PyTorch Mobile Android Example

Quickstart

1. Build Docker image pytorch_android_example

sh ./build_docker_image.sh

Docker image is based on ubuntu and contains all dependencies to build pytorch for linux and pytorch for android, including Vulkan and NNAPI backends.

Result:

└─ $ docker image ls
REPOSITORY                TAG       IMAGE ID       CREATED        SIZE
pytorch_android_example   latest    8cc67601109c   21 hours ago   12.4GB

2. Run model train and building pytorch android

sh ./train_model_and_build_pytorch_in_docker.sh

Result:

└─ $ ls model/output/
mnist-nnapi-ops.yaml	mnist-ops-all.yaml	mnist-quant.pt		mnist-vulkan.pt		mnist.ptl
mnist-nnapi.pt		mnist-ops.yaml		mnist-quant.ptl		mnist-vulkan.ptl
mnist-nnapi.ptl		mnist-quant-ops.yaml	mnist-vulkan-ops.yaml	mnist.pt
└─ $ ls android/application/app/aars/
pytorch_android.aar

3. Install android application on connected android device

sh ./install_android_app.sh

Result: Application MNIST installed.

Content

docker/Dockerfile - docker container definition

model/mnist.py - MNIST model definition and training code. Serializes the model and quantized variant of it for CPU, Vulkan and NNAPI.

pytorch-patches - temporary changes in pytorch that will be eliminated after fixing functionality in pytorch master

sh ./model/build_local_pytorch_for_mnist.sh - runs build of pytorch android, contains specification of android abis for the build.