If you want to run on a GPU you need to make sure your version of Apache MXNet Incubating contains the GPU bindings. Depending on your version of CUDA you can do this by running the following:
where ${CUDA_VERSION} can be 75 (7.5), 80 (8.0), 90 (9.0), or 91 (9.1).
Usage
To train a denoising autoencoder, turn on --source-noise-train with detailed noise options (--source-noise-insertion, --source-noise-insertion-vocab, --source-noise-deletion, --source-noise-permutation). Please put the same training data for both source and target sides and also the same validation data for both sides. Optionally, you can also switch on --source-noise-validation to evaluate your models on a noisy validation set during the training. Example:
Denoising with a trained model can be done with sockeye.translate module in the same way as translating an input sentence. You can use all other modules provided by Sockeye on denoising autoencoder, e.g. sharding the training data (sockeye.prepare_data) or model averaging (sockeye.average). Please refer to the Sockeye documentation for details.