记录一些pytorch的基本使用方法。
import torch
from torch.autograd import Variable # 自动求导变量
from torch import nn, optim # 网络层与优化器
from torch.utils.data import TensorDataset, DataLoader # 数据集迭代器
从列表生成:a = torch.tensor([1,2,3], dtype=int)
全1全0:torch.ones(shape),torch.zeros(shape)
整数序列[m,n):torch.arrange(m,n)
随机数[0,1):torch.rand(3,4)
随机整数[m,n):torch.randint(m,n,shape)
正态分布:torch.randn()
相同形状随机数:b = torch.rand_like(a, dtype=float)
维度:a.ndim;形状:a.shape,a.size()
数值:a[0].item()
类型转换:a.float(),a.type(torch.float32)
array = np.array(tensor)
tensor = torch.from_numpy(array)
tensor = torch.tensor(array)
b = a.view(shape)
b = a.reshape(shape)
a+b,torch.add(a,b[,out=c]), a.add_(b)
a-b,a*b,a/b, a%b,a//b,a**m
平方根:torch.sqrt(a)
矩阵乘法:torch.matmul(a, b)
转置:a.T
求和:torch.sum(a)
均值:.mean(a),中位数:.median(a)
最值:torch.max/min/argmax/argmin(a)
x = torch.ones(shape, requires_grad=True)
y = x * x
y.backward()
grad = x.grad
1、保存模型:
torch.save(model.state_dict(), 'model.pth')
2、载入模型
model.load_state_dict(torch.load('model.pth'))
# 层
nn.Linear(in_features, out_features)
nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding)
nn.MaxPool2d(kernel_size,stride)
nn.LSTM(input_size, hidden_size, num_layers, batch_first)
# 激活函数
self.tanh = nn.Tanh()
self.relu = nn.ReLU()
self.softmax = nn.Softmax(dim=1)
# 模型序列
self.fc = nn.Sequential(nn.Linear(784,128), nn.ReLU())
self.conv = nn.Sequential(nn.Conv2d(32,64,5,1,2), nn.ReLU(), nn.MaxPool2d(2,2))
self.lstm = nn.LSTM(input_size=28, hidden_size=64, num_layers=1, batch_first=True)
nn.MSELoss()
nn.CrossEntropyLoss()
optim.SGD(model.parameters(), lr=0.1)
optim.Adam(model.parameters(), lr=0.001)
pytorch的张量操作与numpy数组基本一致。模型保存只保存参数,加载时需要重新定义模型。采用继承类的形式来定义模型,在初始化中设置模型参数,在方法中进行前向传播,注意按层要求修改数据形状。