Skip to content

Latest commit

 

History

History
executable file
·
127 lines (70 loc) · 3.48 KB

README.md

File metadata and controls

executable file
·
127 lines (70 loc) · 3.48 KB
Raw

mobilenetv3_centernet

there is a pytorch version, trained with mobilenetv2, it is more simple.

introduction

This is a tensorflow implement mobilenetv3-centernet framework, which can be easily deployeed on Android(MNN) and IOS(CoreML) mobile devices, end to end.

Purpose: Light detection algorithms that work on mobile devices is widely used, such as face detection. So there is an easy project contains model training and model converter.

** contact me if u have question [email protected] **

pretrained model , and preformance

mscoco

no test time augmentation.

model input_size map [email protected] [email protected]
mbv3-large-0.75-modified_head 512x512 0.251 0.423 0.258

requirment

  • tensorflow 1.14

  • tensorpack 0.9.9 (for data provider)

  • opencv

  • python 3.6

  • MNNConverter

  • coremltools

useage

MSCOCO

train

  1. download mscoco data, then run python prepare_coco_data.py --mscocodir ./mscoco

  2. download pretrained model from mbv3-large0.75 relese it in the current dir.

  3. then, modify in config=mb3_config in train_config.py, then run:

    python train.py

    and if u want to check the data when training, u could set vis in confifs/mscoco/mbv3_config.py as True

  4. After training, freeze the model as .pb by

    python tools/auto_freeze.py --pretrained_mobile ./model/yourmodel.ckpt

    it will produce a detector.pb

evaluation

python model_eval/custome_eval.py [--model [TRAINED_MODEL]] [--annFile [cocostyle annFile]]
                          [--imgDir [the images dir]] [--is_show [show the result]]

python model_eval/custome_eval.py --model model/detector.pb
                                --annFile ../mscoco/annotations/instances_val2017.json
                                --imgDir ../mscoco/val2017
                                --is_show 1

ps, no test time augmentation is used.

finetune

  1. download the trained model, modify the config config.MODEL.pretrained_model='yourmodel.ckpt', and set config.MODEL.continue_train=True
  2. python train.py

visualization

if u get a trained model and dont need to work on mobile device, run python tools/auto_freeze.py, it will read the checkpoint file in ./model, and produce detector.pb, then

python visualization/vis.py

u can check th code in visualization to make it runable, it's simple.

model convert for mobile device

I have carefully processed the postprocess, and it can works within the model, so it could be deployed end to end.

4.1 MNN

+ 4.1.1 convert model

    just use the MNN converter, for example:
    `./MNNConvert -f TF --modelFile detector.pb --MNNModel centernet.mnn --bizCode biz  --fp16 1`

+ 4.1.2 visualization with mnn python wrapper

    `python visualization/vis_with_mnn.py --mnn_model centernet.mnn --imgDir 'your image dir'`

4.2 coreml

+ 4.2.1 convert

    `python tools/converter_to_coreml.py`

+ 4.2.2 visualization with coreml python wrapper

    `python visualization/vis_with_coreml.py --coreml_model centernet.mlmodel --imgDir 'your image dir'`

ps, if you want to do quantization, please reffer to the official doc, it is easy.

TODO:

  • Android project.