循环神经网络PyTorch版

最新推荐文章于 2024-07-09 09:04:37 发布

coder_jyz

最新推荐文章于 2024-07-09 09:04:37 发布

阅读量396

点赞数

分类专栏：深度学习文章标签：深度学习 python

本文链接：https://blog.csdn.net/coder_jyz/article/details/119523089

版权

循环神经网络(RNN)

申明：本书的理论和代码来自于开源书籍，TensorFlow深度学习

地址：https://github.com/jiajunhua/dragen1860-Deep-Learning-with-TensorFlow-book

1. Pytorch的RNN简单实现

2.1. 单层RNNCell

[s,b,n], s为句子长度，b为句子数量，n为词向量长度，h为状态长度

import  torch
from    torch import nn
from    torch import optim
from    torch.nn import functional as F
# [s,b,n] s:句长，b：句子数，n：词向量长度
x = torch.randn(80, 4, 100)
# param1：input_size  | param2：hidden_size 
cell1 = nn.RNNCell(100, 64)
# [batch, hidden_size]
h1 = torch.zeros(4, 64)
# xt [batch, input_size]
for xt in x:
    h1 = cell1(xt, h1)
print(h1.shape) # [4, 64]

2.2. 多层 RNNCell

x = torch.randn(80, 4, 100)
cell1 = nn.RNNCell(100, 64)
cell2 = nn.RNNCell(64, 20)
h1 = torch.zeros(4, 64)
h2 = torch.zeros(4, 20)
for xt in x:
    h1 = cell1(xt, h1)
    h2 = cell2(h1, h2)
print(h2.shape) # [4, 20]

2.3 RNN

# 单层
rnn = nn.RNN(input_size=100, hidden_size=64, num_layers=1)
x = torch.randn(80, 4, 100)
out, h = rnn(x, torch.zeros(1, 4, 64))
print(out.shape, h.shape) # torch.Size([80, 4, 64]) torch.Size([1, 4, 64])

# 多层
rnn = nn.RNN(input_size=100, hidden_size=64, num_layers=3)
x = torch.randn(80, 4, 100)
out, h = rnn(x, torch.zeros(3, 4, 64))
print(out.shape, h.shape)