【动手学PaddlePaddle2.0系列】模型训练的N种姿势

【动手学Paddle2.0系列】模型训练的N种姿势

最开始接触深度学习的时候，我的码力几乎等于0，所以在最初的时候，几乎都是使用封装好的高层API进行训练，自由度很低。

想通过这个教程，对paddle2.0中的各种开启训练的方式进行一个总结。让我们开始愉快的学（ban）习（zhuan）吧！

1 Paddle2.0 主要思想

那么，用框架来类比，飞桨框架基础API对应方法一，飞桨框架高层API对应方法二。使用基础API，我们可以随心所欲的搭建自己的深度学习模型，不会受到任何限制；而使用方法二，我们可以很快的实现模型，达到自己想要的效果，缺点是少了一些自主性。但是，与制作披萨不同的是，飞桨框架可以做到真正的”鱼与熊掌”可以兼得，我们在飞桨框架中实现了API的高低融合，使我们的开发者既可以享受到基础API的强大，又可以兼顾高层API的快捷。

1.1 数据加载

from paddle.vision.transforms import Compose, Normalize
from paddle.vision.datasets import MNIST


# 数据预处理，这里用到了随机调整亮度、对比度和饱和度
transform = Compose([Normalize(mean=[127.5],
                               std=[127.5],
                               data_format='CHW')])


# 数据加载，在训练集上应用数据预处理的操作
train_dataset = MNIST(mode='train', transform=transform)
test_dataset = MNIST(mode='test')

Cache file /home/aistudio/.cache/paddle/dataset/mnist/train-images-idx3-ubyte.gz not found, downloading https://dataset.bj.bcebos.com/mnist/train-images-idx3-ubyte.gz 
Begin to download

Download finished
Cache file /home/aistudio/.cache/paddle/dataset/mnist/train-labels-idx1-ubyte.gz not found, downloading https://dataset.bj.bcebos.com/mnist/train-labels-idx1-ubyte.gz 
Begin to download
........
Download finished
Cache file /home/aistudio/.cache/paddle/dataset/mnist/t10k-images-idx3-ubyte.gz not found, downloading https://dataset.bj.bcebos.com/mnist/t10k-images-idx3-ubyte.gz 
Begin to download

Download finished
Cache file /home/aistudio/.cache/paddle/dataset/mnist/t10k-labels-idx1-ubyte.gz not found, downloading https://dataset.bj.bcebos.com/mnist/t10k-labels-idx1-ubyte.gz 
Begin to download
..
Download finished

1.2 加载模型

import paddle
from paddle.vision.models import LeNet
import paddle.nn as nn

mnist = LeNet()
# 模型封装，用Model类封装
model = paddle.Model(mnist)

# 模型配置：为模型训练做准备，设置优化器，损失函数和精度计算方式
model.prepare(optimizer=paddle.optimizer.Adam(parameters=model.parameters()),
              loss=nn.CrossEntropyLoss(),
              metrics=paddle.metric.Accuracy())

2 模型训练的N种姿势

2.1 高层API训练

2.1.1 方式一

# 模型训练，
model.fit(train_dataset,
          epochs=5,
          batch_size=64,
          verbose=1)

# 模型评估，
model.evaluate(test_dataset, verbose=1)

Epoch 1/5
step  30/938 [..............................] - loss: 0.5914 - acc: 0.6250 - ETA: 11s - 12ms/st

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:77: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  return (isinstance(seq, collections.Sequence) and


step  70/938 [=>............................] - loss: 0.4902 - acc: 0.7636 - ETA: 6s - 8ms/stepstep 938/938 [==============================] - loss: 0.0476 - acc: 0.9470 - 5ms/step        
Epoch 2/5
step 938/938 [==============================] - loss: 0.0125 - acc: 0.9803 - 5ms/step         
Epoch 3/5
step 938/938 [==============================] - loss: 0.0026 - acc: 0.9844 - 5ms/step         
Epoch 4/5
step 938/938 [==============================] - loss: 0.0061 - acc: 0.9869 - 5ms/step         
Epoch 5/5
step 938/938 [==============================] - loss: 0.0837 - acc: 0.9885 - 5ms/step         
Eval begin...
step 10000/10000 [==============================] - loss: 0.0000e+00 - acc: 0.9684 - 2ms/step         
Eval samples: 10000





{'loss': [0.0], 'acc': 0.9684}

2.1.2 方式二

# 构建训练集数据加载器
train_loader = paddle.io.DataLoader(train_dataset, batch_size=64, shuffle=True)

# 构建测试集数据加载器
test_loader = paddle.io.DataLoader(test_dataset, batch_size=64, shuffle=True)

model.fit(train_loader,
        eval_data = test_loader,
        epochs=5,
        verbose=1,
        )

Epoch 1/5
step 938/938 [==============================] - loss: 0.0493 - acc: 0.9898 - 5ms/step         
Eval begin...
step 157/157 [==============================] - loss: 0.0000e+00 - acc: 0.9672 - 3ms/step     
Eval samples: 10000
Epoch 2/5
step 938/938 [==============================] - loss: 0.0285 - acc: 0.9911 - 5ms/step         
Eval begin...
step 157/157 [==============================] - loss: 0.0000e+00 - acc: 0.9388 - 3ms/step     
Eval samples: 10000
Epoch 3/5
step 938/938 [==============================] - loss: 0.0627 - acc: 0.9913 - 6ms/step         
Eval begin...
step 157/157 [==============================] - loss: 4.0000 - acc: 0.9466 - 3ms/step         
Eval samples: 10000
Epoch 4/5
step 938/938 [==============================] - loss: 0.0184 - acc: 0.9922 - 5ms/step         
Eval begin...
step 157/157 [==============================] - loss: 0.0000e+00 - acc: 0.9529 - 3ms/step     
Eval samples: 10000
Epoch 5/5
step 938/938 [==============================] - loss: 0.0035 - acc: 0.9927 - 5ms/step        
Eval begin...
step 157/157 [==============================] - loss: 3.8146 - acc: 0.9229 - 3ms/step         
Eval samples: 10000

model.evaluate(test_dataset, verbose=1)

Eval begin...
step 10000/10000 [==============================] - loss: 0.0000e+00 - acc: 0.9229 - 2ms/step        
Eval samples: 10000





{'loss': [0.0], 'acc': 0.9229}

2.1.3 方式三

for epoch in range(5):
    for batch_id, data in enumerate(train_loader()):
        x_data = data[0]
        y_data = data[1]
        info = model.train_batch([x_data], [y_data])

        if batch_id % 1 == 0:
            print(info)

    for ebatch_id, edata in enumerate(test_loader()):
        x_data_v = edata[0]
        y_data_v = edata[1]
        info = model.eval_batch([x_data_v], [y_data_v])

        if ebatch_id % 1 == 0:
            print(info)

model.evaluate(test_dataset, verbose=1)

Eval begin...
step 10000/10000 [==============================] - loss: 0.0000e+00 - acc: 0.9880 - 2ms/step        
Eval samples: 10000





{'loss': [0.0], 'acc': 0.9880055555555556}

2.1.4 方式四

import paddle.nn.functional as F


mnist.train()
epochs = 5
optim = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())
# 用Adam作为优化函数
for epoch in range(epochs):
    for batch_id, data in enumerate(train_loader()):
        x_data = data[0]
        y_data = data[1]
       
        predicts = mnist(x_data)
        loss = F.cross_entropy(predicts, y_data)
        # 计算损失
        acc = paddle.metric.accuracy(predicts, y_data, k=2)
        loss.backward()
        if batch_id % 1 == 0:
            print("epoch: {}, batch_id: {}, loss is: {}, acc is: {}".format(epoch, batch_id, loss.numpy(), acc.numpy()))
        optim.step()
        optim.clear_grad()
    for batch_id, data in enumerate(test_loader()):
        x_data = data[0]
        y_data = data[1]
        
        predicts = mnist(x_data)
        loss = F.cross_entropy(predicts, y_data)
        # 计算损失
        acc = paddle.metric.accuracy(predicts, y_data, k=2)
        if batch_id % 1 == 0:
    acc = paddle.metric.accuracy(predicts, y_data, k=2)
        if batch_id % 1 == 0:
            print("epoch: {}, batch_id: {}, loss is: {}, acc is: {}".format(epoch, batch_id, loss.numpy(), acc.numpy()))

总结

在上面的四种训练方式中，实际上每种方式都有其优缺点，完全使用高层API能够方便的开启训练。对于完全的新手用户，推荐使用方式一、方式二。对于已经熟悉Paddle2.0的用户推荐使用方式四。方式四可以更加灵活，并且能够灵活地加入混合精度训练，训练方式更加自由。

一点小小的宣传&下期预告

我会在下一个项目中介绍在Paddle2.0中使用混合精度训练，欢迎大家关注哦。

我目前在上海，感兴趣的领域包括模型压缩、小目标检测、嵌入式，欢迎交流关注。来AI Studio互粉吧_等你哦