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")
x_train.shape #第一个表示样本个数,第二个表示特征28*28*1=784
(50000, 784)
784是mnist数据集每个样本的像素点个数
注意数据需转换成tensor才能参与后续建模训练
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())
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
import torch.nn.functional as F
loss_func = F.cross_entropy
def model(xb):
return xb.mm(weights) + bias
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) #随机初始化w
bs = 64
bias = torch.zeros(10, requires_grad=True) #偏置:常数或随机初始化,对结果影响较小
print(loss_func(model(xb), yb))
创建一个model来更简化代码
1.必须继承nn.Module且在其构造函数中需调用nn.Module的构造函数
2.无需写反向传播函数,nn.Module能够利用autograd自动实现反向传播
3.Module中的可学习参数可以通过named_parameters()或者parameters()返回迭代器
from torch import nn
class Mnist_NN(nn.Module):
def __init__(self): #构造函数
super().__init__()
self.hidden1 = nn.Linear(784, 128) #全连接层,输入784个像素点,输出128个特征
self.hidden2 = nn.Linear(128, 256) #第二个隐层,第一层的输出,是第二层的输入,
#256个神经元
self.out = nn.Linear(256, 10) #10个分类
self.dropout=nn.Dropout(0.5) #按照百分比 自己定
def forward(self, x): #前向传播 自己定义 反向传播是自动的
#输入x,是数据给的,batch 64*784
x = F.relu(self.hidden1(x)) #得到64*128的结果
x=self.dropout(x) #每个全连接层都要加dropout,随机杀除神经元,防止过拟合
x = F.relu(self.hidden2(x)) #得到65*256 的向量
x=self.dropout(x)
x = self.out(x) #输出层,256*10的矩阵
return x
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)
(dropout): Dropout(p=0.5, inplace=False)
)
可以打印我们定义好名字里的权重和偏置项
for name, parameter in net.named_parameters():
print(name, parameter,parameter.size())
使用TensorDataset和DataLoader来简化¶
from torch.utils.data import TensorDataset
from torch.utils.data import DataLoader
train_ds = TensorDataset(x_train, y_train) #做一个封装
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)
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
import numpy as np
def fit(steps, model, loss_func, opt, train_dl, valid_dl): #训练函数
#迭代 net 优化器
for step in range(steps):
model.train() #训练模式,更新每一层权置和参数
for xb, yb in train_dl: #打包好数据一个个去返
loss_batch(model, loss_func, xb, yb, opt) # xb, yb,输入的数据和标签
model.eval() #验证,不更新
with torch.no_grad():
#返回当前损失和数量,一个loss对应一个num
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) #平均损失
print('当前step:'+str(step), '验证集损失:'+str(val_loss))
#zip用法
a=[1,2,3]
b=[4,5,6]
zipped=zip(a,b)
print(list(zipped)) #[(1, 4), (2, 5), (3, 6)]
a2,b2=zip(*zip(a,b))
print(a2) #(1, 2, 3)
print(b2) #(4, 5, 6)
from torch import optim
def get_model():
model = Mnist_NN()
return model, optim.Adam(model.parameters(), lr=0.001) #返回 优化器 lr学习率
def loss_batch(model, loss_func, xb, yb, opt=None): #1.计算loss 2.更新w,b
loss = loss_func(model(xb), yb)# model(xb)把模型放输入当中,得到预测值 yb真实值
if opt is not None: #优化器
loss.backward() #反向传播,算出每一层的权置参数,梯度
opt.step() # 执行更新,沿着梯度方向更新
opt.zero_grad() #torch每次迭代累加记录,把之前迭代清空
return loss.item(), len(xb)
train_dl, valid_dl = get_data(train_ds, valid_ds, bs)
model, opt = get_model()
fit(20, model, loss_func, opt, train_dl, valid_dl)
当前step:0 验证集损失:0.18639104866981507
当前step:1 验证集损失:0.1372520131058991
当前step:2 验证集损失:0.12028736076653004
当前step:3 验证集损失:0.10732126496359706
当前step:4 验证集损失:0.10093651054650545
当前step:5 验证集损失:0.09517242526896298
当前step:6 验证集损失:0.09194612504523247
当前step:7 验证集损失:0.08943103497959673
当前step:8 验证集损失:0.0877913200291805
当前step:9 验证集损失:0.08958465236043557
当前step:10 验证集损失:0.08797709809066728
当前step:11 验证集损失:0.08352214076635428
当前step:12 验证集损失:0.0866958644344937
当前step:13 验证集损失:0.08074293819144368
当前step:14 验证集损失:0.08045687620015815
当前step:15 验证集损失:0.08040115665267222
当前step:16 验证集损失:0.07971061569196172
当前step:17 验证集损失:0.08116058965921402
当前step:18 验证集损失:0.0811522187425755
当前step:19 验证集损失:0.0807436868159566
correct=0
total=0
for xb,yb in valid_dl: #去验证集里取数据
outputs=model(xb) #128*10 每个样本属于各个类别的概率值
_, predicted=torch.max(outputs.data,1) #返回最大的值和索引,算概率值哪个大,沿着1这个维度(行)
total+=yb.size(0)
correct+=(predicted==yb).sum().item() #预测值=真实值.对了多少.转换成数组格式(可绘制图)
print('Accuracy of the network on the 10000 test images:%d %%' %(
100 * correct / total))
Accuracy of the network on the 10000 test images:97 %