segmind / keras-jukebox

UI to control learning rate, take snapshot and perform a few other actions without interrupting your training.

Acknowledgement

This project is supported by Segmind

keras JukeBox

This is a UI based hyper-parameter controller, which let's you control the following.

• start, pause and stop a live training.
• reset the learning rate on dynamically while training is in progress.
• take a snapshot at will

more functionalities are to be added

Dependencies

This package depends on MQTT protocol for communication. So, it is expected that an MQTT broker is up and running in 'localhost' at port 1883(default port).

Install it by :

sudo apt-get update
sudo apt-get install mosquitto
sudo apt-get install mosquitto-clients

Python dependencies:

• python >= 3.6.8
• paho-mqtt
• PyQt5
• tensorflow >= 1.14

Note: This package is intended and tested for tensorflow-keras api and NOT keras with tensorflow 'backend'

Usage

you can try the following example

save the follwing example fashion_mnist_jukebox.py

from __future__ import absolute_import, division, print_function, unicode_literals

import tensorflow as tf
from tensorflow import keras


# import the callback
from keras_jukebox import JukeBoxCallback


fashion_mnist = keras.datasets.fashion_mnist

(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()


train_images = train_images / 255.0

test_images = test_images / 255.0

model = keras.Sequential([
    keras.layers.Flatten(input_shape=(28, 28)),
    keras.layers.Dense(128, activation='relu'),
    keras.layers.Dense(10, activation='softmax')
])


model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# pass the jukebox callback to model.fit method
model.fit(train_images, train_labels, epochs=10, callbacks=[JukeBoxCallback(verbose=1)])

and run it. You will notice that the script starts but training doesn't, which is because it is paused and needs a JukeBox-UI to start.

Now, open a new terminal(Alt+ctrl+T) and start the JukeBox by typing:

start_jukebox

and you should see the UI pop up, note the algorithm is in pause mode by default. Hit the play button to start the training.