In [1]:
'''Train a Siamese MLP on pairs of digits from the MNIST dataset.
It follows Hadsell-et-al.'06 [1] by computing the Euclidean distance on the
output of the shared network and by optimizing the contrastive loss (see paper
for mode details).
[1] "Dimensionality Reduction by Learning an Invariant Mapping"
http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf
Gets to 99.5% test accuracy after 20 epochs.
3 seconds per epoch on a Titan X GPU
'''
Out[1]:
In [16]:
import numpy as np
import random
In [17]:
# Keras imports
from keras.datasets import mnist
from keras.models import Sequential, Model
from keras.layers import Dense, Dropout, Input, Lambda
from keras.optimizers import RMSprop
from keras import backend as K
In [18]:
def euclidean_distance(vects):
x, y = vects
return K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon()))
In [19]:
def eucl_dist_output_shape(shapes):
shape1, shape2 = shapes
return (shape1[0], 1)
In [20]:
def contrastive_loss(y_true, y_pred):
'''Contrastive loss from Hadsell-et-al.'06
http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf
'''
margin = 1
return K.mean(y_true * K.square(y_pred) +
(1 - y_true) * K.square(K.maximum(margin - y_pred, 0)))
In [21]:
def create_pairs(x, digit_indices):
'''Positive and negative pair creation.
Alternates between positive and negative pairs.
'''
pairs = []
labels = []
n = min([len(digit_indices[d]) for d in range(10)]) - 1
for d in range(10):
for i in range(n):
z1, z2 = digit_indices[d][i], digit_indices[d][i + 1]
pairs += [[x[z1], x[z2]]]
inc = random.randrange(1, 10)
dn = (d + inc) % 10
z1, z2 = digit_indices[d][i], digit_indices[dn][i]
pairs += [[x[z1], x[z2]]]
labels += [1, 0]
return np.array(pairs), np.array(labels)
In [22]:
def create_base_network(input_dim):
'''Base network to be shared (eq. to feature extraction).
'''
seq = Sequential()
seq.add(Dense(128, input_shape=(input_dim,), activation='relu'))
seq.add(Dropout(0.1))
seq.add(Dense(128, activation='relu'))
seq.add(Dropout(0.1))
seq.add(Dense(128, activation='relu'))
return seq
In [23]:
def compute_accuracy(predictions, labels):
'''Compute classification accuracy with a fixed threshold on distances.
'''
return labels[predictions.ravel() < 0.5].mean()
In [29]:
# the data, shuffled and split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.reshape(60000, 784)
x_test = x_test.reshape(10000, 784)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
# normalize
x_train /= 255
x_test /= 255
# flatten images
input_dim = 784
epochs = 20
In [25]:
# create training+test positive and negative pairs
digit_indices = [np.where(y_train == i)[0] for i in range(10)]
tr_pairs, tr_y = create_pairs(x_train, digit_indices)
digit_indices = [np.where(y_test == i)[0] for i in range(10)]
te_pairs, te_y = create_pairs(x_test, digit_indices)
In [26]:
# network definition
base_network = create_base_network(input_dim)
input_a = Input(shape=(input_dim,))
input_b = Input(shape=(input_dim,))
In [27]:
# because we re-use the same instance `base_network`,
# the weights of the network
# will be shared across the two branches
processed_a = base_network(input_a)
processed_b = base_network(input_b)
distance = Lambda(euclidean_distance,
output_shape=eucl_dist_output_shape)([processed_a, processed_b])
model = Model([input_a, input_b], distance)
In [28]:
# train
rms = RMSprop()
model.compile(loss=contrastive_loss, optimizer=rms)
model.fit([tr_pairs[:, 0], tr_pairs[:, 1]], tr_y,
batch_size=128,
epochs=epochs,
validation_data=([te_pairs[:, 0], te_pairs[:, 1]], te_y))
Out[28]:
In [14]:
# compute final accuracy on training and test sets
pred = model.predict([tr_pairs[:, 0], tr_pairs[:, 1]])
tr_acc = compute_accuracy(pred, tr_y)
pred = model.predict([te_pairs[:, 0], te_pairs[:, 1]])
te_acc = compute_accuracy(pred, te_y)
print('* Accuracy on training set: %0.2f%%' % (100 * tr_acc))
print('* Accuracy on test set: %0.2f%%' % (100 * te_acc))
In [ ]: