transparent-keras 0.3.1

Provide a drop-in replacement for a Keras Model that allows a look under the hood during training

Transparent Keras aims to provide a very simple way to look under the hood during training of Keras models by defining an extra set of outputs that will be returned by train_on_batch or test_on_batch.

The API is extremely simple all that is provided is a TransparentModel that accepts an extra constructor keyword argument observed_tensors. The created model should behave exactly like a Keras model except for the functions (train|test)_on_batch, which return the extra tensors as after their normal return values.

Example

from keras.layers import Activation, Dense, Dropout, Input
import numpy as np

from transparent_keras import TransparentModel

x0 = Input(shape=(10,))
x = Dense(10, activation="relu")(x0)
x = Dropout(0.5)(x)
y_extra = x = Dense(10)(x)
x = Activation("relu")(x)
x = Dropout(0.5)(x)
y = Dense(1)(x)

m = TransparentModel(inputs=[x0], outputs=[y], observed_tensors=[y_extra])
m.compile(optimizer="sgd", loss="mse")

x_random = np.random.rand(128, 10)
y_random = np.random.rand(128, 1)
loss, y_extra = m.train_on_batch(x_random, y_random)