keruvi (Keras runtime visualisation)

Runtime visualisation of TensorFlow Keras model training with Node.js.

Requirements

• TensorFlow (https://www.tensorflow.org/)
  • deep neural networks
• Node.js (https://nodejs.org/en/)
  • JavaScript runtime

The example models in model directory may require further packages, i.e.:

• NumPy (https://numpy.org/)
  • numerical computations
• Pillow (https://python-pillow.org/)
  • image processing

Install

TensorFlow libraries for Python can be installed via pip. If necessary, install python3 and python3-pip packages first.

# sudo apt install python3
# sudo apt install python3-pip
pip3 install tensorflow

Node project dependencies (e.g. Express framework) can be installed via npm. If necessary, install Node.js and npm first. See https://nodejs.org/en/ and https://www.npmjs.com/ for details.

# install Node.js and npm
cd server/
npm i --production

Usage

During training of Keras models, metrics can be received via Keras callbacks. Use custom KeruviRemoteMonitor callback to send runtime data to a Node server which is able to visualize current training status. Start this server as follows:

cd server/
npm run start

By default, the server will be accessible via port 3000. Use environment variable KERUVI_PORT to change this behavior.

Connect specific Keras model to visualisation server using KeruviRemoteMonitor. In the following example, the callback instance is configured with server URL http://localhost:3000 and ID boston:

from tensorflow.keras.models import Sequential
from tensorflow.keras.optimizers import RMSprop

from callback.remote_monitor import KeruviRemoteMonitor

# Create model.
model = Sequential()

# Add Keras layers as desired.
# ...

# Build model.
model.compile(optimizer=RMSprop(learning_rate=0.01), loss='mse')

# Train model using custom Keras callback.
monitor = KeruviRemoteMonitor(root='http://localhost:3000', model_id='boston')
model.fit(
    your_training_data,
    your_training_labels,
    epochs=200,
    batch_size=16,
    validation_split=0.2,
    callbacks=[monitor],
    verbose=0
)

To get corresponding training metrics, e.g. for each epoch (on_epoch_end callback functionality), visit URL http://localhost:3000/epoch/boston. This endpoint is configured to be used as an EventSource:

let oEventSource = new EventSource('epoch/boston');
oEventSource.onmessage = function(oEvent) {

    // Update visualisation.
    // ...

}

See https://www.tensorflow.org/api_docs/python/tf/keras/callbacks/Callback for more informationen concerning Keras callbacks. Sample models using a custom Keras callback can be found in model folder. Run corresponding _example.py scripts for testing, e.g. boston_housing_example.py to train a model using the Boston Housing dataset. Currently, trainable example models for the following datasets are provided:

• Boston Housing (prediction of house prices)
• GTSRB (classification of traffic sign images)
• MNIST and Chars74K (classification of handwritten digits)
• IMDb (sentiment analysis)