pytorch神经网络分类

Mnist分类任务：

  • 网络基本构建与训练方法，常用函数解析

  • torch.nn.functional模块

  • nn.Module模块

读取Mnist数据集

  • 会自动进行下载
1
%matplotlib inline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
from pathlib import Path
import requests

DATA_PATH = Path("data")
PATH = DATA_PATH / "mnist"

PATH.mkdir(parents=True, exist_ok=True)

URL = "http://deeplearning.net/data/mnist/"
FILENAME = "mnist.pkl.gz"

if not (PATH / FILENAME).exists():
        content = requests.get(URL + FILENAME).content
        (PATH / FILENAME).open("wb").write(content)
1
2
3
4
5
import pickle
import gzip

with gzip.open((PATH / FILENAME).as_posix(), "rb") as f:
        ((x_train, y_train), (x_valid, y_valid), _) = pickle.load(f, encoding="latin-1")

784是mnist数据集每个样本的像素点个数

1
2
3
4
5
from matplotlib import pyplot
import numpy as np

pyplot.imshow(x_train[0].reshape((28, 28)), cmap="gray")
print(x_train.shape)
(50000, 784)

png

FAO FAO

注意数据需转换成tensor才能参与后续建模训练

1
2
3
4
5
6
7
8
9
10
import torch

x_train, y_train, x_valid, y_valid = map(
    torch.tensor, (x_train, y_train, x_valid, y_valid)
)
n, c = x_train.shape
x_train, x_train.shape, y_train.min(), y_train.max()
print(x_train, y_train)
print(x_train.shape)
print(y_train.min(), y_train.max())
tensor([[0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        ...,
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.],
        [0., 0., 0.,  ..., 0., 0., 0.]]) tensor([5, 0, 4,  ..., 8, 4, 8])
torch.Size([50000, 784])
tensor(0) tensor(9)

torch.nn.functional 很多层和函数在这里都会见到

torch.nn.functional中有很多功能，后续会常用的。那什么时候使用nn.Module，什么时候使用nn.functional呢？一般情况下，如果模型有可学习的参数，最好用nn.Module，其他情况nn.functional相对更简单一些

1
2
3
4
5
6
import torch.nn.functional as F

loss_func = F.cross_entropy

def model(xb):
    return xb.mm(weights) + bias
1
2
3
4
5
6
7
8
bs = 64
xb = x_train[0:bs]  # a mini-batch from x
yb = y_train[0:bs]
weights = torch.randn([784, 10], dtype = torch.float,  requires_grad = True) 
bs = 64
bias = torch.zeros(10, requires_grad=True)

print(loss_func(model(xb), yb))
tensor(10.7988, grad_fn=<NllLossBackward>)

创建一个model来更简化代码

  • 必须继承nn.Module且在其构造函数中需调用nn.Module的构造函数
  • 无需写反向传播函数，nn.Module能够利用autograd自动实现反向传播
  • Module中的可学习参数可以通过named_parameters()或者parameters()返回迭代器
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
from torch import nn

class Mnist_NN(nn.Module):
    def __init__(self):
        super().__init__()
        self.hidden1 = nn.Linear(784, 128)
        self.hidden2 = nn.Linear(128, 256)
        self.out  = nn.Linear(256, 10)

    def forward(self, x):
        x = F.relu(self.hidden1(x))
        x = F.relu(self.hidden2(x))
        x = self.out(x)
        return x
        
1
2
3
net = Mnist_NN()
print(net)

Mnist_NN(
  (hidden1): Linear(in_features=784, out_features=128, bias=True)
  (hidden2): Linear(in_features=128, out_features=256, bias=True)
  (out): Linear(in_features=256, out_features=10, bias=True)
)

可以打印我们定义好名字里的权重和偏置项

1
2
for name, parameter in net.named_parameters():
    print(name, parameter,parameter.size())
hidden1.weight Parameter containing:
tensor([[ 0.0018,  0.0218,  0.0036,  ..., -0.0286, -0.0166,  0.0089],
        [-0.0349,  0.0268,  0.0328,  ...,  0.0263,  0.0200, -0.0137],
        [ 0.0061,  0.0060, -0.0351,  ...,  0.0130, -0.0085,  0.0073],
        ...,
        [-0.0231,  0.0195, -0.0205,  ..., -0.0207, -0.0103, -0.0223],
        [-0.0299,  0.0305,  0.0098,  ...,  0.0184, -0.0247, -0.0207],
        [-0.0306, -0.0252, -0.0341,  ...,  0.0136, -0.0285,  0.0057]],
       requires_grad=True) torch.Size([128, 784])
hidden1.bias Parameter containing:
tensor([ 0.0072, -0.0269, -0.0320, -0.0162,  0.0102,  0.0189, -0.0118, -0.0063,
        -0.0277,  0.0349,  0.0267, -0.0035,  0.0127, -0.0152, -0.0070,  0.0228,
        -0.0029,  0.0049,  0.0072,  0.0002, -0.0356,  0.0097, -0.0003, -0.0223,
        -0.0028, -0.0120, -0.0060, -0.0063,  0.0237,  0.0142,  0.0044, -0.0005,
         0.0349, -0.0132,  0.0138, -0.0295, -0.0299,  0.0074,  0.0231,  0.0292,
        -0.0178,  0.0046,  0.0043, -0.0195,  0.0175, -0.0069,  0.0228,  0.0169,
         0.0339,  0.0245, -0.0326, -0.0260, -0.0029,  0.0028,  0.0322, -0.0209,
        -0.0287,  0.0195,  0.0188,  0.0261,  0.0148, -0.0195, -0.0094, -0.0294,
        -0.0209, -0.0142,  0.0131,  0.0273,  0.0017,  0.0219,  0.0187,  0.0161,
         0.0203,  0.0332,  0.0225,  0.0154,  0.0169, -0.0346, -0.0114,  0.0277,
         0.0292, -0.0164,  0.0001, -0.0299, -0.0076, -0.0128, -0.0076, -0.0080,
        -0.0209, -0.0194, -0.0143,  0.0292, -0.0316, -0.0188, -0.0052,  0.0013,
        -0.0247,  0.0352, -0.0253, -0.0306,  0.0035, -0.0253,  0.0167, -0.0260,
        -0.0179, -0.0342,  0.0033, -0.0287, -0.0272,  0.0238,  0.0323,  0.0108,
         0.0097,  0.0219,  0.0111,  0.0208, -0.0279,  0.0324, -0.0325, -0.0166,
        -0.0010, -0.0007,  0.0298,  0.0329,  0.0012, -0.0073, -0.0010,  0.0057],
       requires_grad=True) torch.Size([128])
hidden2.weight Parameter containing:
tensor([[-0.0383, -0.0649,  0.0665,  ..., -0.0312,  0.0394, -0.0801],
        [-0.0189, -0.0342,  0.0431,  ..., -0.0321,  0.0072,  0.0367],
        [ 0.0289,  0.0780,  0.0496,  ...,  0.0018, -0.0604, -0.0156],
        ...,
        [-0.0360,  0.0394, -0.0615,  ...,  0.0233, -0.0536, -0.0266],
        [ 0.0416,  0.0082, -0.0345,  ...,  0.0808, -0.0308, -0.0403],
        [-0.0477,  0.0136, -0.0408,  ...,  0.0180, -0.0316, -0.0782]],
       requires_grad=True) torch.Size([256, 128])
hidden2.bias Parameter containing:
tensor([-0.0694, -0.0363, -0.0178,  0.0206, -0.0875, -0.0876, -0.0369, -0.0386,
         0.0642, -0.0738, -0.0017, -0.0243, -0.0054,  0.0757, -0.0254,  0.0050,
         0.0519, -0.0695,  0.0318, -0.0042, -0.0189, -0.0263, -0.0627, -0.0691,
         0.0713, -0.0696, -0.0672,  0.0297,  0.0102,  0.0040,  0.0830,  0.0214,
         0.0714,  0.0327, -0.0582, -0.0354,  0.0621,  0.0475,  0.0490,  0.0331,
        -0.0111, -0.0469, -0.0695, -0.0062, -0.0432, -0.0132, -0.0856, -0.0219,
        -0.0185, -0.0517,  0.0017, -0.0788, -0.0403,  0.0039,  0.0544, -0.0496,
         0.0588, -0.0068,  0.0496,  0.0588, -0.0100,  0.0731,  0.0071, -0.0155,
        -0.0872, -0.0504,  0.0499,  0.0628, -0.0057,  0.0530, -0.0518, -0.0049,
         0.0767,  0.0743,  0.0748, -0.0438,  0.0235, -0.0809,  0.0140, -0.0374,
         0.0615, -0.0177,  0.0061, -0.0013, -0.0138, -0.0750, -0.0550,  0.0732,
         0.0050,  0.0778,  0.0415,  0.0487,  0.0522,  0.0867, -0.0255, -0.0264,
         0.0829,  0.0599,  0.0194,  0.0831, -0.0562,  0.0487, -0.0411,  0.0237,
         0.0347, -0.0194, -0.0560, -0.0562, -0.0076,  0.0459, -0.0477,  0.0345,
        -0.0575, -0.0005,  0.0174,  0.0855, -0.0257, -0.0279, -0.0348, -0.0114,
        -0.0823, -0.0075, -0.0524,  0.0331,  0.0387, -0.0575,  0.0068, -0.0590,
        -0.0101, -0.0880, -0.0375,  0.0033, -0.0172, -0.0641, -0.0797,  0.0407,
         0.0741, -0.0041, -0.0608,  0.0672, -0.0464, -0.0716, -0.0191, -0.0645,
         0.0397,  0.0013,  0.0063,  0.0370,  0.0475, -0.0535,  0.0721, -0.0431,
         0.0053, -0.0568, -0.0228, -0.0260, -0.0784, -0.0148,  0.0229, -0.0095,
        -0.0040,  0.0025,  0.0781,  0.0140, -0.0561,  0.0384, -0.0011, -0.0366,
         0.0345,  0.0015,  0.0294, -0.0734, -0.0852, -0.0015, -0.0747, -0.0100,
         0.0801, -0.0739,  0.0611,  0.0536,  0.0298, -0.0097,  0.0017, -0.0398,
         0.0076, -0.0759, -0.0293,  0.0344, -0.0463, -0.0270,  0.0447,  0.0814,
        -0.0193, -0.0559,  0.0160,  0.0216, -0.0346,  0.0316,  0.0881, -0.0652,
        -0.0169,  0.0117, -0.0107, -0.0754, -0.0231, -0.0291,  0.0210,  0.0427,
         0.0418,  0.0040,  0.0762,  0.0645, -0.0368, -0.0229, -0.0569, -0.0881,
        -0.0660,  0.0297,  0.0433, -0.0777,  0.0212, -0.0601,  0.0795, -0.0511,
        -0.0634,  0.0720,  0.0016,  0.0693, -0.0547, -0.0652, -0.0480,  0.0759,
         0.0194, -0.0328, -0.0211, -0.0025, -0.0055, -0.0157,  0.0817,  0.0030,
         0.0310, -0.0735,  0.0160, -0.0368,  0.0528, -0.0675, -0.0083, -0.0427,
        -0.0872,  0.0699,  0.0795, -0.0738, -0.0639,  0.0350,  0.0114,  0.0303],
       requires_grad=True) torch.Size([256])
out.weight Parameter containing:
tensor([[ 0.0232, -0.0571,  0.0439,  ..., -0.0417, -0.0237,  0.0183],
        [ 0.0210,  0.0607,  0.0277,  ..., -0.0015,  0.0571,  0.0502],
        [ 0.0297, -0.0393,  0.0616,  ...,  0.0131, -0.0163, -0.0239],
        ...,
        [ 0.0416,  0.0309, -0.0441,  ..., -0.0493,  0.0284, -0.0230],
        [ 0.0404, -0.0564,  0.0442,  ..., -0.0271, -0.0526, -0.0554],
        [-0.0404, -0.0049, -0.0256,  ..., -0.0262, -0.0130,  0.0057]],
       requires_grad=True) torch.Size([10, 256])
out.bias Parameter containing:
tensor([-0.0536,  0.0007,  0.0227, -0.0072, -0.0168, -0.0125, -0.0207, -0.0558,
         0.0579, -0.0439], requires_grad=True) torch.Size([10])

使用TensorDataset和DataLoader来简化

1
2
3
4
5
6
7
8
from torch.utils.data import TensorDataset
from torch.utils.data import DataLoader

train_ds = TensorDataset(x_train, y_train)
train_dl = DataLoader(train_ds, batch_size=bs, shuffle=True)

valid_ds = TensorDataset(x_valid, y_valid)
valid_dl = DataLoader(valid_ds, batch_size=bs * 2)
1
2
3
4
5
def get_data(train_ds, valid_ds, bs):
    return (
        DataLoader(train_ds, batch_size=bs, shuffle=True),
        DataLoader(valid_ds, batch_size=bs * 2),
    )
  • 一般在训练模型时加上model.train()，这样会正常使用Batch Normalization和 Dropout
  • 测试的时候一般选择model.eval()，这样就不会使用Batch Normalization和 Dropout
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
import numpy as np

def fit(steps, model, loss_func, opt, train_dl, valid_dl):
    for step in range(steps):
        model.train()
        for xb, yb in train_dl:
            loss_batch(model, loss_func, xb, yb, opt)

        model.eval()
        with torch.no_grad():
            losses, nums = zip(
                *[loss_batch(model, loss_func, xb, yb) for xb, yb in valid_dl]
            )
        val_loss = np.sum(np.multiply(losses, nums)) / np.sum(nums)
        print('当前step:'+str(step), '验证集损失：'+str(val_loss))
1
2
3
4
from torch import optim
def get_model():
    model = Mnist_NN()
    return model, optim.SGD(model.parameters(), lr=0.001)
1
2
3
4
5
6
7
8
9
def loss_batch(model, loss_func, xb, yb, opt=None):
    loss = loss_func(model(xb), yb)

    if opt is not None:
        loss.backward()
        opt.step()
        opt.zero_grad()

    return loss.item(), len(xb)

三行搞定！

1
2
3
train_dl, valid_dl = get_data(train_ds, valid_ds, bs)
model, opt = get_model()
fit(25, model, loss_func, opt, train_dl, valid_dl)
当前step:0 验证集损失：2.2796445930480957
当前step:1 验证集损失：2.2440698066711424
当前step:2 验证集损失：2.1889826164245605
当前step:3 验证集损失：2.0985311767578123
当前step:4 验证集损失：1.9517273582458496
当前step:5 验证集损失：1.7341805934906005
当前step:6 验证集损失：1.4719875366210937
当前step:7 验证集损失：1.2273896869659424
当前step:8 验证集损失：1.0362271406173706
当前step:9 验证集损失：0.8963696184158325
当前step:10 验证集损失：0.7927186088562012
当前step:11 验证集损失：0.7141492074012756
当前step:12 验证集损失：0.6529350900650024
当前step:13 验证集损失：0.60417300491333
当前step:14 验证集损失：0.5643046331882476
当前step:15 验证集损失：0.5317994566917419
当前step:16 验证集损失：0.5047958114624024
当前step:17 验证集损失：0.4813900615692139
当前step:18 验证集损失：0.4618900228500366
当前step:19 验证集损失：0.4443243554592133
当前step:20 验证集损失：0.4297310716629028
当前step:21 验证集损失：0.416976597738266
当前step:22 验证集损失：0.406348459148407
当前step:23 验证集损失：0.3963301926612854
当前step:24 验证集损失：0.38733808159828187

1

pytorch神经网络分类
https://zhangfuli.github.io/2020/09/02/pytorch神经网络分类/
作者
张富利
发布于
2020年9月2日
许可协议