, logs, results, GPU status, and weight management;
a React/Vite frontend for launching jobs and inspecting progress;
a unified weights/ area with a generated manifest for .pt, .pth, .onnx, and .engine assets.

Layout

Seg_Data_Server_Net/
  backend/            FastAPI API, job runner, module wrappers
  frontend/           React + Vite operator UI
  scripts/            helper scripts for running services and syncing weights
  weights/            copied model weights and manifest.json

Quick Start

cd Seg_Data_Server_Net
cp .env.example .env

# Backend. The deployment env is seg_smp so the API and most task wrappers
# share the same segmentation dependency stack. MMSeg jobs default to the
# separate SEG_MMSEG_CONDA_ENV because full mmcv wheels must match torch/CUDA.
conda run -n seg_smp uvicorn app.main:app --app-dir backend --host 0.0.0.0 --port 8010

# Frontend.
cd frontend
npm install
npm run dev -- --host 0.0.0.0

Open the Vite URL shown in the terminal. The frontend expects the backend at http://localhost:8010 by default.

The web UI includes a dataset bench for creating upload workspaces, uploading images/labels/masks, and jumping into the existing rename, PNG conversion, resize, pair-check, label rebuild, transparent overlay, stitch, and video-frame jobs. Selecting an uploaded dataset fills task JSON with its images, labels, and masks directories. Segmentation previews, YOLO heatmaps, and loss/metric artifacts are grouped on the results dashboard, and YOLO-style results.csv files are parsed into lightweight training curves.

The coverage panel calls GET /api/coverage and verifies that the user-facing scripts from the existing Seg/ workspace are mapped to web jobs. MMSeg vendored internals, docs, build outputs, converters, and config templates are classified as supporting artifacts rather than direct web actions.

The same panel can run POST /api/acceptance/smoke, a lightweight live smoke that creates an upload dataset, uploads a label, downloads it through the artifact API, runs a mock job, checks SSE log streaming, and executes one legacy image/label overlay job on tiny generated PNGs. It also runs model family readiness checks: a SegModel/SMP forward pass, a YOLO segmentation prediction on a tiny image, MMSeg config parsing, and local MMSeg pretrained weight discovery. MMSeg full-model readiness is validated in SEG_MMSEG_CONDA_ENV by importing mmcv._ext and building a local MMSeg EncoderDecoder from the existing config tree.

For stronger runtime proof, POST /api/acceptance/deep runs minimal training loops for the three model families: one SegModel optimizer step, one YOLO segmentation epoch on a synthetic 64x64 dataset, and one MMSeg optimizer step through the full mmcv._ext runtime. The latest report is available from GET /api/acceptance/deep/latest and is surfaced in the coverage panel.

Current seg_smp uses mmcv-lite because no torch 2.6/cu124 full mmcv wheel is available on this machine and nvcc is not installed for source builds. A dedicated seg_mmcv environment is used for MMSeg tasks and has torch 2.1.2+cu121, mmcv 2.1.0, mmsegmentation 1.2.2, and NumPy 1.26. If rebuilding the environment, keep these versions aligned:

conda create -n seg_mmcv python=3.10 -y
conda run -n seg_mmcv python -m pip install -U pip
conda run -n seg_mmcv python -m pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu121
conda run -n seg_mmcv python -m pip install mmengine==0.10.7 mmsegmentation==1.2.2 'mmcv==2.1.0' -f https://download.openmmlab.com/mmcv/dist/cu121/torch2.1/index.html
conda run -n seg_mmcv python -m pip install 'numpy<2' 'opencv-python<4.12' ftfy regex matplotlib pandas scikit-learn scipy seaborn tqdm tensorboard

Weight Sync

The current workspace contains tens of GB of pretrained and trained weights. They are copied into weights/files/<original-relative-path> and indexed in weights/manifest.json.

cd Seg_Data_Server_Net
python scripts/sync_weights.py --mode copy --hash

Weights must remain local to the deployment machine. Do not push .pt, .pth, .onnx, .engine, or weights/files/ into Gitea. The repository stores only code, weights/manifest.json, and helper scripts. Before pushing, run:

scripts/check_no_weight_git.sh

If a deployment machine needs weights, run the sync command locally on that machine after cloning the code.

Job Types

The backend exposes all current Seg capabilities as job types. Examples:

dataset.rename, dataset.resize, dataset.pair, dataset.rebuild_labels, dataset.stack, dataset.stitch, dataset.video_frames, dataset.yolo_check_pairs, dataset.yolo_stack, dataset.yolo_txt_sort, dataset.yolo_convert_png, dataset.yolo_resize
segmodel.train, segmodel.batch_train, segmodel.predict, segmodel.batch_predict, segmodel.flops, segmodel.params_flops, segmodel.benchmark, segmodel.raw_mask_check
yolo.train, yolo.batch_train, yolo.predict, yolo.batch_predict, yolo.heatmap, yolo.compare, yolo.raw_mask_check, yolo.video_visible
mmseg.generate_data, mmseg.generate_alg, mmseg.train, mmseg.metrics, mmseg.flops_fps, mmseg.draw, mmseg.extract_loss_miou
visual.train, visual.inference, visual.fps, visual.yolo11_heatmap_v1, visual.yolo11_heatmap_v2, visual.deal_labels
analysis.all, system.backup, mock.echo

Use GET /api/catalog to inspect supported models, algorithms, datasets, and task types discovered from the existing Seg/ workspace. Use GET /api/coverage to inspect script-to-task coverage and task buildability. Use GET /api/results/curves to inspect parsed training curves discovered from YOLO, SegModel, MMSeg, visual-tool, and analysis output directories.

Agents

Run the local evaluation and validation agents before publishing changes:

PYTHONPATH=backend conda run -n seg_smp python scripts/run_agents.py --build

The validation agent checks catalog coverage, the seg_smp task env, the seg_mmcv MMSeg env, GPU visibility, no-weight Git safety, backend tests, frontend build, and live backend/frontend endpoints when the services are running. With live validation enabled it also runs the lightweight acceptance smoke above. By default it also runs the deep training acceptance; set SEG_VALIDATE_DEEP=0 when a quick non-training validation pass is needed.