Keras runtime visualisation.
Runtime visualisation of TensorFlow Keras model training with Node.js.
The example models in model directory may require further packages, i.e.:
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-pippip3 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 npmcd server/npm i --production
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 Sequentialfrom tensorflow.keras.optimizers import RMSpropfrom 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: