independence.py

# import WRN_old as WRN #For previous runs
import WRN
import tensorflow as tf
from utils import *
import os
import random
import time
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
# from keras import backend as K

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

def plot_variance_activations(model, layer_name, x_train):
    layer_output=model.get_layer(layer_name).output
    intermediate_model=tf.keras.models.Model(inputs=model.input,outputs=layer_output)

    x_train = list(x_train)
    ex1, ex2 = random.sample(x_train, 2)
    act_1, act_2, act_3 = [], [], []

    for _ in range(5):
        ex3 = random.sample(x_train, 1)[0]
        example1 = [tf.convert_to_tensor(ex3),
                    tf.convert_to_tensor(ex1),
                    tf.convert_to_tensor(ex2)]
        example2 = [tf.convert_to_tensor(ex1),
                    tf.convert_to_tensor(ex3),
                    tf.convert_to_tensor(ex2)]
        example3 = [tf.convert_to_tensor(ex1),
                    tf.convert_to_tensor(ex2),
                    tf.convert_to_tensor(ex3)]
        stack1 = tf.stack(example1)[None, :, :,:]
        stack2 = tf.stack(example2)[None, :, :,:]
        stack3 = tf.stack(example3)[None, :, :,:]

        act_1.append(tf.reshape(intermediate_model(stack1),[-1]))
        act_2.append(tf.reshape(intermediate_model(stack2),[-1]))
        act_3.append(tf.reshape(intermediate_model(stack3),[-1]))


    act_tfs = [tf.stack(act) for act in [act_1, act_2, act_3]]
    var_1, var_2, var_3 = [tf.math.reduce_std(act_tf, axis=0) for act_tf in act_tfs]
    print(var_1, var_2, var_3)
    fig = plt.figure()
    ax = fig.add_subplot()
    plt.scatter(var_1, var_2, s=1)
    plt.title("Variance of the activation of the neurons in the NN")
    plt.xlabel("Variance when changing input 1")
    plt.ylabel("Variance when changing input 2")
    ax.set_aspect('equal', adjustable='box')
    plt.savefig('plots/variance/Activation_Variance_0.png')
    plt.show()

    fig = plt.figure()
    ax = fig.add_subplot(projection='3d')
    ax.scatter(var_1, var_2, var_3, s=1)
    plt.title("Variance of the activation of the neurons in the NN")
    ax.set_xlabel('Variance changing input 1')
    ax.set_ylabel('Variance changing input 2')
    ax.set_zlabel('Variance changing input 3')
    plt.savefig('plots/variance/Activation_Variance_3D_0.png')
    plt.show()

def main():
    n, k, M = 28, 10, 3
    classes = 10
    input_shape = [M]+[32, 32, 3]
    model = WRN.build_model(input_shape, classes, n, k, M)
    checkpoint_path="run/Cifar10/0_0_0_3/weights_0.h5"
    model.load_weights(checkpoint_path)

    # print(model.summary())

    layer_name="conv2d_10"

    (x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()

    plot_variance_activations(model, layer_name, x_train)

if __name__ == "__main__":
    main()