How lutao2021/pidray works
Project Type and Purpose
• PIDray is an open‐source research codebase built as an extension of the mmdetection framework. Its goal is to train, evaluate, and benchmark object detection/segmentation models on the PIDray X-ray prohibited items dataset.
High‐Level Components
• CLI Layer (entrypoints):
• shell scripts (tools/dist_train.sh, tools/dist_test.sh)
• Python entrypoints (tools/train.py, tools/test.py)
• Configuration Manager:
• configs/… (custom cascade_mask_rcnn_r101_with_R0R1.py, etc.)
• mmdet/.mim/configs/_base_ (dataset, model, schedule, runtime defaults)
• Dataset & Pipeline:
• mmdet/datasets (dataset builders, wrappers, pipelines for loading + augmentations)
• Custom PIDray dataset integration via config overrides
• Model Builder:
• mmdet/models (backbones, necks, heads, detectors, ROI‐heads, seg‐heads)
• mmdet/models/utils for building blocks
• Core Engine & Utilities:
• mmdet/apis/train.py & inference.py (training/testing loops)
• mmdet/core (anchor generation, bbox coder, sampling, losses, evaluation hooks)
• mmdet/core/hook (logging, checkpointing, EMA, lr schedulers)
• mmdet/utils & mmcv (logging, config parsing, dist utilities)
• Evaluation & Export:
• mmdet/apis/test.py
• mmdet/core/evaluation (mean AP, IoU, panoptic utils)
• model export tools (onnx, TorchServe, TensorRT) in mmdet/tools/deployment
• External Services & Storage:
• GPU cluster / distributed training via PyTorch DDP & Slurm scripts
• Dataset storage (local disk or network drives)
• Model weights & logs (file system or remote buckets)
Relationships & Data Flow
1. User invokes shell script → Python entrypoint
2. Entrypoint loads experiment config via Config Manager
3. Config Manager builds: Dataset objects + DataLoader + Model + Optimizer + Hooks
4. Training loop (Engine) pulls batches from DataLoader → forward pass through Model → losses computed via Core modules → backward pass → parameter update
5. Hooks handle logging, checkpoint save, lr scheduling
6. After training or on‐the‐fly, Evaluation module computes metrics on validation/test splits
7. Results (model weights, metrics, inference outputs) written to disk
Architectural Patterns & Design Principles
• Modular separation: CLI, Configs, Data, Model, Core, Hooks, Deployment
• Factory/Builder pattern: “build_dataset”, “build_detector”, “build_optimizer” in mmdet/models/builder.py
• Hook system: decouples training loop from logging, checkpointing, EMA, lr schedule
• Layered directory structure: clear demarcation between configs, code, demos, tests