Emoji TensorFlow-iOS

This is a TensorFlow demo that can be run on iOS. It implements a text classifier that can predict emoji from short text (like tweets).

Presentation: TensorFlow on iOS.pdf

License

Prerequests

• TensorFlow (installation & source code)
• Keras
• Python 3
• Xcode
• Jupyter (for viewing ipynb files)
• Twitter data

How to train the model?

I have included a pretrained Keras model in this repository (p5-40-test.hdf5) that you can play with. But in case you want to train it by yourself, here is a brief guide.

1. Prepare training data.
   1. Downlaod and unzip twitter training data.
   2. Modify the $INPUT directory in extract_all.sh and run the script. Then you will get data/extracted.list.
   3. Run stats_top.py to get the top emojis stored at data/stat.txt.
   4. Open Jupyter notebook and run build_dataset.ipynb. It produces data/dataset.pickle with all sampled training data.
   5. Run tokenize_dataset.ipynb to produce the tokenized dataset data/plain_dataset.pickle as well as metadata data/plain_dataset_meta.pickle.
2. Run train.py to train the model. High end Cuda GPUs are recommended. The trained Keras model will be saved as p5-40-test.hdf5.
3. (Optional) You can try the model on arbitrary input with replayer.ipynb.
4. (Optional) You can also try to visualize the training process by tensorboard --log_dir=. if you have trained a model.

How to compile TensorFlow for iOS?

Follow the official compile guide here.

How to run the model on iOS?

Unfortunately, TensorFlow for iOS is still an alpha version. So we have to tweak a little bit to make it work.

Add additional integer Ops for LSTM.

Navigate to tensorflow/core/kernels directory and change the code like this:

At cwise_op_add_1.cc:

    // -- Original
    REGISTER5(BinaryOp, CPU, "Add", functor::add, float, Eigen::half, double, int32,
              int64);
    #if TENSORFLOW_USE_SYCL

    // -- Change to
    REGISTER5(BinaryOp, CPU, "Add", functor::add, float, Eigen::half, double, int32,
              int64);
    #if defined(__ANDROID_TYPES_SLIM__)
    REGISTER(BinaryOp, CPU, "Add", functor::add, int32);
    #endif  // __ANDROID_TYPES_SLIM__
    #if TENSORFLOW_USE_SYCL

At cwise_op_less.cc:

    // -- Original
    REGISTER8(BinaryOp, CPU, "Less", functor::less, float, Eigen::half, double,
              int32, int64, uint8, int8, int16);
    #if GOOGLE_CUDA

    // -- Change to
    REGISTER8(BinaryOp, CPU, "Less", functor::less, float, Eigen::half, double,
              int32, int64, uint8, int8, int16);
    #if defined(__ANDROID_TYPES_SLIM__)
    REGISTER(BinaryOp, CPU, "Less", functor::less, int32);
    #endif  // __ANDROID_TYPES_SLIM__
    #if GOOGLE_CUDA

Then compile TensorFlow again and you won't encounter "No OpsKernel found" issue.

Convert the Keras model to TensorFlow model.

Run export_tf_model.ipynb to convert Keras model file p5-40-test.hdf5 to TensorFlow model:

• GraphDef: export/p5-40-test-serving/graph-serving.pb
• Checkpoint: export/p5-40-test-serving/model-ckpt-*

Navigate to export/p5-40-test-serving directory and run the following command to convert the model to mobile version: python3 -m tensorflow.python.tools.freeze_graph
--input_graph="graph-serving.pb" --input_checkpoint="model-ckpt"
--output_graph="frozen.pb" --output_node_names="dense_2/Softmax"

Finally you will get forzen.pb file which will be used later.

Run the model on iOS.

Copy the Xcode project emoji_demo to tensorflow/tensorflow/contrib/ios_examples. You should be able to compile and run it on iOS now. The demo itself includes a pretrained model at data/forzen.pb. To run your own model, you have to replace it with yours.