/MNN-yolov3

Primary LanguagePython

MNN-yolov3

Introduction

MNN demo of YOLOv3(converted from Stronger-Yolo).

Quick Start

  1. Install MNN following the corresponding guide.
  2. Setup an environment following Stronger-Yolo.
  3. run v3/pb.py to convert tensorflow checkpoint into portable model.
  4. (optional) Fold constants using TF tools. (Recommended by MNN.) 
    bazel-bin/tensorflow/tools/graph_transforms/transform_graph --transforms=fold_constants(ignore_errors=true)
  5. Converting model (remember to build convert tools first) 
    cd {MNN dir}/tools/converter/build/
./MNNConvert -f TF --modelFile {MNN-yolov3 project dir}/v3/port/yolov3_opti_fc.pb --MNNModel yolo_opti_fc.mnn --bizCode MNN
  6. Copy MNN-demo/yolo.cpp in to {MNN dir}/demo/exec and Modify {MNN dir}/demo/exec/CmakeLists.txt like MNN-demo/CmakeLists.txt.
  7. Run cpp execution.

Quantitative Analysis

Note:
1.Inference time is tested using MNN official Test Tool with scorethreshold 0.2 And 0.7849 is the original tensorflow result.
2.All MAP results are evaluated using the first 300 testing images in order to save time.
3.-quant model is quantized using official MNN tool. The poor inference speed is due to arm-specified optimization. Check this.

Model InputSize Thread Inference(ms) Params MAP(VOC)
Yolov3 544 4/6/8 2962/2975/3043 27M 0.7803(0.7849)
Yolov3 320 4/6/8 828/668/821 27M 0.7127(0.7249)
Yolov3-quant 320 4/6/8 1636/1445/1472 7.7M 0.7082(0.7249)

Qualitative Comparison

  • Testing Result in Tensorflow(top) and MNN(down).
    Result of Tensorflow Result of Tensorflow

Important Notes during model converting

  1. Following this issue to remove/replace some op.
  2. Remove condition op which is related to BatchNormalization and training Flag. Otherwise it will cause MNN converting failure. 
    Identity's input node num. != 1
  3. Replace v3/model/head/build_nework with build_nework_MNN, which replaces tf.shape with static inputshape and replace 
    [:, tf.newaxis] -> tf.expand_dims // currently strided_slice op is not very well supported in MNN.

TODO

  • Speed analyse.
  • Model Quantization.
  • Android Support.

Reference

stronger-yolo

MNN

NCNN