刘二大人《PyTorch深度学习实践》完结合集——第12课：循环神经网络（基础篇）——课后作业

油条与肉包

已于 2023-10-21 11:09:03 修改

阅读量236

点赞数

文章标签：深度学习 pytorch rnn

于 2023-10-20 18:14:09 首次发布

本文链接：https://blog.csdn.net/qq_44738211/article/details/133929229

版权

1：使用RNNcell+onehot vector

2：RNNcell+embedding层

3：RNN+embedding效果不好就不跑了

4：LSTM+onehot vector

5：LSTM+Embedding

6：使用GRUcell+embedding

7：使用GRU +embedding

8：总结：

1：使用RNNcell+onehot vector

# 使用RNN+onehot
import torch
import matplotlib.pyplot as plt

input_size = 4
hidden_size =4
batch_size=1

idx2char = ['e','h','l','o']
x_data = [1,0,2,2,3]           #hello
y_data = [3,1,2,3,2]           #ohlol

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] #x_one_hot是list:[[0, 1, 0, 0], [1, 0, 0, 0], [0, 0, 1, 0], [0, 0, 1, 0], [0, 0, 0, 1]]
inputs = torch.Tensor(x_one_hot).view(-1,batch_size,input_size) #torch.Size([5, 1, 4])
labels = torch.LongTensor(y_data).view(-1,1) #torch.Size([5, 1])

class Model(torch.nn.Module):
    def __init__(self):
        super(Model,self).__init__()
        self.rnncell = torch.nn.RNNCell(input_size=input_size,hidden_size=hidden_size)

    def forward(self,input,hidden):
        hidden = self.rnncell(input,hidden)
        return hidden
    def init_hidden(self):                                              #构建初始的隐层h0 [batch_size,hidden_size]
        return torch.zeros(batch_size,hidden_size)

model = Model()

#Construct loss functions and optimizer.................Use Torch API...................................................#
criterion  =  torch.nn.CrossEntropyLoss()
optimizer =   torch.optim.Adam(model.parameters(),lr=0.1) #lr为学习率，因为0.01太小了，我改成了0.1

#plot
x_axis = []
y_axis = []

for epoch in range(15):
    optimizer.zero_grad()
    hidden = model.init_hidden()    #hidden_size = torch.tensor([1,4])
    loss = 0
    print('Predicted string:',end='')
    for input,label in zip(inputs,labels):              #torch.Size([1, 4]) torch.Size([1])
        hidden = model(input,hidden)                    #epoch=1:按seqlen第一次返回h1_1,第二次返回h1_1,h1_2,....返回所有的hidden
        loss += criterion(hidden,label)                 #label 在交叉熵损失中会自动转化为one_hot vector
        _,idx = hidden.max(dim=1)                       #按dim=1就是batch_size或者说按行取最大值和最大值所在的下标索引值，
        print(idx2char[idx.item()],end='')
    loss.backward()
    optimizer.step()
    print(', Epoch [%d/15] loss=%.4f' % (epoch+1,loss.item()))
    x_axis.append(epoch)
    y_axis.append(loss.item())
    # drawing.....................................................................................................#
plt.figure(figsize=(7, 7), dpi=80)  # 创建画布
plt.plot(x_axis, y_axis, color='b', linestyle='-')  # 绘制折线图，点划线
plt.xlabel('epoch')  # 设置图x轴标签
plt.ylabel('loss rate')  # 设置图y轴标签
plt.legend(["loss"], title='loss&epoch', loc='upper right', fontsize=15)  # 设置图列
plt.show()

D:\Anaconda\envs\study\python.exe "D:\python pycharm learning\刘二大人课程\P\p12.2.py"
Predicted string:ooooo, Epoch [1/15] loss=7.1597
Predicted string:ooooo, Epoch [2/15] loss=5.9397
Predicted string:ooool, Epoch [3/15] loss=5.2407
Predicted string:ollll, Epoch [4/15] loss=4.9466
Predicted string:ollll, Epoch [5/15] loss=4.7458
Predicted string:ollll, Epoch [6/15] loss=4.5285
Predicted string:ohlll, Epoch [7/15] loss=4.2802
Predicted string:ohlll, Epoch [8/15] loss=3.9900
Predicted string:ohlll, Epoch [9/15] loss=3.6928
Predicted string:ohlol, Epoch [10/15] loss=3.4294
Predicted string:ohlol, Epoch [11/15] loss=3.2167
Predicted string:ohlol, Epoch [12/15] loss=3.0618
Predicted string:ohlol, Epoch [13/15] loss=2.9525
Predicted string:ohlol, Epoch [14/15] loss=2.8609
Predicted string:ohlol, Epoch [15/15] loss=2.7697