手写数字识别(mxnet官网例子)

.手写数字识别简介：通过MNIST数据集建立一个手写数字分类器。（MNIST对于手写数据分类任务是一个广泛使用的数据集）。1.前提：mxnet 0.10及以上、python、jupyter notebook（有时间可以jupyter notebook的用法，如：PPT的制作）pip install requests jupyter python下jupyter notebook的安装2.加载数据集：import mxnet as mxmnist = mx.test_utils.get_mnist()此时MXNET数据集已完全加载到内存中（注：此法对于大型数据集不适用）考虑要素：快速高效地从源直接流数据+输入样本的顺序图像通常用4维数组来表示：（batch_size，num_channels，width，height）对于MNIST数据集，因为是28*28灰度图像，所以只有1个颜色通道，width=28，height=28，本例中batch=100（批处理100），即输入形状是（batch_size，1，28，28）数据迭代器通过随机的调整输入来解决连续feed相同样本的问题。测试数据的顺序无关紧要。batch_size = 100train_iter=mx.io.NDArrayIter(mnisttrain_data,mnisttrain_label, batch_size, shuffle=True)val_iter = mx.io.NDArrayIter(mnisttest_data, mnisttest_label, batch_size)初始化MNIST数据集的数据迭代器（2个：训练数据+测试数据）3.训练+预测：（2种方法） （CNN优于MLP） 1）传统深度神经网络结构MLP（多层神经网络）MLPMXNET的符号接口为输入的数据创建一个占位符变量 data = mx.sym.var(data)data = mx.sym.flatten(data=data)将数据从4维变成2维(batch_size,num_channel*width*height)fc1 = mx.sym.FullyConnected(data=data, num_hidden=128)act1 = mx.sym.Activation(data=fc1, act_type=relu)第一个全连接层及相应的激活函数fc2 = mx.sym.FullyConnected(data=act1, num_hidden = 64)act2 = mx.sym.Activation(data=fc2, act_type=relu)第二个全连接层及相应的激活函数（声明2个全连接层，每层有128个和64个神经元）fc3 = mx.sym.FullyConnected(data=act2, num_hidden=10)声明大小10的最终完全连接层mlp = mx.sym.SoftmaxOutput(data=fc3, name=softmax)softmax的交叉熵损失 MNIST的MLP网络结构以上，已完成了数据迭代器和神经网络的申明，下面可以进行训练。超参数：处理大小、学习速率import logginglogging.getLogger().setLevel(logging.DEBUG) 记录到标准输出mlp_model = mx.mod.Module(symbol=mlp, context=mx.cpu()在CPU上创建一个可训练的模块mlp_model.fit(train_iter 训练数据 eval_data=val_iter, 验证数据 optimizer=sgd, 使用SGD训练 optimizer_params=learning_rate:0.1, 使用固定的学习速率 eval_metric=acc, 训练过程中报告准确性 batch_end_callback=mx.callback.Speedometer(batch_size, 100), 每批次100数据输出的进展 num_epoch=10) 训练至多通过10个数据预测：test_iter = mx.io.NDArrayIter(mnisttest_data, None, batch_size)prob = mlp_model.predict(test_iter)assert prob.shape = (10000, 10)计算每一个测试图像可能的预测得分（probij第i个测试图像包含j输出类）test_iter = mx.io.NDArrayIter(mnisttest_data, mnisttest_label, batch_size) 预测精度的方法acc = mx.metric.Accuracy()mlp_model.score(test_iter, acc)print(acc)assert acc.get()1 0.96如果一切顺利的话，我们将看到一个准确的值大约是0.96，这意味着我们能够准确地预测96%的测试图像中的数字。 2）卷积神经网络（CNN） 卷积层+池化层data = mx.sym.var(data)conv1 = mx.sym.Convolution(data=data, kernel=(5,5), num_filter=20)tanh1 = mx.sym.Activation(data=conv1, act_type=tanh)pool1 = mx.sym.Pooling(data=tanh1, pool_type=max, kernel=(2,2), stride=(2,2)第一个卷积层、池化层conv2 = mx.sym.Convolution(data=pool1, kernel=(5,5), num_filter=50) 第二个卷积层tanh2 = mx.sym.Activation(data=conv2, act_type=tanh)pool2 = mx.sym.Pooling(data=tanh2, pool_type=max, kernel=(2,2), stride=(2,2)flatten = mx.sym.flatten(data=pool2) 第一个全连接层fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=500)tanh3 = mx.sym.Activation(data=fc1, act_type=tanh)fc2 = mx.sym.FullyConnected(data=tanh3, num_hidden=10)第二个全连接层lenet = mx.sym.SoftmaxOutput(data=fc2, name=softmax) Softmax损失 LeNet第一个卷积层+池化层lenet_model = mx.mod.Module(symbol=lenet, context=mx.cpu()在GPU上创建一个可训练的模块lenet_model.fit(train_iter, eval_data=val_iter, optimizer=sgd, optimizer_params=learning_rate:0.1, eval_metric=acc, batch_end_callback = mx.callback.Speedometer(batch_size, 100), num_epoch=10)训练（同MLP）test_iter = mx.io.NDArrayIter(mnisttest_data, None, batch_size)prob = lenet_model.predict(test_iter)test_iter = mx.io.NDArrayIter(mnisttest_data, mnisttest_label, batch_size)acc = mx.metric.Accuracy() 预测LeNet的准确性lenet_model.score(test_iter, acc)print(acc)assert acc.get()1 0.98使用CNN，我们能够正确地预测所有测试图像的98%左右。附：完整代码1)MLPimport mxnet as mxmnist = mx.test_utils.get_mnist()batch_size = 100train_iter = mx.io.NDArrayIter(mnisttrain_data, mnisttrain_label, batch_size, shuffle=True)val_iter = mx.io.NDArrayIter(mnisttest_data, mnisttest_label, batch_size)data = mx.sym.var(data)data = mx.sym.flatten(data=data)fc1 = mx.sym.FullyConnected(data=data, num_hidden=128)act1 = mx.sym.Activation(data=fc1, act_type=relu)fc2 = mx.sym.FullyConnected(data=act1, num_hidden = 64)act2 = mx.sym.Activation(data=fc2, act_type=relu)fc3 = mx.sym.FullyConnected(data=act2, num_hidden=10)mlp = mx.sym.SoftmaxOutput(data=fc3, name=softmax)import logginglogging.getLogger().setLevel(logging.DEBUG) mlp_model = mx.mod.Module(symbol=mlp, context=mx.cpu()mlp_model.fit(train_iter, eval_data=val_iter, optimizer=sgd, optimizer_params=learning_rate:0.1, eval_metric=acc, batch_end_callback = mx.callback.Speedometer(batch_size, 100), num_epoch=10) test_iter = mx.io.NDArrayIter(mnisttest_data, None, batch_size)prob = mlp_model.predict(test_iter)assert prob.shape = (10000, 10)test_iter = mx.io.NDArrayIter(mnisttest_data, mnisttest_label, batch_size)acc = mx.metric.Accuracy()mlp_model.score(test_iter, acc)print(acc)assert acc.get()1 0.962)CNNimport mxnet as mxmnist = mx.test_utils.get_mnist()batch_size = 100train_iter = mx.io.NDArrayIter(mnisttrain_data, mnisttrain_label, batch_size, shuffle=True)val_iter = mx.io.NDArrayIter(mnisttest_data, mnisttest_label, batch_size)data = mx.sym.var(data)conv1 = mx.sym.Convolution(data=data, kernel=(5,5), num_filter=20)tanh1 = mx.sym.Activation(data=conv1, act_type=tanh)pool1 = mx.sym.Pooling(data=tanh1, pool_type=max, kernel=(2,2), stride=(2,2)conv2 = mx.sym.Convolution(data=pool1, kernel=(5,5), num_filter=50)tanh2 = mx.sym.Activation(data=conv2, act_type=tanh)pool2 = mx.sym.Pooling(data=tanh2, pool_type=max, kernel=(2,2), stride=(2,2)flatten = mx.sym.flatten(data=pool2)fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=500)tanh3 = mx.sym.Activation(data=fc1, act_type=tanh)fc2 = mx.sym.FullyConnected(data=tanh3, num_hidden=10)lenet = mx.sym.SoftmaxOutput(data=fc2, name=softmax)lenet_model = mx.mod.Module(symbol=lenet, context=mx.cpu()lenet_model.fit(train_iter, eval_data=val_iter, optimizer=sgd, optimizer_params=learning_rate:0.1, eval_metric=acc, batch_end_callback = mx.callback.Speedometer(batch_size, 100), num_epoch=10)test_iter = mx.io.NDArrayIter(mnisttest_data, None, batch_size)prob = lenet_model.predict(test_iter)test_iter = mx.io.NDArrayIter(mnisttest_data, mnisttest_label, batch_size)acc = mx.metric.Accuracy()lenet_model.score(test_iter, acc)print(acc)assert acc.get()1 0.98.