MXnet代码实战之CNN

介绍

CNN指的是卷积神经网络，这个介绍网上资料多的很，我就不介绍了，我这里主要是针对沐神教程的CNN代码做一个笔记。理解有不对的地方欢迎指出。

卷积神经网络里面最重要也是最基本的概念就是卷积层、池化层、全连接层、卷积核、参数共享等。

图：

这个图是对下面代码的一个描述，对于一张图片，首先处理成28*28，通过第一层卷积层，得到20个通道的输出（每个输出为24*24），通过第一个池化层，得到20个12*12的输出。第二层卷积层：输入有20个通道，输出有50个通道，所以这里有50个filter（卷积核），每一个filter都会去卷积20个通道，然后求和，得到一个输出。所以50个filter就得到50个输出啦（输出为10*10）。通过第二个池化层得到50个5*5的输出。然后就是flatten层，意思是把它们拉开成一条向量，所以就是5*5*50=1250啦。后面又接了二个全连接层，最终输出为1*10，正好对应10类（有多少类别都就通过全连接层处理成多少类）。其实这个就是常用的LeNet。

从0开始实现CNN

代码：

#!/usr/bin/env python
# -*- coding:utf-8 -*-
#Author: yuquanle
#2017/10/28
#沐神教程实战之CNN
# 其实就是简单的LeNet雏形
#本例子使用FashionMNIST数据集

from mxnet import nd
from mxnet import gluon

# 处理输入数据
def transform_mnist(data, label):
    # change data from height x weight x channel to channel x height x weight
    return nd.transpose(data.astype('float32'), (2,0,1))/255, label.astype('float32')


def load_data_fashion_mnist(batch_size, transform=transform_mnist):
    """download the fashion mnist dataest and then load into memory"""
    mnist_train = gluon.data.vision.FashionMNIST(
        train=True, transform=transform)
    mnist_test = gluon.data.vision.FashionMNIST(
        train=False, transform=transform)
    train_data = gluon.data.DataLoader(
        mnist_train, batch_size, shuffle=True)
    test_data = gluon.data.DataLoader(
        mnist_test, batch_size, shuffle=False)
    return (train_data, test_data)

# 载入数据
batch_size = 256
train_data, test_data = load_data_fashion_mnist(batch_size)

import mxnet as mx
# 优先使用GPU运算
try:
    ctx = mx.gpu()
    _ = nd.zeros((1,), ctx=ctx)
except:
    ctx = mx.cpu()


weight_scale = .01
# output channels = 20, kernel = (5,5)
W1 = nd.random_normal(shape=(20,1,5,5), scale=weight_scale, ctx=ctx)
b1 = nd.zeros(W1.shape[0], ctx=ctx)
# output channels = 50, kernel = (3,3)
W2 = nd.random_normal(shape=(50,20,3,3), scale=weight_scale, ctx=ctx)
b2 = nd.zeros(W2.shape[0], ctx=ctx)
# output dim = 128
W3 = nd.random_normal(shape=(1250, 128), scale=weight_scale, ctx=ctx)
b3 = nd.zeros(W3.shape[1], ctx=ctx)
# output dim = 10
W4 = nd.random_normal(shape=(W3.shape[1], 10), scale=weight_scale, ctx=ctx)
b4 = nd.zeros(W4.shape[1], ctx=ctx)
params = [W1, b1, W2, b2, W3, b3, W4, b4]
for param in params:
    param.attach_grad()

# 定义模型
def net(X, verbose=False):
    X = X.as_in_context(W1.context)
    # 第⼀层卷积，kernel为5*5，filter有20个，所以卷积输出有20个通道
    h1_conv = nd.Convolution( data=X, weight=W1, bias=b1, kernel=W1.shape[2:], num_filter=W1.shape[0])
    # 激活函数
    h1_activation = nd.relu(h1_conv)
    # 池化
    h1 = nd.Pooling(data=h1_activation, pool_type="max", kernel=(2, 2), stride=(2, 2))
    # 第⼆层卷积，kernel为3*3，filter有50个，所以卷积输出有50个通道。输入有20个通道，所以每一个filter会卷积20个通道，然后求和
    h2_conv = nd.Convolution(data=h1, weight=W2, bias=b2, kernel=W2.shape[2:], num_filter=W2.shape[0])
    h2_activation = nd.relu(h2_conv)
    h2 = nd.Pooling(data=h2_activation, pool_type="max", kernel=(2, 2), stride=(2, 2))
    # flatten成向量
    h2 = nd.flatten(h2)
    # 第⼀层全连接
    h3_linear = nd.dot(h2, W3) + b3
    h3 = nd.relu(h3_linear)
    # 第⼆层全连接
    h4_linear = nd.dot(h3, W4) + b4

    if verbose:
        print('X:',X.shape)
        print('h1_conv:', h1_conv.shape)
        print('1st conv block:', h1.shape)
        print('h2_conv:', h2_conv.shape)
        print('2nd conv block:', h2.shape)
        print('1st dense:', h3.shape)
        print('2nd dense:', h4_linear.shape)
        print('output:', h4_linear)
    return h4_linear

# 查看信息
for data, _ in train_data:
    print(data.shape)
    net(data, verbose=True)
    break


def evaluate_accuracy(data_iterator, net, ctx=mx.cpu()):
    acc = 0.
    for data, label in data_iterator:
        output = net(data.as_in_context(ctx))
        acc += accuracy(output, label.as_in_context(ctx))
    return acc / len(data_iterator)

from mxnet import autograd as autograd
from utils import SGD, accuracy
from mxnet import gluon
#
#
softmax_cross_entropy = gluon.loss.SoftmaxCrossEntropyLoss()
learning_rate = .2
for epoch in range(5):
    train_loss = 0.
    train_acc = 0.
    for data, label in train_data:
        label = label.as_in_context(ctx)
        with autograd.record():
            output = net(data)
            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, ctx)
    print("Epoch %d. Loss: %f, Train acc %f, Test acc %f" % (
            epoch, train_loss / len(train_data),
            train_acc / len(train_data), test_acc))

使用Gluon

代码：

#!/usr/bin/env python
# -*- coding:utf-8 -*-
#Author: yuquanle
#2017/10/28
#沐神教程实战之CNN，使用Gluon
#本例子使用FashionMNIST数据集

from mxnet import gluon
from time import time

# 处理输入数据
def transform_mnist(data, label):
    # change data from height x weight x channel to channel x height x weight
    return nd.transpose(data.astype('float32'), (2,0,1))/255, label.astype('float32')

def load_data_fashion_mnist(batch_size, transform=transform_mnist):
    """download the fashion mnist dataest and then load into memory"""
    mnist_train = gluon.data.vision.FashionMNIST(
        train=True, transform=transform)
    mnist_test = gluon.data.vision.FashionMNIST(
        train=False, transform=transform)
    train_data = gluon.data.DataLoader(
        mnist_train, batch_size, shuffle=True)
    test_data = gluon.data.DataLoader(
        mnist_test, batch_size, shuffle=False)
    return (train_data, test_data)

# 获取数据
batch_size = 256
train_data, test_data = load_data_fashion_mnist(batch_size)

# 定义模型
from mxnet.gluon import nn
net = nn.Sequential()
with net.name_scope():
    net.add(
        nn.Conv2D(channels=20, kernel_size=5, activation='relu'),
        nn.MaxPool2D(pool_size=2, strides=2),
        nn.Conv2D(channels=50, kernel_size=3, activation='relu'),
        nn.MaxPool2D(pool_size=2, strides=2),
        nn.Flatten(),
        nn.Dense(128, activation="relu"),
        nn.Dense(10)
    )

# 初始化
from mxnet import autograd
from mxnet import nd
import mxnet as mx
# 优先使用GPU运算
try:
    ctx = mx.gpu()
    _ = nd.zeros((1,), ctx=ctx)
except:
    ctx = mx.cpu()

net.initialize(ctx=ctx)
print('initialize weight on', ctx)



def _get_batch(batch, ctx):
    """return data and label on ctx"""
    if isinstance(batch, mx.io.DataBatch):
        data = batch.data[0]
        label = batch.label[0]
    else:
        data, label = batch
    return (gluon.utils.split_and_load(data, ctx),
            gluon.utils.split_and_load(label, ctx),
            data.shape[0])

def evaluate_accuracy(data_iterator, net, ctx=[mx.cpu()]):
    if isinstance(ctx, mx.Context):
        ctx = [ctx]
    acc = nd.array([0])
    n = 0.
    if isinstance(data_iterator, mx.io.MXDataIter):
        data_iterator.reset()
    for batch in data_iterator:
        data, label, batch_size = _get_batch(batch, ctx)
        for X, y in zip(data, label):
            acc += nd.sum(net(X).argmax(axis=1)==y).copyto(mx.cpu())
        acc.wait_to_read() # don't push too many operators into backend
        n += batch_size
    return acc.asscalar() / n


def train(train_data, test_data, net, loss, trainer, ctx, num_epochs, print_batches=None):
    """Train a network"""
    print("Start training on ", ctx)
    if isinstance(ctx, mx.Context):
        ctx = [ctx]
    for epoch in range(num_epochs):
        train_loss, train_acc, n = 0.0, 0.0, 0.0
        if isinstance(train_data, mx.io.MXDataIter):
            train_data.reset()
        start = time()
        for i, batch in enumerate(train_data):
            data, label, batch_size = _get_batch(batch, ctx)
            losses = []
            with autograd.record():
                outputs = [net(X) for X in data]
                losses = [loss(yhat, y) for yhat, y in zip(outputs, label)]
            for l in losses:
                l.backward()
            train_acc += sum([(yhat.argmax(axis=1)==y).sum().asscalar()
                              for yhat, y in zip(outputs, label)])
            train_loss += sum([l.sum().asscalar() for l in losses])
            trainer.step(batch_size)
            n += batch_size
            if print_batches and (i+1) % print_batches == 0:
                print("Batch %d. Loss: %f, Train acc %f" % (
                    n, train_loss/n, train_acc/n
                ))

        test_acc = evaluate_accuracy(test_data, net, ctx)
        print("Epoch %d. Loss: %.3f, Train acc %.2f, Test acc %.2f, Time %.1f sec" % (
            epoch, train_loss/n, train_acc/n, test_acc, time() - start
        ))


# 训练和测试
loss = gluon.loss.SoftmaxCrossEntropyLoss()
trainer = gluon.Trainer(net.collect_params(),'sgd', {'learning_rate': 0.5})
train(train_data, test_data, net, loss, trainer, ctx, num_epochs=5)


结果：
initialize weight on gpu(0)
Start training on  gpu(0)
[10:50:57] d:\program files (x86)\jenkins\workspace\mxnet\mxnet\src\operator\./cudnn_algoreg-inl.h:106: Running performance tests to find the best convolution algorithm, this can take a while... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
Epoch 0. Loss: 1.140, Train acc 0.57, Test acc 0.80, Time 14.0 sec
Epoch 1. Loss: 0.457, Train acc 0.83, Test acc 0.82, Time 13.3 sec
Epoch 2. Loss: 0.380, Train acc 0.86, Test acc 0.77, Time 17.2 sec
Epoch 3. Loss: 0.342, Train acc 0.87, Test acc 0.87, Time 17.2 sec
Epoch 4. Loss: 0.311, Train acc 0.88, Test acc 0.88, Time 16.9 sec

Process finished with exit code 0