瞭解神經網路原理的同學們應該都知道，隱藏層越多，最終預測結果的準確度越高，但是計算量也越大，在上一篇的基礎上，我們手動新增一個隱藏層，程式碼如下（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
#輸入層的引數
W1 = nd.random_normal(shape=(num_inputs, num_hidden), scale=weight_scale)
b1 = nd.zeros(num_hidden)
#隱藏層的引數
W2 = nd.random_normal(shape=(num_hidden, num_outputs), scale=weight_scale)
b2 = nd.zeros(num_outputs)
#引數變多了
params = [W1, b1, W2, b2]
for param in params:
param.attach_grad()
#啟用函式
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和交叉熵損失函式
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和交叉熵損失函式
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，都加了註釋，主要改動點有5個：
1. 手動新增了1個隱藏層，該層有256個節點
2. 多了一層，所以引數也變多了
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合併了(這樣避免了單獨對softmax求導，理論上講更穩定些）
5. 另外啟用函式換成了收斂速度更快的relu（參考：Deep learning系列（七）啟用函式 ）
執行效果：
相對原始純手動版本，準確率提升了不少！
tips：類似的思路，我們可以再手動新增第2層隱藏層，關鍵程式碼參考下面
...
#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
#輸入層的引數
W1 = nd.random_normal(shape=(num_inputs, num_hidden1), scale=weight_scale1)
b1 = nd.zeros(num_hidden1)
#隱藏層的引數
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)
#引數變多了
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