Supply a requirements file with the model-archive.

version

torch                1.5.1
torch-model-archiver 0.2.0b20200731
torchserve           0.1.1
torchtext            0.6.0
torchvision          0.6.0a0+35d732a

export command

 torch-model-archiver --model-name craft --version 1.0 --model-file ./craft.py  --serialized-file ./model/craft_mlt_25k.pth  --handler  ./craft_handler.py -r requirements.txt

requirements.txt

-i https://pypi.tuna.tsinghua.edu.cn/simple
pillow
opencv_python==3.4.1.15

docerk command

docker run --rm -it \                                                                                                              
           -p 8080:8080 -p 8081:8081 \
           --name torch_serve_cpu \
           -v (pwd)/model_store:/home/model-server/model-store \
           pytorch/torchserve \
           torchserve --start --ncs --model-store model-store --models craft.mar

logs

2020-07-31 08:22:18,079 [INFO ] W-9003-craft_1.0-stdout org.pytorch.serve.wlm.WorkerLifeCycle -     import cv2
2020-07-31 08:22:18,079 [INFO ] W-9003-craft_1.0-stdout org.pytorch.serve.wlm.WorkerLifeCycle - ModuleNotFoundError: No module named 'cv2'

[IamMohitM]

Actually, it worked at some later point with a different installation (tensorflow) in the requirements file. Haven’t tested with opencv as shown above. Used the --requirements-file option instead of -r

[festinais]

@IamMohitM can you try with the command: --requirements-file instead of -r. For more details: https://github.com/pytorch/serve/tree/master/model-archiver#torch-model-archiver-command-line-interface.

I had the same issue and I tried with --extra-files and -r, it didn’t work. With --requirements-file it works!