例子 Examples
下面给出的所有例子程序都能在examples文件夹中找到：GitHub地址。
example1_or_nn.py
用Hamaa构建单层神经网络来解决”或问题”的一个简洁的例子。 其中包含了创建、训练、测试一个神经网络必备的所有函数。
from hamaa.datasets import datasets
from hamaa.layers import Dense, Activation
from hamaa.models import Sequential
from hamaa.optimizers import SGD
# 1. create a model
model = Sequential()
# 2. add a full connected layer to model
model.add(Dense(input_dim=2, output_dim=2, init='uniform'))
# 3. add a activation layer to model
model.add(Activation('sigmoid'))
# 4. use "mean square error" as the objective of model
model.set_objective('mse')
# 5. use "stochastic gradient descent" as the optimizerof model
model.set_optimizer(SGD(lr=0.9, momentum=0.9, decay=1e-6))
# 6. print the summary of model
print model.summary()
# 7. load "or" data, note that the label "y" is one hot
# x = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
# y = np.array([[1, 0], [0, 1], [0, 1], [0, 1]])
x, y = datasets.load_or_data()
# 8. train the neural network
model.train(training_data=(x, y), nb_epochs=10)
# 9. evaluate the accuracy on data
print 'test accuracy: ', model.evaluate_accuracy(x, y)
example2_moons_nn.py
构建一个神经元数目为 [2->4->2] 的多层神经网络来对线性不可分数据集进行分类。
from hamaa.datasets import datasets
from hamaa.layers import Dense, Activation
from hamaa.models import Sequential
from hamaa.optimizers import SGD
from hamaa.utils.np_utils import split_training_data
model = Sequential()
model.add(Dense(input_dim=2, output_dim=4, init='normal'))
model.add(Activation('sigmoid'))
model.add(Dense(output_dim=2))
model.add(Activation('softmax'))
model.set_objective('categorical_crossentropy')
model.set_optimizer(SGD(lr=0.01, momentum=0.9))
print model.summary()
x, y = datasets.load_moons_data(nb_data=2000, noise=0.1)
# split nine in tenth of original data as training data, and the rest as validation data
training_data, validation_data = split_training_data(data=(x, y), split_ratio=0.9)
# "verbose" means display mode of training information
# "log_epoch" means display training information every log_epoch times
model.train(training_data=training_data,
nb_epochs=30,
mini_batch_size=100,
verbose=1,
validation_data=validation_data,
log_epoch=1)
print 'test accuracy: ', model.evaluate_accuracy(x, y)
# plot the prediction on training_data and validation_data
model.plot_prediction(data=training_data)
model.plot_prediction(data=validation_data)
# plot a line chart about accuracy and loss with epoch.
model.plot_training_iteration()
example3_mnist_nn.py
构建一个多层神经网络来对MNIST数据集进行分类。 使用进度条功能来显示训练过程。
from hamaa.datasets.datasets import load_mnist_data
from hamaa.layers import Dense, Activation
from hamaa.models import Sequential
from hamaa.optimizers import SGD
from hamaa.utils.np_utils import split_training_data
print 'loading MNIST dataset...'
# "preprocess" means normalization
training_data, test_data = load_mnist_data(nb_training=60000, nb_test=10000, preprocess=True, flatten=True)
training_data, validation_data = split_training_data(training_data, split_ratio=0.95)
print 'training_data:', training_data[0].shape
print 'validation_data:', validation_data[0].shape
print 'test_data:', test_data[0].shape
model = Sequential()
model.add(Dense(input_dim=784, output_dim=100, init='glorot_normal'))
model.add(Activation('sigmoid'))
model.add(Dense(output_dim=10, init='glorot_normal'))
model.add(Activation('softmax'))
model.set_objective('categorical_crossentropy')
model.set_optimizer(SGD(lr=0.01, momentum=0.9))
print model.summary()
model.train(training_data=training_data,
nb_epochs=10,
mini_batch_size=100,
verbose=2,
validation_data=validation_data,
log_epoch=1)
print 'test accuracy:', model.evaluate_accuracy(test_data[0], test_data[1])
model.plot_training_iteration()
example4_mnist_cnn.py
构建一个卷积神经网络来对MNIST数据集进行分类。 使用进度条功能来显示训练过程， 并使用可视化工具对卷积层权重进行可视化。
from hamaa.datasets.datasets import load_mnist_data
from hamaa.layers import Dense, Activation, Convolution2D, Flatten, MeanPooling2D
from hamaa.models import Sequential
from hamaa.optimizers import SGD
from hamaa.utils import vis_utils
from hamaa.utils.np_utils import split_training_data
print 'loading MNIST dataset...'
training_data, test_data = load_mnist_data(nb_training=6000, nb_test=10000, preprocess=True, flatten=False)
training_data, validation_data = split_training_data(training_data, nb_validation=1000)
print 'training_data:', training_data[0].shape
print 'validation_data:', validation_data[0].shape
print 'test_data:', test_data[0].shape
model = Sequential()
model.add(Convolution2D(nb_kernel=5, kernel_height=5, kernel_width=5, activation='relu', input_shape=(1, 28, 28)))
model.add(MeanPooling2D(pooling_size=(2, 2)))
model.add(Convolution2D(nb_kernel=5, kernel_height=5, kernel_width=5, activation='relu'))
model.add(MeanPooling2D(pooling_size=(2, 2)))
model.add(Flatten())
model.add(Dense(output_dim=100, init='glorot_normal'))
model.add(Activation('sigmoid'))
model.add(Dense(output_dim=10, init='glorot_normal'))
model.add(Activation('softmax'))
model.set_objective('categorical_crossentropy')
model.set_optimizer(SGD(lr=0.002, momentum=0.9))
print model.summary()
model.train(training_data=training_data,
nb_epochs=20,
mini_batch_size=64,
verbose=2,
validation_data=validation_data,
log_epoch=1)
print model.evaluate_accuracy(test_data[0], test_data[1])
# visualize the kernels of convolution layer
vis_utils.visualize_convolution_kernel(model.layers[0], title='layer 0')
vis_utils.visualize_convolution_kernel(model.layers[2], title='layer 2')
model.plot_training_iteration()
在MNIST数据集上测试准确率达到99.3%的卷积神经网络卷积核可视化示意图： (注: 非来自上述程序)
第一个卷积层:
第二个卷积层:
