PaddlePaddle 学习心得（第三周学习）

进入第三周学习后，开始使用完整案例进行训练测试。

与自学的最大区别是有小伙伴一起学习会效率更好。

先看结果

 

本次作业是使用衰减学习率，通过多次调参数。

查找 API得知有提供多个api可选择点击查看

学习率调度器¶

当我们使用诸如梯度下降法等方式来训练模型时，一般会兼顾训练速度和损失(loss)来选择相对合适的学习率。但若在训练过程中一直使用一个学习率，训练集的损失下降到一定程度后便不再继续下降，而是在一定范围内震荡。其震荡原理如下图所示，即当损失函数收敛到局部极小值附近时，会由于学习率过大导致更新步幅过大，每步参数更新会反复越过极小值而出现震荡。

这个做法是为了更快下降到稳定值，减小震荡

 

本次选择了使用piecewise_decay这个api进行测试。

代码节选如下：

def trainadj(model):
    countindex=0
    boundaries = [40, 80]
    values = [0.5, 0.1,0.05]
    with fluid.dygraph.guard():
        print('start training ... ')
        model.train()
        epoch_num = 2
        # 定义优化器
        opt = fluid.optimizer.Momentum(
            learning_rate=fluid.layers.piecewise_decay(boundaries=boundaries, values=values),
            momentum=0.9,
            parameter_list=model.parameters(),
            regularization=fluid.regularizer.L2Decay(1e-4)
            )
        # 定义数据读取器，训练数据读取器和验证数据读取器
        train_loader = data_loader(DATADIR, batch_size=10, mode='train')
        valid_loader = valid_data_loader(DATADIR2, CSVFILE)
        for epoch in range(epoch_num):
            for batch_id, data in enumerate(train_loader()):
                x_data, y_data = data
                img = fluid.dygraph.to_variable(x_data)
                label = fluid.dygraph.to_variable(y_data)
                # 运行模型前向计算，得到预测值
                logits = model(img)
                # 进行loss计算
                loss = fluid.layers.sigmoid_cross_entropy_with_logits(logits, label)
                avg_loss = fluid.layers.mean(loss)

                #if batch_id % 10 == 0:

                print("{} #:epoch: {}, batch_id: {}, loss is: {}".format(opt.current_step_lr(),epoch, batch_id, avg_loss.numpy()))
                #lossindex.append(countindex)
                lossdataadj.append(avg_loss.numpy())
                countindex=countindex+1
                # 反向传播，更新权重，清除梯度
                avg_loss.backward()
                opt.minimize(avg_loss)
                model.clear_gradients()

            model.eval()
            accuracies = []
            losses = []
            for batch_id, data in enumerate(valid_loader()):
                x_data, y_data = data
                img = fluid.dygraph.to_variable(x_data)
                label = fluid.dygraph.to_variable(y_data)
                # 运行模型前向计算，得到预测值
                logits = model(img)
                # 二分类，sigmoid计算后的结果以0.5为阈值分两个类别
                # 计算sigmoid后的预测概率，进行loss计算
                pred = fluid.layers.sigmoid(logits)
                loss = fluid.layers.sigmoid_cross_entropy_with_logits(logits, label)
                # 计算预测概率小于0.5的类别
                pred2 = pred * (-1.0) + 1.0
                # 得到两个类别的预测概率，并沿第一个维度级联
                pred = fluid.layers.concat([pred2, pred], axis=1)
                acc = fluid.layers.accuracy(pred, fluid.layers.cast(label, dtype='int64'))
                accuracies.append(acc.numpy())
                losses.append(loss.numpy())
            print("[validation] accuracy/loss: {}/{}".format(np.mean(accuracies), np.mean(losses)))
            model.train()

        # save params of model
        fluid.save_dygraph(model.state_dict(), 'palm')
        # save optimizer state
        fluid.save_dygraph(opt.state_dict(), 'palm')

关键修改部分：

原代码优化器部分：

  # 定义优化器
        opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9, parameter_list=model.parameters())

修改为

#参数部分

boundaries = [40, 80]
values = [0.5, 0.1,0.05]

# 定义优化器
        opt = fluid.optimizer.Momentum(
            learning_rate=fluid.layers.piecewise_decay(boundaries=boundaries, values=values),
            momentum=0.9,
            parameter_list=model.parameters(),
            regularization=fluid.regularizer.L2Decay(1e-4)
            )

运行后可观察到loss开始时变动很大，后期就平稳下降到，小范围调整

0.5 #:epoch: 0, batch_id: 0, loss is: [0.7019217]
0.5 #:epoch: 0, batch_id: 1, loss is: [12.024451]
0.5 #:epoch: 0, batch_id: 2, loss is: [157.16641]
0.5 #:epoch: 0, batch_id: 3, loss is: [95.9813]
0.5 #:epoch: 0, batch_id: 4, loss is: [4.2797537]
0.5 #:epoch: 0, batch_id: 5, loss is: [7.3096175]
0.5 #:epoch: 0, batch_id: 6, loss is: [4.384579]
0.5 #:epoch: 0, batch_id: 7, loss is: [0.4799278]
0.5 #:epoch: 0, batch_id: 8, loss is: [7.04195]
0.5 #:epoch: 0, batch_id: 9, loss is: [2.5763474]
0.5 #:epoch: 0, batch_id: 10, loss is: [0.71913904]
0.5 #:epoch: 0, batch_id: 11, loss is: [0.6301126]
0.5 #:epoch: 0, batch_id: 12, loss is: [13.080458]
0.5 #:epoch: 0, batch_id: 13, loss is: [0.5763507]
0.5 #:epoch: 0, batch_id: 14, loss is: [0.4996148]
0.5 #:epoch: 0, batch_id: 15, loss is: [0.26528448]
0.5 #:epoch: 0, batch_id: 16, loss is: [19.885382]
0.5 #:epoch: 0, batch_id: 17, loss is: [1.20226]
0.5 #:epoch: 0, batch_id: 18, loss is: [8.281421]
0.5 #:epoch: 0, batch_id: 19, loss is: [10.232064]
0.5 #:epoch: 0, batch_id: 20, loss is: [1.1040335]
0.5 #:epoch: 0, batch_id: 21, loss is: [13.693789]
0.5 #:epoch: 0, batch_id: 22, loss is: [0.7570822]
0.5 #:epoch: 0, batch_id: 23, loss is: [22.65204]
0.5 #:epoch: 0, batch_id: 24, loss is: [0.73202336]
0.5 #:epoch: 0, batch_id: 25, loss is: [0.5780096]
0.5 #:epoch: 0, batch_id: 26, loss is: [11.400303]
0.5 #:epoch: 0, batch_id: 27, loss is: [9.84152]
0.5 #:epoch: 0, batch_id: 28, loss is: [0.60879284]
0.5 #:epoch: 0, batch_id: 29, loss is: [1.6898936]
0.5 #:epoch: 0, batch_id: 30, loss is: [2.4156432]
0.5 #:epoch: 0, batch_id: 31, loss is: [1.0774876]
0.5 #:epoch: 0, batch_id: 32, loss is: [0.34068987]
0.5 #:epoch: 0, batch_id: 33, loss is: [1.7986888]
0.5 #:epoch: 0, batch_id: 34, loss is: [0.7363323]
0.5 #:epoch: 0, batch_id: 35, loss is: [0.6592141]
0.5 #:epoch: 0, batch_id: 36, loss is: [0.42114687]
0.5 #:epoch: 0, batch_id: 37, loss is: [0.74656934]
0.5 #:epoch: 0, batch_id: 38, loss is: [0.7371367]
0.5 #:epoch: 0, batch_id: 39, loss is: [2.3921227]

后期可对

boundaries = [40, 80]
values = [0.5, 0.1,0.05]

这两部分变量进行修改高度，再观察效果即可。

当然，以上结果并不是最好的情况，上述只是记录本次调试参数过程。 

 

最后，百度paddlepaddle学习课程对入门了解与使用带来了很大帮助，值得推荐！