【CNN从零开始系列之三】网络的训练

训练模型

前面一章中已经定义好了网络模型，我们就用定义好的模型和数据集开始训练。具体实现如下：

avg_loss = -1
base_lr = 0.01
for i in range(10000):
    learning_rate = base_lr * (1 + 0.0001 * i) ** (-0.75)
    if i % 100 == 0:
        print("Testing ...")
        data.set_mode(data.TEST)
        acc_sum = 0
        for j in range(100):
            data.next_batch_test_data()
            accuracy = calc_accuracy(softmax.forward(), data.get_label())
            acc_sum = acc_sum + accuracy
        acc_sum = acc_sum / 100
        print("Accuracy = %.4f" % (acc_sum))
    data.set_mode(data.TRAIN)
    data.next_batch_train_data()
    softmax.forward()
    loss = softmax.calc_loss(data.get_label())
    if avg_loss == -1:
        avg_loss = loss
    else:
        avg_loss = avg_loss * 0.9 + 0.1 * loss
    print("iter = %-5d\tLoss = %.4f\tAvg loss = %.4f\tLearning rate = %f" % (i + 1, loss, avg_loss, learning_rate))
    softmax.backward(learning_rate)

#训练过程如下
最后能达到98%的准确度。

项目demo

源码链接