keras的siamese(孪生网络)实现案例

本是古典何须时尚 2023-02-18 09:53 16阅读 0赞

更多python视频教程请到菜鸟教程https://www.piaodoo.com/

代码位于keras的[官方样例][Link 1]，并做了微量修改和大量学习?。

最终效果：

![20200612141852.jpg][]

![20200612141901.jpg][]

import keras
    import numpy as np
    import matplotlib.pyplot as plt

import random

from keras.callbacks import TensorBoard  
from keras.datasets import mnist  
from keras.models import Model  
from keras.layers import Input, Flatten, Dense, Dropout, Lambda  
from keras.optimizers import RMSprop  
from keras import backend as K

num\_classes = 10  
epochs = 20

def euclidean\_distance(vects):  
x, y = vects  
sum\_square = K.sum(K.square(x - y), axis=1, keepdims=True)  
return K.sqrt(K.maximum(sum\_square, K.epsilon()))

def eucl\_dist\_output\_shape(shapes):  
shape1, shape2 = shapes  
return (shape1\[0\], 1)

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  
sqaure\_pred = K.square(y\_pred)  
margin\_square = K.square(K.maximum(margin - y\_pred, 0))  
return K.mean(y\_true \* sqaure\_pred + (1 - y\_true) \* margin\_square)

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(num\_classes)\]) - 1  
for d in range(num\_classes):  
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, num\_classes)  
dn = (d + inc) % num\_classes  
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)

def create\_base\_network(input\_shape):  
‘’‘Base network to be shared (eq. to feature extraction).  
‘’’  
input = Input(shape=input\_shape)  
x = Flatten()(input)  
x = Dense(128, activation=‘relu’)(x)  
x = Dropout(0.1)(x)  
x = Dense(128, activation=‘relu’)(x)  
x = Dropout(0.1)(x)  
x = Dense(128, activation=‘relu’)(x)  
return Model(input, x)

def compute\_accuracy(y\_true, y\_pred): \# numpy上的操作  
‘’‘Compute classification accuracy with a fixed threshold on distances.  
‘’’  
pred = y\_pred.ravel() < 0.5  
return np.mean(pred == y\_true)

def accuracy(y\_true, y\_pred): \# Tensor上的操作  
‘’‘Compute classification accuracy with a fixed threshold on distances.  
‘’’  
return K.mean(K.equal(y\_true, K.cast(y\_pred < 0.5, y\_true.dtype)))

def plot\_train\_history(history, train\_metrics, val\_metrics):  
plt.plot(history.history.get(train\_metrics), ‘-o’)  
plt.plot(history.history.get(val\_metrics), ‘-o’)  
plt.ylabel(train\_metrics)  
plt.xlabel(‘Epochs’)  
plt.legend(\[‘train’, ‘validation’\])

# the data, split between train and test sets #

(x\_train, y\_train), (x\_test, y\_test) = mnist.load\_data()  
x\_train = x\_train.astype(‘float32’)  
x\_test = x\_test.astype(‘float32’)  
x\_train /= 255  
x\_test /= 255  
input\_shape = x\_train.shape\[1:\]

# create training+test positive and negative pairs #

digit\_indices = \[np.where(y\_train == i)\[0\] for i in range(num\_classes)\]  
tr\_pairs, tr\_y = create\_pairs(x\_train, digit\_indices)

digit\_indices = \[np.where(y\_test == i)\[0\] for i in range(num\_classes)\]  
te\_pairs, te\_y = create\_pairs(x\_test, digit\_indices)

# network definition #

base\_network = create\_base\_network(input\_shape)

input\_a = Input(shape=input\_shape)  
input\_b = Input(shape=input\_shape)

# 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)  
keras.utils.plot\_model(model,“siamModel.png”,show\_shapes=True)  
model.summary()

# train #

rms = RMSprop()  
model.compile(loss=contrastive\_loss, optimizer=rms, metrics=\[accuracy\])  
history=model.fit(\[tr\_pairs\[:, 0\], tr\_pairs\[:, 1\]\], tr\_y,  
batch\_size=128,  
epochs=epochs,verbose=2,  
validation\_data=(\[te\_pairs\[:, 0\], te\_pairs\[:, 1\]\], te\_y))

plt.figure(figsize=(8, 4))  
plt.subplot(1, 2, 1)  
plot\_train\_history(history, ‘loss’, ‘val\_loss’)  
plt.subplot(1, 2, 2)  
plot\_train\_history(history, ‘accuracy’, ‘val\_accuracy’)  
plt.show()

# compute final accuracy on training and test sets #

y\_pred = model.predict(\[tr\_pairs\[:, 0\], tr\_pairs\[:, 1\]\])  
tr\_acc = compute\_accuracy(tr\_y, y\_pred)  
y\_pred = model.predict(\[te\_pairs\[:, 0\], te\_pairs\[:, 1\]\])  
te\_acc = compute\_accuracy(te\_y, y\_pred)

print(’\* Accuracy on training set: %0.2f%%’ % (100 \* tr\_acc))  
print(’\* Accuracy on test set: %0.2f%%’ % (100 \* te\_acc))

以上这篇keras的siamese(孪生网络)实现案例就是小编分享给大家的全部内容了，希望能给大家一个参考，也希望大家多多

茂名论坛https://www.hnthzk.com/

化州橘红http://www.sfkyty.com/

茂名论坛http://www.nrso.net/

源码搜藏网http://www.cntkd.net/

[Link 1]: https://github.com/keras-team/keras/blob/master/examples/mnist_siamese.py
[20200612141852.jpg]: https://img.jbzj.com/file_images/article/202006/20200612141852.jpg
[20200612141901.jpg]: https://img.jbzj.com/file_images/article/202006/20200612141901.jpg