RNN

LSTM

import torch
lstm:torch.nn.LSTM=torch.nn.LSTM(5,3,num_layers=2)
param_ls=list(lstm.named_parameters())
end=True

在这里插入图片描述

import torch
lstm:torch.nn.LSTM=torch.nn.LSTM(input_size=5,hidden_size=3,num_layers=1)
param_ls=list(lstm.named_parameters())
end=True

在这里插入图片描述

mnist lstm

import torch
import torch.nn.functional

IMG_H=28
IMG_W=28
MNIST_CLASS_CNT=10
TRAINING_SAMPLE_CNT=5
learning_rate=0.1

class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.lstm_layer_cnt=1
        self.lstm_hidden_size=16
        self.lstm_input_size=IMG_W
        self.lstm=torch.nn.LSTM(input_size=self.lstm_input_size,hidden_size=self.lstm_hidden_size,num_layers=self.lstm_layer_cnt,batch_first=True)
        self.fc = torch.nn.Linear(in_features=self.lstm_hidden_size, out_features=MNIST_CLASS_CNT)

    def forward(self,x):
        batch_size=x.size(0)
        h0=torch.zeros(self.lstm_layer_cnt,batch_size,self.lstm_hidden_size)
        c0 = torch.zeros(self.lstm_layer_cnt, batch_size, self.lstm_hidden_size)
        out,hidden=self.lstm(x,(h0,c0))
        end_out=out[:,-1,:]
        fc_out=self.fc(end_out)
        ? = torch.log_softmax(input=fc_out,dim=1)
        return ?

x=torch.randn((TRAINING_SAMPLE_CNT,IMG_H,IMG_W),dtype=torch.float)
y=torch.zeros((TRAINING_SAMPLE_CNT),dtype=torch.long)
net=Net()
optimizer=torch.optim.Adam(params=net.parameters(),lr=learning_rate)
print(f"net:{net}")
"""
Net(
  (lstm): LSTM(28, 16, batch_first=True)
  (fc): Linear(in_features=16, out_features=10, bias=True)
)"""

#train:
#forward
?=net(x)
loss:torch.Tensor=torch.nn.functional.nll_loss(input=?,target=y)

#backward
optimizer.zero_grad()
loss.backward()
optimizer.step()
end=True