小白也能读懂的ConvLSTM！(开源pytorch代码) - ToB企服应用市场:ToB评测及商务社交产业平台

import torch.nn as nn
import torch
class ConvLSTMCell(nn.Module):
def __init__(self, input_dim, hidden_dim, kernel_size, bias):
"""
初始化卷积 LSTM 单元。
参数:
----------
input_dim: int
输入张量的通道数。
hidden_dim: int
隐藏状态的通道数。
kernel_size: (int, int)
卷积核的大小。
bias: bool
是否添加偏置项。
"""
super(ConvLSTMCell, self).__init__()
self.input_dim = input_dim
self.hidden_dim = hidden_dim
self.kernel_size = kernel_size
# 计算填充大小以保持输入和输出尺寸一致
self.padding = kernel_size[0] // 2, kernel_size[1] // 2
self.bias = bias
# 定义卷积层，输入是输入维度加上隐藏维度，输出是4倍的隐藏维度（对应i, f, o, g）
self.conv = nn.Conv2d(in_channels=self.input_dim + self.hidden_dim,
out_channels=4 * self.hidden_dim,
kernel_size=self.kernel_size,
padding=self.padding,
bias=self.bias)
def forward(self, input_tensor, cur_state):
h_cur, c_cur = cur_state
# 沿着通道轴进行拼接
combined = torch.cat([input_tensor, h_cur], dim=1)
combined_conv = self.conv(combined)
# 将输出分割成四个部分，分别对应输入门、遗忘门、输出门和候选单元状态
cc_i, cc_f, cc_o, cc_g = torch.split(combined_conv, self.hidden_dim, dim=1)
i = torch.sigmoid(cc_i)
f = torch.sigmoid(cc_f)
o = torch.sigmoid(cc_o)
g = torch.tanh(cc_g)
# 更新单元状态
c_next = f * c_cur + i * g
# 更新隐藏状态
h_next = o * torch.tanh(c_next)
return h_next, c_next
def init_hidden(self, batch_size, image_size):
height, width = image_size
# 初始化隐藏状态和单元状态为零
return (torch.zeros(batch_size, self.hidden_dim, height, width, device=self.conv.weight.device),
torch.zeros(batch_size, self.hidden_dim, height, width, device=self.conv.weight.device))
class ConvLSTM(nn.Module):
"""
卷积 LSTM 层。
参数:
----------
input_dim: 输入通道数
hidden_dim: 隐藏通道数
kernel_size: 卷积核大小
num_layers: LSTM 层的数量
batch_first: 批次是否在第一维
bias: 卷积中是否有偏置项
return_all_layers: 是否返回所有层的计算结果
输入:
------
一个形状为 B, T, C, H, W 或者 T, B, C, H, W 的张量
输出:
------
元组包含两个列表（长度为 num_layers 或者长度为 1 如果 return_all_layers 为 False）：
0 - layer_output_list 是长度为 T 的每个输出的列表
1 - last_state_list 是最后的状态列表，其中每个元素是一个 (h, c) 对应隐藏状态和记忆状态
示例:
>>> x = torch.rand((32, 10, 64, 128, 128))
>>> convlstm = ConvLSTM(64, 16, 3, 1, True, True, False)
>>> _, last_states = convlstm(x)
>>> h = last_states[0][0] # 0 表示层索引，0 表示 h 索引
"""
def __init__(self, input_dim, hidden_dim, kernel_size, num_layers,
batch_first=False, bias=True, return_all_layers=False):
super(ConvLSTM, self).__init__()
# 检查 kernel_size 的一致性
self._check_kernel_size_consistency(kernel_size)
# 确保 kernel_size 和 hidden_dim 的长度与层数一致
kernel_size = self._extend_for_multilayer(kernel_size, num_layers)
hidden_dim = self._extend_for_multilayer(hidden_dim, num_layers)
if not len(kernel_size) == len(hidden_dim) == num_layers:
raise ValueError('不一致的列表长度。')
self.input_dim = input_dim
self.hidden_dim = hidden_dim
self.kernel_size = kernel_size
self.num_layers = num_layers
self.batch_first = batch_first
self.bias = bias
self.return_all_layers = return_all_layers
# 创建 ConvLSTMCell 列表
cell_list = []
for i in range(0, self.num_layers):
cur_input_dim = self.input_dim if i == 0 else self.hidden_dim[i - 1]
cell_list.append(ConvLSTMCell(input_dim=cur_input_dim,
hidden_dim=self.hidden_dim[i],
kernel_size=self.kernel_size[i],
bias=self.bias))
self.cell_list = nn.ModuleList(cell_list)
def forward(self, input_tensor, hidden_state=None):
"""
前向传播函数。
参数:
----------
input_tensor: 输入张量，形状为 (t, b, c, h, w) 或者 (b, t, c, h, w)
hidden_state: 初始隐藏状态，默认为 None
返回:
-------
last_state_list, layer_output
"""
if not self.batch_first:
# 改变输入张量的顺序，如果 batch_first 为 False
input_tensor = input_tensor.permute(1, 0, 2, 3, 4)
b, _, _, h, w = input_tensor.size()
# 实现状态化的 ConvLSTM
if hidden_state is not None:
raise NotImplementedError()
else:
# 初始化隐藏状态
hidden_state = self._init_hidden(batch_size=b,
image_size=(h, w))
layer_output_list = []
last_state_list = []
seq_len = input_tensor.size(1)
cur_layer_input = input_tensor
for layer_idx in range(self.num_layers):
h, c = hidden_state[layer_idx]
output_inner = []
for t in range(seq_len):
# 在每个时间步上更新状态
h, c = self.cell_list[layer_idx](input_tensor=cur_layer_input[:, t, :, :, :],
cur_state=[h, c])
output_inner.append(h)
# 将输出堆叠起来
layer_output = torch.stack(output_inner, dim=1)
cur_layer_input = layer_output
layer_output_list.append(layer_output)
last_state_list.append([h, c])
if not self.return_all_layers:
# 如果不需要返回所有层，则只返回最后一层的输出和状态
layer_output_list = layer_output_list[-1:]
last_state_list = last_state_list[-1:]
return layer_output_list, last_state_list
def _init_hidden(self, batch_size, image_size):
init_states = []
for i in range(self.num_layers):
# 初始化每一层的隐藏状态
init_states.append(self.cell_list[i].init_hidden(batch_size, image_size))
return init_states
@staticmethod
def _check_kernel_size_consistency(kernel_size):
if not (isinstance(kernel_size, tuple) or
(isinstance(kernel_size, list) and all([isinstance(elem, tuple) for elem in kernel_size]))):
raise ValueError('`kernel_size` 必须是 tuple 或者 list of tuples')
@staticmethod
def _extend_for_multilayer(param, num_layers):
if not isinstance(param, list):
param = [param] * num_layers
return param

复制代码