Tensorflow implementation of Deep Gamblers.
$ pip install git+https://github.com/simaki/deep-gamblersExample: MNIST classification with abstention
import tensorflow as tf
from deep_gamblers import coverage, GamblerLoss
x_tr, y_tr = ... # Fetch MNIST
model = tf.models.Sequential([
Conv2D(10, 4, activation="relu"),
Conv2D(10, 4, activation="relu"),
Conv2D(10, 4, activation="relu"),
Conv2D(10, 4, activation="relu"),
Flatten(),
Dense(10 + 1, activation="relu"),
])
model.compile(optimizer="adam", loss=GamblerLoss(6.0), metrics=[coverage, "accuracy"])
model.fit(x_tr, y_tr, epochs=10)