機器學習筆記(5)：多類邏輯回歸-手動添加隱藏層

了解神經網絡原理的同學們應該都知道，隱藏層越多，最終預測結果的准確度越高，但是計算量也越大，在上一篇的基礎上，我們手動添加一個隱藏層，代碼如下（主要參考自多層感知機 — 從0開始）：

from mxnet import gluon
from mxnet import ndarray as nd
import matplotlib.pyplot as plt
import mxnet as mx
from mxnet import autograd
 
def transform(data, label):
    return data.astype('float32')/255, label.astype('float32')
 
mnist_train = gluon.data.vision.FashionMNIST(train=True, transform=transform)
mnist_test = gluon.data.vision.FashionMNIST(train=False, transform=transform)
 
def show_images(images):
    n = images.shape[0]
    _, figs = plt.subplots(1, n, figsize=(15, 15))
    for i in range(n):
        figs[i].imshow(images[i].reshape((28, 28)).asnumpy())
        figs[i].axes.get_xaxis().set_visible(False)
        figs[i].axes.get_yaxis().set_visible(False)
    plt.show()

def get_text_labels(label):
    text_labels = [
        'T 恤', '長 褲', '套頭衫', '裙 子', '外 套',
        '涼 鞋', '襯 衣', '運動鞋', '包 包', '短 靴'
    ]
    return [text_labels[int(i)] for i in label]
 
data, label = mnist_train[0:10]
 
print('example shape: ', data.shape, 'label:', label)
show_images(data)
print(get_text_labels(label))
 
batch_size = 256
train_data = gluon.data.DataLoader(mnist_train, batch_size, shuffle=True)
test_data = gluon.data.DataLoader(mnist_test, batch_size, shuffle=False)
 
num_inputs = 784
num_outputs = 10
 
#增加一層包含256個節點的隱藏層
num_hidden = 256
weight_scale = .01
 
#輸入層的參數
W1 = nd.random_normal(shape=(num_inputs, num_hidden), scale=weight_scale)
b1 = nd.zeros(num_hidden)
 
#隱藏層的參數
W2 = nd.random_normal(shape=(num_hidden, num_outputs), scale=weight_scale)
b2 = nd.zeros(num_outputs)
 
#參數變多了
params = [W1, b1, W2, b2]
 
for param in params:
    param.attach_grad()
    
#激活函數
def relu(X):
    return nd.maximum(X, 0)
 
#計算模型
def net(X):
    X = X.reshape((-1, num_inputs))
    #先計算到隱藏層的輸出
    h1 = relu(nd.dot(X, W1) + b1)
    #再利用隱藏層計算最終的輸出
    output = nd.dot(h1, W2) + b2
    return output
 
#Softmax和交叉熵損失函數
softmax_cross_entropy = gluon.loss.SoftmaxCrossEntropyLoss()

#梯度下降法
def SGD(params, lr):
    for param in params:
        param[:] = param - lr * param.grad
    
def accuracy(output, label):
    return nd.mean(output.argmax(axis=1) == label).asscalar()
 
def _get_batch(batch):
    if isinstance(batch, mx.io.DataBatch):
        data = batch.data[0]
        label = batch.label[0]
    else:
        data, label = batch
    return data, label
 
def evaluate_accuracy(data_iterator, net):
    acc = 0.
    if isinstance(data_iterator, mx.io.MXDataIter):
        data_iterator.reset()
    for i, batch in enumerate(data_iterator):
        data, label = _get_batch(batch)
        output = net(data)
        acc += accuracy(output, label)
    return acc / (i+1)
 
learning_rate = .5
 
for epoch in range(5):
    train_loss = 0.
    train_acc = 0.
    for data, label in train_data:
        with autograd.record():
            output = net(data)
            #使用Softmax和交叉熵損失函數
            loss = softmax_cross_entropy(output, label)
        loss.backward()
        SGD(params, learning_rate / batch_size)
        train_loss += nd.mean(loss).asscalar()
        train_acc += accuracy(output, label)
 
    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))

data, label = mnist_test[0:10]
show_images(data)
print('true labels')
print(get_text_labels(label))
 
predicted_labels = net(data).argmax(axis=1)
print('predicted labels')
print(get_text_labels(predicted_labels.asnumpy()))

 有變化的地方，都加了注釋，主要改動點有5個：

1. 手動添加了1個隱藏層，該層有256個節點

2. 多了一層，所以參數也變多了

3. 計算y=wx+b模型時，就要一層層來算了

4. 將softmax與交叉熵CrossEntropy合並了(這樣避免了單獨對softmax求導，理論上講更穩定些）

5. 另外激活函數換成了收斂速度更快的relu（參考：Deep learning系列（七）激活函數 ）

運行效果：

相對原始純手動版本，准確率提升了不少！

 

tips：類似的思路，我們可以再手動添加第2層隱藏層，關鍵代碼參考下面

...

#增加一層包含256個節點的隱藏層
num_hidden1 = 256
weight_scale1 = .01

#再增加一層包含512個節點的隱藏層
num_hidden2 = 512
weight_scale2 = .01
 
#輸入層的參數
W1 = nd.random_normal(shape=(num_inputs, num_hidden1), scale=weight_scale1)
b1 = nd.zeros(num_hidden1)
 
#隱藏層的參數
W2 = nd.random_normal(shape=(num_hidden1, num_hidden2), scale=weight_scale1)
b2 = nd.zeros(num_hidden2)

W3 = nd.random_normal(shape=(num_hidden2, num_outputs), scale=weight_scale2)
b3 = nd.zeros(num_outputs)
 
#參數變多了
params = [W1, b1, W2, b2, W3, b3]

...

#計算模型
def net(X):
    X = X.reshape((-1, num_inputs))
    #先計算到隱藏層的輸出
    h1 = relu(nd.dot(X, W1) + b1)
    h2 = relu(nd.dot(h1,W2) + b2)
    #再利用隱藏層計算最終的輸出
    output = nd.dot(h2, W3) + b3
    return output