OV DeformableConvolution support for DEKR

README.md

OV DeformableConvolution support for DEKR

Pytorch layer documentation

OpenVINO layer documentation

Follow these steps to get OpenVINO IR:

• You can find all the changes needed to support the DeformableConvolution layer here.

1. Create .onnx model with python3 tools/convert.py --cfg path/to/file.yaml TEST.MODEL_FILE path/to/file.yaml:
2. Create OpenVINO IR with OpenVINO master branch:

python3 <INSTALL_DIR>/openvino/tools/mo/openvino/tools/mo/mo.py --input_model model.onnx

3. Open model.xml file and replace all the values of bilinear_interpolation_pad parameter with true:

<layer id="..." name="..." type="DeformableConvolution" ...>
    <data auto_pad="explicit" bilinear_interpolation_pad="true" ...>

4. Check the result with python3 compare.py.

Expected difference:

1. For COCO:

python3 tools/convert.py \
    --cfg experiments/coco/inference_demo_coco.yaml \
    TEST.MODEL_FILE model/pose_coco/pose_dekr_hrnetw32_coco.pth

the result is:

Heatmap reference range: [-0.016255718, 0.047695834]
Offset reference range: [-50.128445, 53.03814]

Heatmap diff: 8.2841143e-07
Offset diff: 0.0043911934

2. For CrowdPose:

python3 tools/convert.py \
    --cfg experiments/crowdpose/inference_demo_crowdpose.yaml \
    TEST.MODEL_FILE model/pose_crowdpose/pose_dekr_hrnetw32_crowdpose.pth

the result is:

Heatmap reference range: [-0.0047741896, 0.028738622]
Offset reference range: [-70.21522, 52.001522]

Heatmap diff: 7.105991e-07
Offset diff: 0.0013594627