TorchServe KeyError

Discussed in https://github.com/open-mmlab/mmdetection/discussions/9200

<div type='discussions-op-text'>

^{Originally posted by austinmw October 31, 2022} Hi, I used MMDetection 3.0.0rc2 to train a Deformable DETR model and a Faster RCNN model, and then I used the mmdet2torchserve tool to create two model.mar. I tried testing local deployment with my own TorchServe Dockerfile, but I get errors with both of the TorchServe archives:

KeyError: "class DeformableDETR in mmdet/models/detectors/deformable_detr.py: 'DetDataPreprocessor is not in the model registry. Please check whether the value of DetDataPreprocessor is correct or it was registered as expected.

and

KeyError: "class FasterRCNN in mmdet/models/detectors/faster_rcnn.py: 'DetDataPreprocessor is not in the model registry. Please check whether the value of DetDataPreprocessor is correct or it was registered as expected.

My Dockerfile looks like this:

# Use Sagemaker PyTorch inference container as base image
ARG REGION
FROM 763104351884.dkr.ecr.${REGION}.amazonaws.com/pytorch-inference:1.12.1-gpu-py38-cu113-ubuntu20.04-sagemaker

# Install MMEngine, MMCV, MMDetection
WORKDIR /opt/ml/code
RUN pip install cython nvgpu && \
    pip install -U pip openmim torch-model-archiver && \
    pip install git+https://github.com/open-mmlab/mmengine.git@v0.2.0 && \
    mim install "mmcv>=2.0.0rc2" -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.12.0/index.html && \
    pip install git+https://github.com/open-mmlab/mmdetection.git@v3.0.0rc2 && \
    rm -rf /var/lib/apt/lists/*

COPY config.properties /home/model-server/config.properties

WORKDIR /
ENTRYPOINT ["python", "/usr/local/bin/dockerd-entrypoint.py"]
CMD ["torchserve", "--start", "--ts-config", "/home/model-server/config.properties", "--model-store", "/opt/ml/model"]

I placed the model.mars into /opt/ml/model.

Anyone know how to resolve this KeyError? Is mmdet_handler.py compatible with v3.0.0rc2?

</div>

[RunningLeon]

@lvhan028 to take look. MMDeploy should already be compatible with mmdet 3.x in mmdeploy 1.x. We just need to have an entry point in mmdet.

Hi, torchserve is not included in mmdeploy. Deployment for torchserve would be kept in mmdet. As for the error, it’s because of the BC-breaking of mmdet2.x and mmdet3.x. A fixing of this would come later.

[austinmw]

Hi, thanks for response, okay I’ll see if I can get it working for PR, otherwise can wait till official 3.0 release