MXNet是基礎，Gluon是封裝，兩者猶如TensorFlow和Keras，不過得益於動態圖機制，兩者交互比TensorFlow和Keras要方便得多，其基礎操作和pytorch極為相似，但是方便不少，有pytorch基礎入門會很簡單。

庫導入寫法，

from mxnet import ndarray as nd
from mxnet import autograd
from mxnet import gluon
import mxnet as mx

MXNet

mxnet.ndarray是整個科學計算系統的基礎，整體API和numpy的nparray一致，這一點類似於pytorch，不過不同於pytorch內置變量、張量等不同數據類型，mxnet簡化了只有ndarray一種，通過mxnet.autograd

自動求導

x = nd.arange(4).reshape((4, 1))
# 標記需要自動求導的量
x.attach_grad()
# 有自動求導就需要記錄計算圖
with autograd.record():
    y = 2 * nd.dot(x.T, x)
# 反向傳播輸出
y.backward()
# 獲取梯度
print(‘x.grad: ‘, x.grad)

nd轉化為數字

nd.asscalar()

nd與np數組互化

y = nd.array(x) # NumPy轉換成NDArray。

z = y.asnumpy() # NDArray轉換成NumPy。

節約內存的加法

nd.elemwise_add(x, y, out=z)

層實現

relu激活

def relu(X):
    return nd.maximum(X, 0)

全連接層

# 變量生成
w = nd.random.normal(scale=1, shape=(num_inputs, 1))
b = nd.zeros(shape=(1,))
params = [w, b]

# 變量掛載梯度
for param in params:
    param.attach_grad()

# 實現全連接
def net(X, w, b):
    return nd.dot(X, w) + b

SGD實現

def sgd(params, lr, batch_size):
    for param in params:
        param[:] = param - lr * param.grad / batch_size

Gluon

內存數據集加載

import mxnet as mx
from mxnet import autograd, nd
import numpy as np

num_inputs = 2
num_examples = 1000
true_w = [2, -3.4]
true_b = 4.2
features = nd.random.normal(scale=1, shape=(num_examples, num_inputs))
labels = true_w[0] * features[:, 0] + true_w[1] * features[:, 1] + true_b
labels += nd.random.normal(scale=0.01, shape=labels.shape)


from mxnet.gluon import data as gdata

batch_size = 10
dataset = gdata.ArrayDataset(features, labels)
data_iter = gdata.DataLoader(dataset, batch_size, shuffle=True)


for X, y in data_iter:
    print(X, y)
    break

[[-1.74047375  0.26071024]
 [ 0.65584248 -0.50490594]
 [-0.97745866 -0.01658815]
 [-0.55589193  0.30666101]
 [-0.61393601 -2.62473822]
 [ 0.82654613 -0.00791582]
 [ 0.29560572 -1.21692061]
 [-0.35985938 -1.37184834]
 [-1.69631028 -1.74014604]
 [ 1.31199837 -1.96280086]]
<NDArray 10x2 @cpu(0)>
[ -0.14842382   7.22247267   2.30917668   2.0601418   11.89551163
   5.87866735   8.94194221   8.15139961   6.72600317  13.50252151]
<NDArray 10 @cpu(0)>

模型定義

• 序列模型生成
• 層填充
• 初始化模型參數

net = gluon.nn.Sequential()
with net.name_scope():
    net.add(gluon.nn.Dense(1))
net.collect_params().initialize(mx.init.Normal(sigma=1))  # 模型參數初始化選擇normal分布

優化器

wd參數為模型添加了L2正則化，機制為：w = w - lr*grad - wd*w

trainer = gluon.Trainer(net.collect_params(), ‘sgd‘, {
        ‘learning_rate‘: learning_rate, ‘wd‘: weight_decay})

trainer.step(batch_size)需要運行在每一次反向傳播之後，會更新參數，一次模擬的訓練過程如下，

for e in range(epochs):
        for data, label in data_iter_train:
            with autograd.record():
                output = net(data)
                loss = square_loss(output, label)
            loss.backward()
            trainer.step(batch_size)
        train_loss.append(test(net, X_train, y_train))
        test_loss.append(test(net, X_test, y_test))

層函數API

拉伸

nn.Flatten()

全連接層

gluon.nn.Dense(256, activation="relu")

參數表示輸出節點數

損失函數class API

交叉熵

loss = gloss.SoftmaxCrossEntropyLoss()

『MXNet』第一彈_基礎操作以及常見層實現