README.md

NanoDet MNN Demo

This fold provides NanoDet inference code using Alibaba's MNN framework. Most of the implements in this fold are same as demo_ncnn.

Install MNN

Python library

Just run:

pip install MNN

C++ library

Please follow the official document to build MNN engine.

Convert model

1. Export ONNX model

    python tools/export_onnx.py --cfg_path ${CONFIG_PATH} --model_path ${PYTORCH_MODEL_PATH}

2. Convert to MNN

   python -m MNN.tools.mnnconvert -f ONNX --modelFile sim.onnx --MNNModel nanodet.mnn

It should be note that the input size does not have to be fixed, it can be any integer multiple of strides, since NanoDet is anchor free. We can adapt the shape of dummy_input in ./tools/export_onnx.py to get ONNX and MNN models with different input sizes.

Here are converted model Download Link.

Build

For C++ code, replace libMNN.so under ./mnn/lib with the one you just compiled, modify OpenCV path at CMake file, and run

mkdir build && cd build
cmake ..
make

Note that a flag at main.cpp is used to control whether to show the detection result or save it into a fold.

#define __SAVE_RESULT__ // if defined save drawed results to ../results, else show it in windows

Run

Python

The multi-backend python demo is still working in progress.

C++

C++ inference interface is same with NCNN code, to detect images in a fold, run:

./nanodet-mnn "1" "../imgs/*.jpg"

For speed benchmark

./nanodet-mnn "3" "0"

Custom model

If you want to use custom model, please make sure the hyperparameters in nanodet_mnn.h are the same with your training config file.

int input_size[2] = {416, 416}; // input height and width
int num_class = 80; // number of classes. 80 for COCO
int reg_max = 7; // `reg_max` set in the training config. Default: 7.
std::vector<int> strides = { 8, 16, 32, 64 }; // strides of the multi-level feature.

Reference

Ultra-Light-Fast-Generic-Face-Detector-1MB

ONNX Simplifier

NanoDet NCNN

MNN

Example results