In this project I explore all phase of a Deep Network develops:

• First, we train different RNN, using different techniques and hyperparameter and we choose the best: 2 layer GRU + 1 Dense. We use Keras for this phase.
• Use the network to generate sequences in local with python.
• Translate the network and weight to TensorFlow and save.
• Deploy using tensorflow-serving and get predictions using python.
• Deploy in Google Cloud ML and get sequences with python (you need and authorized google account).
• Serve predictions using a REST API in Google App Engine and get sequences with python.
• Deploy a Web Page in Google App Engine that uses this. API.

We use text from the Galician politician Beiras and keras for defining and training.

This work is based on aind2-rnn

We develog a Recurrent Neural Network (RNN) to create a Galician language sequence generator.

We test diferents RNN network:

• LSTM
• GRU
• GRU + Dropout

We also test ModelCheckpoint in training.

The best network for this is case is a GRU with 3 layers.

We develop a GAN network with Recurrent Neural Network (RNN) to create a Galician language sequence generator.

A GAN network does not work in this case

The GAN implementation is based on MNIST_CNN_GAN

We develog a Recurrent Neural Network (RNN) with TimeDistributed to create a Galician language sequence generator.

**Get good result only using 100000 carazters for training, but it use a lot of memory **

In beiras-rnn/Beiras-rnn.ipynb we choose GRU as our best network, here we try some different hyperparameters:

Move the learning rate: With 0.001 work perfect, when we reduce the net does not learn.

Change to Adam Optimizer: We get similar results than using RSMprop

Reduce batch size from 500 to 32.- This increase the time and I could not finish the learning process.

It normally takes 10 hours to train the models in a g2.2xlarge AWS machine.

• predict-local.py : Create the model using the weights and make a sequence from the sentence introduce in the command line.
• Gru2TensorfFlow .- Jupyter notebook to convert keras model to tensorflow model and how do deploy with tensorflow-serving.
• predict-tf-serving.py : Make a secuence from the sentence introduce in the command line using tensorflow-serving.

• Google ML Python 2.7.ipynb .- How to deploy to Google Cloud ML
• predict-google-cloud-ml.py : Make a secuence from the sentence introduce in the command line using Google Gloud ML.
• predict/ - Google App engine API and Web page.
• predict-app-engine.py : Make a secuence from the sentence introduce in the command line using API app engine.

In this part we train our model in Google cloud: we change our model to be able to use it in the cloud, do a local train for test and then train the Google cloud using tensorflow, gcloud and google-ml. I tried also to use a machine with multi GPU, but it was slower. I tried also to train on multiple machines but It does not work.

conda env create -f environment.yml

source activate beiras-rnn

echo "deb [arch=amd64] http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | sudo tee /etc/apt/sources.list.d/tensorflow-serving.list
curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | sudo apt-key add -
sudo apt-get update && sudo apt-get install tensorflow-model-server

jupyter notebook

cd predict
python predict-local.py [sentence]

tensorflow_model_server --port=9000 --model_name=default --model_base_path=/home/aind2/beiras-rnn/export-tf &
cd predict
python predict-remote.py [sentence]

• Keras

• Tensorflow

• Jupyter

• Jose Manuel Fernandez Lorenzo - jotavaladouro-
• Based in aind2-rnn