Integration of Libtorch in DeepDetect for image classification

[Bycob]

• Libtorch backend
• Libtorch image input connector
• Image classification using pytorch traced models
• Downloading and compiling libtorch automatically during DeepDetect compilation
• Unit tests on libtorch backend

Building instructions:

cmake .. -DUSE_TORCH=ON

[beniz]

This code only works for traced model, right ?

[Bycob]

Yes, only traced models are currently supported.

[beniz]

FYI, a build on a CPU only machine with cmake -DUSE_CAFFE=OFF -DUSE_TORCH=ON .. fails with the error below:

[ 32%] Built target pytorch
[ 92%] Built target ddetect
make[2]: *** No rule to make target 'pytorch/src/pytorch/torch/lib/libc10_cuda.so', needed by 'main/dede'.  Stop.
CMakeFiles/Makefile2:177: recipe for target 'main/CMakeFiles/dede.dir/all' failed

However, the build works with cmake -DUSE_CAFFE=OFF -DUSE_TORCH=ON -DUSE_CPU_ONLY=ON ...

I believe that if there's no CUDA automatically detected, USE_CPU_ONLY should be forced ON automatically.

[beniz]

FYI, to work on CPU at least, it is required to install libiomp5(This is a transitional package on Ubuntu 18.04). So we should add it to the README I believe. However, Pytorch appers to come with its own version of that lib, in pytorch/src/pytorch/torch/lib/libiomp5.so.

I get this error when starting dede:

./dede: error while loading shared libraries: libiomp5.so: cannot open shared object file: No such file or directory

[Bycob]

Fixed everything mentioned above, except the libiomp loading error.

[Bycob]

Script to trace resnet and other models.

Requirements:

pip3 install --user pytorch torchvision

[Bycob]

<!> A segfault may occur when running a model with CUDA

[beniz]

@BynaryCobweb Confirmed that the final linking is fixed on CPU builds!

[Bycob]

Example: Image classification

• Download the model

wget https://www.deepdetect.com/dd/examples/torch/resnet50_torch.tar.gz
tar xvf resnet50_torch.tar.gz

• Run dede
• Start the service

curl -X PUT "http://localhost:8080/services/torch_resnet" -d '{
    "description": "image classification service",
    "mllib": "torch",
    "model": {
        "repository": "./resnet50_torch/"
    },
    "parameters": {
        "input": {
            "connector": "image"
        }
    },
    "type": "supervised"
}
'

• Run predict to guess the class of an image

curl -X POST "http://localhost:8080/predict" -d '{
    "service": "torch_resnet",
    "parameters": {
        "input": {
            "width":224,
            "height":224,
            "rgb":true,
            "std":255.0
        },
        "output": {
            "best":3
        }
    },
    "data":["./resnet50_torch/cat.jpg"]
}
'

[Bycob]

• squashed all commits in one