动手学深度学习(三)——丢弃法

import mxnet as mx
from mxnet import nd
from mxnet import gluon
from mxnet import autograd
from utils import load_data_fashion_mnist, accuracy, evaluate_accuracy, SGD

# 设置随机种子
mx.random.seed(2)

# 实现dropout
def dropout(X, drop_probability):
    # 计算保留数据的比例
    keep_probability = 1 - drop_probability
    # 确保drop_probability的输入合法
    assert 0 <= keep_probability <= 1
    # 丢弃所有元素
    if keep_probability == 0:
        return X.zeros_like()
    # 随机生成一个相同纬度的矩阵, 根据随机值和keep_probability的对比确定是否丢弃该元素
    mask = nd.random.uniform(0, 1.0, X.shape, ctx=X.context) < keep_probability
    # 保证 E[dropout(X)] == X, 对剩下的数据进行缩放
    scale = 1 / keep_probability
    return mask * X * scale

# 测试dropout
A = nd.arange(20).reshape((5,4))
dropout(A, 0.0)

[[  0.   1.   2.   3.]
 [  4.   5.   6.   7.]
 [  8.   9.  10.  11.]
 [ 12.  13.  14.  15.]
 [ 16.  17.  18.  19.]]
<NDArray 5x4 @cpu(0)>

dropout(A, 1.0)

[[ 0.  0.  0.  0.]
 [ 0.  0.  0.  0.]
 [ 0.  0.  0.  0.]
 [ 0.  0.  0.  0.]
 [ 0.  0.  0.  0.]]
<NDArray 5x4 @cpu(0)>

dropout(A, 0.5)

[[  0.   2.   4.   0.]
 [  8.   0.  12.   0.]
 [ 16.  18.   0.   0.]
 [  0.   0.   0.  30.]
 [  0.  34.  36.   0.]]
<NDArray 5x4 @cpu(0)>

# 批数据大小
batch_size = 256

# 加载数据
train_data, test_data = load_data_fashion_mnist(batch_size)

# 模型输入大小
num_inputs = 28 * 28

# 模型输出大小
num_outputs = 10

# 第一个隐藏层节点数量
num_hidden1 = 256

# 第二个隐藏层节点数量
num_hidden2 = 256

# 随机数据时的标准差
weight_scale = 0.01

# 第一个隐藏层权重
W1 = nd.random_normal(shape=(num_inputs, num_hidden1), scale=weight_scale)
# 第一个隐藏层偏置
b1 = nd.zeros(num_hidden1)

# 第二个隐藏层权重
W2 = nd.random_normal(shape=(num_hidden1, num_hidden2), scale=weight_scale)
# 第二个隐藏层偏置
b2 = nd.zeros(num_hidden2)

# 输出层权重
W3 = nd.random_normal(shape=(num_hidden2, num_outputs), scale=weight_scale)
# 输出层偏置
b3 = nd.zeros(num_outputs)

# 参数数组
params = [W1, b1, W2, b2, W3, b3]

# 需要计算梯度, 添加自动求导
for param in params:
    param.attach_grad()

# 第一个隐藏层的丢弃概率
drop_prob1 = 0.2
# 第二个隐藏层的丢弃概率
drop_prob2 = 0.5

# 定义网络
def net(X):
    X = X.reshape((-1, num_inputs))
    # 第一层全连接
    h1 = nd.relu(nd.dot(X, W1) + b1)
    # 在第一层全连接后添加丢弃层
    h1 = dropout(h1, drop_prob1)
    # 第二层全连接
    h2 = nd.relu(nd.dot(h1, W2) + b2)
    # 在第二层全连接后添加丢弃层
    h2 = dropout(h2, drop_prob2)
    # 返回输出
    return nd.dot(h2, W3) + b3

# 定义交叉熵损失
softmax_cross_entropy = gluon.loss.SoftmaxCrossEntropyLoss()

# 定义学习率
learning_rate = 0.5

# 训练
for epoch in range(5):
    # 训练损失
    train_loss = 0.0
    # 训练准确率
    train_acc = 0.0
    # 迭代训练
    for data, label in train_data:
        with autograd.record():
            # 计算输出
            output = net(data)
            # 计算损失
            loss = softmax_cross_entropy(output, label)
        # 梯度反向传播
        loss.backward()
        # SGD更新梯度
        SGD(params, learning_rate / batch_size)
        # 记录训练损失
        train_loss += nd.mean(loss).asscalar()
        # 记录训练准确率
        train_acc += accuracy(output, label)
    # 计算测试准确率
    test_acc = evaluate_accuracy(test_data, net)
    print("Epoch %d. Loss: %f, Train acc %f, Test acc %f" % (epoch, train_loss / len(train_data), train_acc / len(train_data), test_acc))

Epoch 0. Loss: 1.221062, Train acc 0.528746, Test acc 0.754006
Epoch 1. Loss: 0.598503, Train acc 0.774890, Test acc 0.813101
Epoch 2. Loss: 0.499490, Train acc 0.818493, Test acc 0.840244
Epoch 3. Loss: 0.457343, Train acc 0.832699, Test acc 0.835036
Epoch 4. Loss: 0.426575, Train acc 0.846070, Test acc 0.849159