以往的网络DNN(Dense稠密网络)

在序列数据中,处理数据过大,linear层比卷积核的运算类要大很多

RNN模型

h0先验条件,如果用于图像生成文本,可以在h0前面加上cnn+fc

g:三维到五维(h0三维,输出五维),本质上是线性层

用tanh是因为取值在+1和-1之间

pytorch中的RNN Cell

只需要输入特征数,和输出特征数就行了,因为本质上是一个线性层

RNN的输入即输出

利用numlayers构建多层

例子

step1 按字母序建立词典再转换为独热向量

因为文字非数字,无法计算,因此需要转换

inputsize最后一个表格的列数

输入向量是一个维度是4的独热向量,输出向量也是个维度是4的概率向量

step2 loss

参数设置

seq_len序列长度(x1, x2, x3)

input_size输入数据每一个(x1)都是一个四维的向量

hidden_size每个隐层都是有两个元素

数据定义

要把inputs和labels重新view,-1为自适应

inputs的格式为(seqlen, batchsize, inputsize)

lables的格式为 (seqlen,1)

seqlen其实就是循环次数

代码,将hello变成ohlol(RNNCell)

import torchinput_size = 4hidden_size = 4batch_size = 1idx2char = ['e', 'h', 'l', 'o']x_data = [1, 0, 2, 2, 3]y_data = [3, 1, 2, 3, 2]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]inputs = torch.Tensor(x_one_hot).view(-1, batch_size, input_size)lables = torch.LongTensor(y_data).view(-1, 1)class Model(torch.nn.Module):    def __init__(self, input_size, hidden_size, batch_size):        super(Model, self).__init__()        self.input_size = input_size        self.hidden_size = hidden_size        self.batch_size = batch_size        self.rnncell = torch.nn.RNNCell(self.input_size, self.hidden_size)    def forward(self, inputs, hidden):        hidden = self.rnncell(inputs, hidden)        return hidden    def init_hidden(self):        return torch.zeros(self.batch_size, self.hidden_size)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    hidden = net.init_hidden()    print('Predicted string: ', end='\n')    for input, lable in zip(inputs, lables):        hidden = net(input, hidden)        loss += criterion(hidden, lable)        _, idx = hidden.max(dim=1)        print(idx2char[idx.item()], end='')    optimizer.zero_grad()    loss.backward()    optimizer.step()    print(', Epoch[%d/15] loss = %.4f' %(epoch + 1, loss.item()))

代码,采用pytorch中的RNN

改变了out的维度

改变了lables的维度

import torchinput_size = 4hidden_size = 4batch_size = 1seq_len = 5idx2char = ['e', 'h', 'l', 'o']x_data = [1, 0, 2, 2, 3]y_data = [3, 1, 2, 3, 2]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]inputs = torch.Tensor(x_one_hot).view(seq_len, batch_size, input_size)lables = torch.LongTensor(y_data)class Model(torch.nn.Module):    def __init__(self, input_size, hidden_size, batch_size, num_layers=1):        super(Model, self).__init__()        self.input_size = input_size        self.hidden_size = hidden_size        self.batch_size = batch_size        self.num_layers = num_layers        self.rnn = torch.nn.RNN(input_size=self.input_size, hidden_size=self.hidden_size, num_layers=self.num_layers)    def forward(self, inputs):        hidden = torch.zeros(self.num_layers, self.batch_size, self.hidden_size)        out,_ = self.rnn(inputs, hidden)        return out.view(-1, self.hidden_size)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):    optimizer.zero_grad()    outputs = net(inputs)    loss = criterion(outputs, lables)    loss.backward()    optimizer.step()    _, idx = outputs.max(dim=1)    idx = idx.data.numpy()    print('Predicted: ', ''.join([idx2char[x] for x in idx]), end='')    print(',Epoch [%d / 15] loss = %.3f' %(epoch + 1, loss.item()))

采用Embedding vectors

独热向量降维为Embedding vectors

四维转换为5维

例如原来维度中的第二个,找到第二行,然后输出就行了

改变网络结构

改变维度