Open4

mmdetection + UniverseNet

$ git clone https://github.com/open-mmlab/mmcv.git && cd mmcv
$ MMCV_WITH_OPS=1 pip3 install -e .
$ sudo pip3 install mmdet

ONNXのエクスポーターの動作確認

$ git clone https://github.com/shinya7y/UniverseNet.git && UniverseNet
$ python3 tools/pytorch2onnx.py -h
usage: pytorch2onnx.py
    [-h]
    [--input-img INPUT_IMG]
    [--show]
    [--output-file OUTPUT_FILE]
    [--opset-version OPSET_VERSION]
    [--verify]
    [--shape SHAPE [SHAPE ...]]
    [--mean MEAN [MEAN ...]]
    [--std STD [STD ...]]
    config checkpoint

Convert MMDetection models to ONNX

positional arguments:
  config                test config file path
  checkpoint            checkpoint file

optional arguments:
  -h, --help            show this help message and exit
  --input-img INPUT_IMG
                        Images for input
  --show                show onnx graph
  --output-file OUTPUT_FILE
  --opset-version OPSET_VERSION
  --verify              verify the onnx model output against pytorch output
  --shape SHAPE [SHAPE ...]
                        input image size
  --mean MEAN [MEAN ...]
                        mean value used for preprocess input data
  --std STD [STD ...]   variance value used for preprocess input data
$ wget https://github.com/shinya7y/UniverseNet/releases/download/20.12/universenet50_2008s_fp16_4x16_micro_mstrain_128_256_2x_coco_20201111_epoch_24-2655e5d3.pth && \
  wget https://github.com/shinya7y/UniverseNet/releases/download/20.12/universenet50_2008s_fp16_4x16_mini_mstrain_320_640_2x_coco_20201110_epoch_24-d2655d05.pth

$ sed -i "s/mmcv_maximum_version = '1.2'/mmcv_maximum_version = '9.0.0'/g" mmdet/__init__.py
$ sed -i "s/one_img = torch.from_numpy(one_img).unsqueeze(0).float()/one_img = torch.from_numpy(one_img).unsqueeze(0).float().cpu()/g" tools/pytorch2onnx.py
$ python3 tools/pytorch2onnx.py \
  --shape 224 224 \
  --input-img demo/demo.jpg \
  --output-file universenet50_2008s_224x224_float32.onnx \
  configs/universenet/universenet50_2008s_fp16_4x16_micro_mstrain_128_256_2x_coco.py \
  universenet50_2008s_fp16_4x16_micro_mstrain_128_256_2x_coco_20201111_epoch_24-2655e5d3.pth
ログインするとコメントできます