用pytorch实现Our network has 3 layers of LSTM units followed by a final fully-connected layer of rectified linear units (ReLUs). A probability distribution for the next price move is produced by applying a softmax activation function.

可以使用PyTorch实现您所描述的神经网络，代码如下：

import torch
import torch.nn as nn

class LSTMPredictor(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(LSTMPredictor, self).__init__()
        self.hidden_size = hidden_size
        self.lstm1 = nn.LSTM(input_size, hidden_size)
        self.lstm2 = nn.LSTM(hidden_size, hidden_size)
        self.lstm3 = nn.LSTM(hidden_size, hidden_size)
        self.fc = nn.Linear(hidden_size, output_size)
        self.softmax = nn.Softmax(dim=1)

    def forward(self, input):
        h1 = torch.zeros(1, input.size(1), self.hidden_size)
        c1 = torch.zeros(1, input.size(1), self.hidden_size)
        h2 = torch.zeros(1, input.size(1), self.hidden_size)
        c2 = torch.zeros(1, input.size(1), self.hidden_size)
        h3 = torch.zeros(1, input.size(1), self.hidden_size)
        c3 = torch.zeros(1, input.size(1), self.hidden_size)
        
        out1, (h1, c1) = self.lstm1(input, (h1, c1))
        out2, (h2, c2) = self.lstm2(out1, (h2, c2))
        out3, (h3, c3) = self.lstm3(out2, (h3, c3))
        out4 = self.fc(out3[-1])
        out5 = self.softmax(out4)

        return out5

在这个模型中，我们使用了三个LSTM层，每个LSTM层的输出作为下一个LSTM层的输入。最后，我们使用一个全连接层和softmax激活函数来生成下一个价格移动的概率分布。

您可以根据您的数据集和需要调整输入大小（input_size）、隐藏层大小（hidden_size）和输出大小（output_size）。