Pytorch开发者手册[更新中...]

Pytorch开发者手册

pytorch维度解释
input: seq_lenbatch_sizeinput_size

self.embedding=torch.nn.Embedding(input_size,emb_size)

output：seq_lenbatch_sizehidden_size

self.gru=torch.nn.GRU(emb_size,hidden_size,num_layers,  bidirectional=True,dropout=0.8)

output,hidden=self.gru(embedd)

output: seq_lenbatch_size (hidden_sizenum_directions)
hidden:(n_layersnum_directions) * batch_size *hidden_size

一 、函数
（1）nn.MaxPool1d
class torch.nn.MaxPool1d(kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)

kernel_size(int or tuple) - max pooling的窗口大小

stride(int or tuple, optional) - max pooling的窗口移动的步长。默认值是kernel_size

padding(int or tuple, optional) - 输入的每一条边补充0的层数

dilation(int or tuple, optional) – 一个控制窗口中元素步幅的参数

return_indices - 如果等于True，会返回输出最大值的序号，对于上采样操作会有帮助

ceil_mode - 如果等于True，计算输出信号大小的时候，会使用向上取整，代替默认的向下取整的操作

eg:

import torch
import torch.nn as nun

m = nn.MaxPool1d(3, stride=2)
input = torch.randn(2, 4, 5)
output = m(input)

print(input)
print(output)

tensor([[[ 1.0000, -0.2401, -1.1694,  0.3162,  0.5495],
         [-0.0668, -0.1455, -0.6538,  0.7396, -1.2040],
         [ 1.9750,  0.9449,  0.6117,  0.9990, -0.8750],
         [ 0.5644, -0.2105,  0.1103,  1.2369,  0.3041]],

        [[-0.6587,  0.5686, -0.4883,  1.4393,  2.0761],
         [ 0.7883,  0.9219,  3.0924,  0.0957,  0.3198],
         [ 1.3996,  0.7220,  0.8044,  1.4365, -0.9084],
         [ 0.8209,  0.6308,  1.7780, -0.2451, -0.7157]]])

 tensor([[[ 1.0000,  0.5495],
         [-0.0668,  0.7396],
         [ 1.9750,  0.9990],
         [ 0.5644,  1.2369]],

        [[ 0.5686,  2.0761],
         [ 3.0924,  3.0924],
         [ 1.3996,  1.4365],
         [ 1.7780,  1.7780]]])

二、Tensor维度变换
维度变换的的主要函数有
torch.reshape() 或者torch.view()
torch.reshape() 可以改变张量的形状
torch.squeeze() 改变维度
torch.unsqueeze() 增加维度
torch.transpose() 可以交换维度

a=torch.Tensor(1,2,3,3)
print(a.size())
b=a.view([3,-1])
print(b.size())
c=torch.squeeze(a)
print(c.size())
e=torch.transpose(a,0,1)
print(e.size())
'''f=a.t()  只有小于等于2个维度才可以转置
print(f.size())'''
d=torch.unsqueeze(a,3)
print(d.size())

torch.Size([1, 2, 3, 3])
torch.Size([3, 6])
torch.Size([2, 3, 3])
torch.Size([2, 1, 3, 3])
torch.Size([1, 2, 3, 1, 3])

三、pytorch进行多GPU的计算

device=torch.device("cuda" if torch.cuda.is_available() else "cpu")
model=model.to(device)