This project compiles libtorch to WebAssembly using Emscripten. Includes an example that loads a pre-trained ResNet50 model that classifies images in the browser.
Exporting the example model requires PyTorch and Torchvision. Setting up a new conda environment is recommended:
conda create -n libtorch-wasm python=3.11
conda activate libtorch-wasm
conda install pytorch torchvision cpuonly -c pytorchTo export the pre-trained example ResNet50 model, run the following command in the root of the project directory:
python export-model.pyThis will generate a resnet50.pt file in the same directory.
The project uses Docker. The primary WebAssembly build is located in the root directory. To compile it, run the following command:
docker build -t resnet-torch-wasm .Note: This is slow.
Once the WebAssembly image is built, you can start a server to host the example page:
docker run -it --rm -p 8080:8080 resnet-torch-wasmAfter running this command, the server will be accessible at http://localhost:8080, where you can see the example page in action.
For testing purposes you may also compile a Ubuntu native binary. This will compile an executable linked against a pre-built libtorch which is quick and practical for prototyping. To compile the native build, run the following command:
docker build -f native/Dockerfile -t resnet-torch-native .The native image will just run the example code and exit. Use the following command:
docker run -it --rm resnet-torch-native- The WebAssembly build is slow, possible improvements would be getting around the limitations of
-Wl,--whole-archiveforlibtorch.aand improving caching in the CMake build. - JIT optimizations like
torch.jit.freezedecrease the model size and improve performance but seem to depend on MKL-DNN which is not being compiled by default. - Most of the included optimization libraries are disabled since they complicate the build process so don't expect native speeds.
This project is licensed under the MIT License.