发布时间:2024-02-03 12:00
B站 刘二大人:循环神经网络(基础篇)
目录
1、RNN概念
2、numLayers含义
3、RNN使用
4、利用RNN Cell训练hello转换到ohlol
5、Embedding编码方式
RNN Cell是线性层。
隐层是RNN Cell里线性层矩阵w的行数。
使用RNN Cell:
import torch
batch_size = 1 # 批处理大小
seq_len = 3 # 序列长度
input_size = 4 # 输入维度
hidden_size = 2 # 隐层维度
cell = torch.nn.RNNCell(input_size=input_size, hidden_size=hidden_size) # 初始化
# (seq, batch, features)
dataset = torch.randn(seq_len, batch_size, input_size)
hidden = torch.zeros(batch_size, hidden_size)
# 这个循环就是处理seq_len长度的数据
for idx, data in enumerate(dataset):
print(\'=\' * 20, idx, \'=\' * 20)
print(\'Input size:\', data.shape, data)
hidden = cell(data, hidden)
print(\'hidden size:\', hidden.shape, hidden)
print(hidden)
input_size和hidden_size: 输入维度和隐层维度
batch_size: 批处理大小
seq_len: 序列长度
num_layers: 隐层数目
使用RNN:
import torch
batch_size = 1 # batch_size: 批处理大小
seq_len = 3 # seq_len: 序列长度
input_size = 4 # input_size:输入维度
hidden_size = 2 # hidden_size: 隐层维度
num_layers = 1 # num_layers: 隐层数目
cell = torch.nn.RNN(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers)
# (seqLen, batchSize, inputSize)
inputs = torch.randn(seq_len, batch_size, input_size)
hidden = torch.zeros(num_layers, batch_size, hidden_size)
out, hidden = cell(inputs, hidden)
print(\'Output size:\', out.shape) # (seq_len, batch_size, hidden_size)
print(\'Output:\', out)
print(\'Hidden size:\', hidden.shape) # (num_layers, batch_size, hidden_size)
print(\'Hidden:\', hidden)
代码如下:
import torch
input_size = 4
hidden_size = 4
batch_size = 1
idx2char = [\'e\', \'h\', \'l\', \'o\']
x_data = [1, 0, 2, 3, 3] # hello中各个字符的下标
y_data = [3, 1, 2, 3, 2] # ohlol中各个字符的下标
one_hot_lookup = [[1, 0, 0, 0],
[0, 1, 0, 0],
[0, 0, 1, 0],
[0, 0, 0, 1]]
x_one_hot = [one_hot_lookup[x] for x in x_data] # (seqLen, inputSize)
inputs = torch.Tensor(x_one_hot).view(-1, batch_size, input_size)
labels = torch.LongTensor(y_data).view(-1, 1)
# torch.Tensor默认是torch.FloatTensor是32位浮点类型数据,torch.LongTensor是64位整型
print(inputs.shape, labels.shape)
class Model(torch.nn.Module):
def __init__(self, input_size, hidden_size, batch_size):
super(Model, self).__init__()
self.batch_size = batch_size
self.input_size = input_size
self.hidden_size = hidden_size
self.rnncell = torch.nn.RNNCell(input_size=self.input_size, hidden_size=self.hidden_size)
def forward(self, inputs, hidden):
hidden = self.rnncell(inputs, hidden) # 输入和隐层转换为下一个隐层
# shape of inputs:(batchSize, inputSize),shape of hidden:(batchSize, hiddenSize),
return hidden
def init_hidden(self):
return torch.zeros(self.batch_size, self.hidden_size) # 生成全0的h0
net = Model(input_size, hidden_size, batch_size)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.1)
for epoch in range(15):
loss = 0
optimizer.zero_grad()
hidden = net.init_hidden()
print(\'Predicted string:\', end=\'\')
for input, label in zip(inputs, labels):
hidden = net(input, hidden)
# 注意交叉熵在计算loss的时候维度关系,这里的hidden是([1, 4]), label是 ([1])
loss += criterion(hidden, label)
_, idx = hidden.max(dim = 1)
print(idx2char[idx.item()], end=\'\')
loss.backward()
optimizer.step()
print(\', Epoch [%d/15] loss=%.4f\' % (epoch+1, loss.item()))
结果:
独热编码向量维度过高;
独热编码向量稀疏,每个向量是一个为1其余为0;
独热编码是硬编码,编码情况与数据特征无关;
采用一种低维度的、稠密的、可学习数据的编码方式:Embedding。
代码:
import torch
input_size = 4
num_class = 4
hidden_size = 8
embedding_size = 10
batch_size = 1
num_layers = 2
seq_len = 5
idx2char_1 = [\'e\', \'h\', \'l\', \'o\']
idx2char_2 = [\'h\', \'l\', \'o\']
x_data = [[1, 0, 2, 2, 3]]
y_data = [3, 1, 2, 2, 3]
# inputs 维度为(batchsize,seqLen)
inputs = torch.LongTensor(x_data)
# labels 维度为(batchsize*seqLen)
labels = torch.LongTensor(y_data)
class Model(torch.nn.Module):
def __init__(self):
super(Model, self).__init__()
self.emb = torch.nn.Embedding(input_size, embedding_size)
self.rnn = torch.nn.RNN(input_size=embedding_size,
hidden_size=hidden_size,
num_layers=num_layers,
batch_first=True)
self.fc = torch.nn.Linear(hidden_size, num_class)
def forward(self, x):
hidden = torch.zeros(num_layers, x.size(0), hidden_size)
x = self.emb(x) # 进行embedding处理
x, _ = self.rnn(x, hidden)
x = self.fc(x)
return x.view(-1, num_class)
net = Model()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.05)
for epoch in range(15):
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
_, idx = outputs.max(dim=1)
idx = idx.data.numpy()
print(\'Predicted string: \', \'\'.join([idx2char_1[x] for x in idx]), end=\'\')
print(\", Epoch [%d/15] loss = %.3f\" % (epoch + 1, loss.item()))
结果: