How linzaer/ultra-light-fast-generic-face-detector-1mb works
Project Type & Purpose
• This is a machine-learning framework and toolkit for ultra-lightweight face detection. It combines:
• A core PyTorch-based training and model definition pipeline (vision package, train.py)
• A data-preparation stage (VOC conversion of WIDER FACE)
• Export/conversion pipelines to multiple inference frameworks (ONNX, Caffe, MNN, NCNN, TensorFlow, TFLite, Paddle)
• Inference and evaluation code in C++ and Python for each backend
High-Level File-Structure Analysis
• Root scripts: train.py, train-version-*.sh, detect_imgs.py, convert_to_onnx.py, cal_flops.py, etc.
• vision/: PyTorch model definitions (mb_tiny, mb_tiny_RFB, SSD head, losses, preprocessing)
• data/: VOC-format dataset generation scripts
• models/: converted weights (ONNX, PyTorch .pth) & labels
• Framework directories (each containing conversion code, backend C++/Python inference, sample main.cpp or .py):
– caffe/onnx2caffe + ultra_face_caffe_inference.py + prototxt/.caffemodel
– MNN/ (C++), MNN/python/ (Python) + .mnn models
– ncnn/ + .param/.bin + C++ inference
– tf/ + TensorFlow saved_model + conversion scripts
– tflite/ + TFLiteFaceDetector.py + .tflite models
– paddle/ + Paddle training & inference modules
– opencv_dnn/ + cv_dnn_ultraface.cpp
• widerface_evaluate/: evaluation scripts
Key Architectural Patterns
• Pipeline architecture: Data → Training → Model Export → Framework-specific Converters → Inference Runtimes → Evaluation
• Plugin-like backends: each inference framework lives in its own directory, implements a common inference interface
• Artifacts as first-class citizens: model files (.pth, .onnx, .mnn, .param/.bin, .tflite) flow between steps
A. Data Preparation
– Input: Raw WIDER FACE dataset
– Script: data/wider_face_2_voc_add_landmark.py → VOC dataset
B. Model Definition & Training
– vision/nn/… + train.py + train-version-*.sh
– Output: PyTorch .pth models (version-slim, version-RFB)
C. ONNX Export
– convert_to_onnx.py → model.onnx
D. Conversion Pipelines (for each backend)
– ONNX → Caffe (onnx2caffe) → .prototxt/.caffemodel
– ONNX → MNN (MNN/src + converter) → .mnn
– ONNX → NCNN (ncnn tools) → .param/.bin
– ONNX → TensorFlow (tf/convert_tensorflow.py) → SavedModel
– ONNX → TFLite (tflite converter) → .tflite
– PyTorch → Paddle (paddle/train.py & paddle/vision)
E. Inference & Demo Applications
– Python demos: detect_imgs_onnx.py, run_video_face_detect.py, MNN/python/ultraface_py_mnn.py, etc.
– C++ demos: MNN/src/main.cpp, ncnn/src/main.cpp, opencv_dnn/cv_dnn_ultraface.cpp
F. Evaluation
– widerface_evaluate/evaluation_on_widerface.py → accuracy metrics
Diagramming Details
• Diamonds or parallelograms for scripts (data prep, conversion tools)
• Layer Separation
• Detail Level
• Entry points: main.cpp or .py demo scripts