Support for convolutional neural network

[JonvoWoo]

Hi, This library is very convenient.
Does this library support convolutional neural networks, such as tfp.layers.Convolution2DFlipout?

nnetwork = inf.layers.Sequential([
tfp.layers.Convolution2DFlipout(.....),
tf.keras.layers.BatchNormalization(),
tf.keras.layers.Activation('relu'),
tf.keras.layers.MaxPooling2D(......),
tf.keras.layers.Flatten(),
tfp.layers.DenseFlipout(......),
tfp.layers.DenseFlipout(........)
])

[rcabanasdepaz]

Yes, InferPy is compatible with tfp.layers.Convolution2DFlipout. As an example, I give you a small example for MNIST binary classification. Note that when the NN contains a variational layer, inf.layers.Sequential must be used so that the inference could access to the loss of such layer. By contrast, if we were using tf.keras.layers.Conv2D, we could use directly tf.keras.Sequential.

Note as well, that an extra dimension is added to the input of the NN with tf.expand_dims. Convolutional layers in Keras require 4 dimensions (#batch, #channel, #dim1, #dim2) while the output of InferPy variable is (#batch, #dim1, #dim2).

import tensorflow as tf
import inferpy as inf
from inferpy.data import mnist
import tensorflow_probability as tfp
import numpy as np

N = 1000 # data size
(x_train, y_train), (x_test, y_test) = mnist.load_data(num_instances=N,
                                  digits=[0, 1], vectorize=False)

S = np.shape(x_train)[1:]

@inf.probmodel
def cnn_flipout_classifier(S):
    with inf.datamodel():
        x = inf.Normal(tf.ones(S), 1, name="x")

        nn = inf.layers.Sequential([
            tfp.layers.Convolution2DFlipout(4, kernel_size=(10,10), padding="same", activation="relu"),
            tf.keras.layers.GlobalMaxPool2D(),
            tf.keras.layers.Dense(1, activation='sigmoid')
        ])


        y = inf.Normal(nn(tf.expand_dims(x, 1)), 0.001, name="y")



p = cnn_flipout_classifier(S)


# Empty Q model
@inf.probmodel
def qmodel():
    pass
q = qmodel()

# set the inference algorithm
VI = inf.inference.VI(q, epochs=2000)

# learn the parameters
p.fit({"x": x_train, "y":y_train}, VI)


# evaluate the model
def evaluate(p, x, y):
    N = np.shape(x)[0]
    output = p.posterior_predictive("y", {"x": x[:N]}).sample()
    x_pred = np.reshape(1* (output>0.5), (N,))
    return np.sum(x_pred == y)/N

acc = evaluate(p, x_test[:N], y_test[:N])
print(f"accuracy = {acc}")