The PyTorch tutorial NLP FROM SCRATCH: TRANSLATION WITH A SEQUENCE TO SEQUENCE NETWORK AND ATTENTION is an easy-to-follow tutorial that has step-by-step instructions on how to train sequence to sequence encoder and decode networks with an attention mechanism for neural machine translation. Although the quality of the translation between English and French using the model in the tutorial may not be as good as a leading-edge transformer based model, it presents a great example of how to deploy an interesting model trained from scratch in an Android app.
This Android demo app shows:
- The Python code that saves the encoder and attention-enhanced decoder models, trained based on the code in the PyTorch NMT tutorial;
- The Python code that loads the saved encoder and decoder models, optimizes them and saves again for mobile apps to use;
- The Android app that uses the encoder and decoder models to do French-English translation.
- PyTorch 1.7 or later (Optional)
- Python 3.8 (Optional)
- Android Pytorch library 1.6, 1.7 or later
- Android Studio 4.0.1 or later
To Test Run the Object Detection Android App, follow the steps below:
If you don't have the PyTorch environment set up to run the script, you can download the PyTorch trained and optimized NMT encoder and decoder models compressed in a zip here, unzip it, copy them to the Android app's assets folder, and skip the rest of this step and go to step 2 directly.
Be aware that the downloadable model file was created with PyTorch 1.7.0, matching the PyTorch Android library 1.7.0 specified in the project's build.gradle file as implementation 'org.pytorch:pytorch_android:1.7.0'. If you use a different version of PyTorch to create your model by following the instructions below, make sure you specify the same PyTorch Android library version in the build.gradle file to avoid possible errors caused by the version mismatch. Furthermore, if you want to use the latest PyTorch master code to create the model, follow the steps at Building PyTorch Android from Source and Using the PyTorch Android Libraries Built on how to use the model in Android.
If you have a good GPU and want to train your model from scratch, uncomment the line trainIters(encoder, decoder, 450100, print_every=5000) in seq2seq_nmt.py before running python seq2seq2_nmt.py to go through the whole process of training, saving, loading, optimizing and saving the final mobile-ready models.
To just convert a pre-trained model seq2seq_mt_150000.pt to the TorchScript model used on mobile, download seq2seq_mt_150000.pt first to the same directory as seq2seq2_nmt.py, then run python seq2seq2_nmt.py. After optimized_encoder_150k.pth and optimized_decoder_150k.pth are generated, copy them to the Android app's assets folder. Note that dynamic quantization is applied to the decoder in seq2seq2_nmt.py for its nn.Linear parameters to reduce the decoder model size from 29MB to 18MB.
Start Android Studio, open the project located in android-demo-app/Seq2SeqNMT, and run on your AVD or real Android device. See this video for a screencast of the app running. Some example translation results are:
