How bichenwuucb/squeezedet works
Project Type and Purpose
• This is an open-source, TensorFlow-based object-detection framework.
• It provides:
• Data ingestion (KITTI / Pascal VOC)
• Model definitions (SqueezeDet, SqueezeDet+, VGG16-ConvDet, ResNet50-ConvDet)
• Training and evaluation pipelines
• Demo and utility scripts
Break Down the Main Components
A. Data Layer
• Raw dataset (images + labels under data/KITTI, Pascal VOC)
• random_split_train_val.py for train/val split
• Dataset classes (src/dataset/imdb.py, kitti.py, pascal_voc.py, voc_eval.py)
• External C++ KITTI evaluation tool (src/dataset/kitti-eval/cpp)
B. Configuration Layer
• Central config module (src/config/config.py)
• Network-specific configs (kitti_res50_config.py, kitti_squeezeDet_config.py, etc.)
C. Model Layer
• Network definitions in src/nets/ (squeezeDet.py, squeezeDetPlus.py, vgg16_convDet.py, resnet50_convDet.py)
• Weight conversion utility (caffemodel2pkl.py)
D. Training Pipeline
• train.py orchestrates data loading, model instantiation, loss computation, optimizer steps, checkpointing
• scripts/train.sh wraps train.py with flags (-net, ‑train_dir, ‑gpu)
E. Evaluation Pipeline
• eval.py loads a checkpoint, runs forward pass on train/val sets, computes metrics
• scripts/eval.sh wraps eval.py (-net, ‑eval_dir, ‑image_set, ‑gpu)
• C++ kitti-eval for official metrics
F. Demo & Visualization
• demo.py for running inference on images/videos
• TensorBoard integration via log events in train_dir/eval_dir
G. Utilities & Helpers
• src/utils/util.py for common helpers (logging, file I/O)
• random_split_train_val.py, shell scripts, Makefile
Relationships and Data Flows
1. Data Ingestion:
Raw images/labels → Dataset classes → Batches of tensors
2. Configuration:
Experiment parameters from config/*.py → Model instantiation + Training/Eval scripts
3. Training Loop:
Batches → Network → Loss → Optimizer → Checkpoints → Logs
4. Evaluation Loop:
Checkpoints → Network → Predictions → C++ kitti-eval → AP metrics → Logs
5. Visualization & Monitoring:
Logs → TensorBoard
6. Demo Inference:
Input images → Trained Checkpoint → Network → Output detections
Architectural Patterns & Design Principles
• Modularity: separation of concerns among config, dataset, model, train/eval, utils
• Plugin-style networks: “-net” flag selects network and corresponding config
• Reusable dataset interface (imdb) for different sources
• Thin wrapper scripts for reproducible experiments
• External tool integration (C++ eval, TensorBoard)