Skip to content

Instantly share code, notes, and snippets.

@inderpartap

inderpartap/NeuralNetVisualize.py

Created February 1, 2020 07:20
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save inderpartap/e9571c62de6b1af31f03767ad718b3dc to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save inderpartap/e9571c62de6b1af31f03767ad718b3dc to your computer and use it in GitHub Desktop.
Download ZIP
This code visualizes how a Multi-layer Perceptron (MLP) Neural Network recognizes hand-written digits from the MNIST data set.
Raw
NeuralNetVisualize.py
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib.animation
from sklearn.model_selection import train_test_split
plt.rcParams["animation.html"] = "jshtml"
from keras.models import Sequential
from keras.layers import Dense, Activation, Dropout
from keras.utils import to_categorical, plot_model
from keras.datasets import mnist
from keras import backend as K
# load mnist dataset
data = pd.read_csv('../input/train.csv')
# split data into train and test sample
x_train, x_test, y_train, y_test = train_test_split(data.iloc[:,1:785], data.iloc[:,0],
test_size = 0.1, random_state = 42)
x_train = x_train.values.reshape(37800, 784)
x_test = x_test.values.reshape(4200, 784)
# compute the number of labels
num_labels = len(np.unique(y_train))
# convert to one-hot vector
y_train = to_categorical(y_train)
y_test = to_categorical(y_test)
# normalize
x_train = x_train.astype('float32') / 255
x_test = x_test.astype('float32') / 255
# network parameters
input_size = x_train.shape[1]
batch_size = 64
dropout = 0.45
# this model is a 3-layer MLP with ReLU and dropout after each layer
model = Sequential()
model.add(Dense(256, input_dim=input_size))
model.add(Activation('relu'))
model.add(Dropout(dropout))
model.add(Dense(64))
model.add(Activation('relu'))
model.add(Dropout(dropout))
model.add(Dense(num_labels))
model.add(Activation('softmax'))
model.summary()
# loss function for one-hot vector using adam optimizer
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# train the network
model.fit(x_train, y_train, epochs=20, batch_size=batch_size)
# validate the model on test dataset to determine generalization
loss, acc = model.evaluate(x_test, y_test, batch_size=batch_size)
print("\nTest accuracy: %.1f%%" % (100.0 * acc))
get_layer_output = K.function([model.layers[0].input, model.layers[0].input, model.layers[0].input],
[model.layers[1].output, model.layers[4].output, model.layers[7].output])
layer1_output, layer2_output, layer3_output = get_layer_output([x_train])
train_ids = [np.arange(len(y_train))[y_train[:,i] == 1] for i in range(10)]
%%capture
%matplotlib inline
# digit to be plotted
digit = 5
# indices of frames to be plotted for this digit
n = range(50)
# initialize plots
f, (ax1, ax2, ax3, ax4) = plt.subplots(1, 4, figsize=(15,4))
# prepare plots
ax1.set_title('Input Layer', fontsize=16)
ax1.axes.get_xaxis().set_visible(False)
ax1.axes.get_yaxis().set_visible(False)
ax2.set_title('Hidden Layer 1', fontsize=16)
ax2.axes.get_xaxis().set_visible(False)
ax2.axes.get_yaxis().set_visible(False)
ax3.set_title('Hidden Layer 2', fontsize=16)
ax3.axes.get_xaxis().set_visible(False)
ax3.axes.get_yaxis().set_visible(False)
ax4.set_title('Output Layer', fontsize=16)
ax4.axes.get_xaxis().set_visible(False)
ax4.axes.get_yaxis().set_visible(False)
# add numbers to the output layer plot to indicate label
for i in range(3):
for j in range(4):
text = ax4.text(j, i, [[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, '', '']][i][j],
ha="center", va="center", color="w", fontsize=16)
def animate(id):
# plot elements that are changed in the animation
digit_plot = ax1.imshow(x_train[train_ids[digit][id]].reshape((28,28)), animated=True)
layer1_plot = ax2.imshow(layer1_output[train_ids[digit][id]].reshape((16,16)), animated=True)
layer2_plot = ax3.imshow(layer2_output[train_ids[digit][id]].reshape((8,8)), animated=True)
output_plot = ax4.imshow(np.append(layer3_output[train_ids[digit][id]],
[np.nan, np.nan]).reshape((3,4)), animated=True)
return digit_plot, layer1_plot, layer2_plot, output_plot,
# define animation
ani = matplotlib.animation.FuncAnimation(f, animate, frames=n, interval=100)
plt.show()
%%capture
%matplotlib inline
# digit to be plotted
digit = 6
# numbers of frames to be summed over
n = np.append([1], np.linspace(5, 100, 20, dtype=int))
# initialize plots
f, (ax1, ax2, ax3, ax4) = plt.subplots(1, 4, figsize=(15,4))
# add a counter indicating the number of frames used in the summation
counter = ax1.text(1, 2, 'n={}'.format(0), color='white', fontsize=16, animated=True)
# prepare plots
ax1.set_title('Input Layer', fontsize=16)
ax1.axes.get_xaxis().set_visible(False)
ax1.axes.get_yaxis().set_visible(False)
ax2.set_title('Hidden Layer 1', fontsize=16)
ax2.axes.get_xaxis().set_visible(False)
ax2.axes.get_yaxis().set_visible(False)
ax3.set_title('Hidden Layer 2', fontsize=16)
ax3.axes.get_xaxis().set_visible(False)
ax3.axes.get_yaxis().set_visible(False)
ax4.set_title('Output Layer', fontsize=16)
ax4.axes.get_xaxis().set_visible(False)
ax4.axes.get_yaxis().set_visible(False)
# add numbers to the output layer plot to indicate label
for i in range(3):
for j in range(4):
text = ax4.text(j, i, [[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, '', '']][i][j],
ha="center", va="center", color="w", fontsize=16)
def animate(id):
# plot elements that are changed in the animation
digit_plot = ax1.imshow(np.sum(x_train[train_ids[digit][:id]], axis=0).reshape((28,28)), animated=True)
layer1_plot = ax2.imshow(np.sum(layer1_output[train_ids[digit][:id]], axis=0).reshape((16,16)), animated=True)
layer2_plot = ax3.imshow(np.sum(layer2_output[train_ids[digit][:id]], axis=0).reshape((8,8)), animated=True)
output_plot = ax4.imshow(np.append(np.sum(layer3_output[train_ids[digit][:id]], axis=0),
[np.nan, np.nan]).reshape((3,4)), animated=True)
counter.set_text('n={}'.format(id))
return digit_plot, layer1_plot, layer2_plot, output_plot, counter,
# define animation
ani = matplotlib.animation.FuncAnimation(f, animate, frames=n, interval=100)
plt.show()
f, ax_arr = plt.subplots(2, 5, figsize=(15,10))
f.subplots_adjust(wspace=0.05, bottom=0.5, top=0.95)
for i, ax in enumerate(np.ravel(ax_arr)):
ax.axes.get_xaxis().set_visible(False)
ax.axes.get_yaxis().set_visible(False)
if i <= 10:
ax.set_title('- {} -'.format(i), fontsize=16)
layer1_plot = ax.imshow(np.sum(layer1_output[train_ids[i]], axis=0).reshape((16,16)))
similarity_layer1 = np.zeros((10,10))
for i in range(10):
for j in range(10):
sum_i_normalized = np.sqrt(np.sum(layer1_output[train_ids[i]], axis=0)/np.sum(layer1_output[train_ids[i]]))
sum_j_normalized = np.sqrt(np.sum(layer1_output[train_ids[j]], axis=0)/np.sum(layer1_output[train_ids[j]]))
similarity_layer1[i,j] = np.sum(sum_i_normalized*sum_j_normalized)
f, ax = plt.subplots()
similarity_layer1_plot = ax.imshow(similarity_layer1, origin='lower')
plt.colorbar(similarity_layer1_plot);
f, ax_arr = plt.subplots(2, 5, figsize=(15,10))
f.subplots_adjust(wspace=0.05, bottom=0.5, top=0.95)
for i, ax in enumerate(np.ravel(ax_arr)):
ax.axes.get_xaxis().set_visible(False)
ax.axes.get_yaxis().set_visible(False)
if i <= 10:
ax.set_title('- {} -'.format(i), fontsize=16)
layer2_plot = ax.imshow(np.sum(layer2_output[train_ids[i]], axis=0).reshape((8,8)))
similarity_layer2 = np.zeros((10,10))
for i in range(10):
for j in range(10):
sum_i_normalized = np.sqrt(np.sum(layer2_output[train_ids[i]], axis=0)/np.sum(layer2_output[train_ids[i]]))
sum_j_normalized = np.sqrt(np.sum(layer2_output[train_ids[j]], axis=0)/np.sum(layer2_output[train_ids[j]]))
similarity_layer2[i,j] = np.sum(sum_i_normalized*sum_j_normalized)
f, ax = plt.subplots()
similarity_layer2_plot = ax.imshow(similarity_layer2, origin='lower')
plt.colorbar(similarity_layer2_plot);
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment