tensorflow的probability在pytorch中有没有对应的包

Question

请问下大家tensorflow中的probability在pytorch中有没有对应的包，我非科班学了好久pytorch，结果发现probability中的convolution1dflipout在pytorch中找不到😭

Answer 1

pytorch里面还没有与convolution1dflipout相对应的卷积层，你要不然就结合pyro这种概率编程库自己实现这个层，要不然就要用pytorch里的自定义层的功能自己实现。

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.parameter import Parameter

# 定义 Conv1dFlipout 层
class Conv1dFlipout(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
        super(Conv1dFlipout, self).__init__()

        # 初始化各种参数
        self.in_channels = in_channels          # 输入通道数
        self.out_channels = out_channels        # 输出通道数
        self.kernel_size = kernel_size          # 卷积核大小
        self.stride = stride                    # 步幅
        self.padding = padding                  # 填充
        self.dilation = dilation                # 空洞
        self.groups = groups                    # 分组卷积
        self.bias = bias                        # 是否使用偏置项

        # 定义权重均值和对数标准差，这些参数是需要学习的
        self.weight_mean = Parameter(torch.Tensor(out_channels, in_channels // groups, kernel_size))
        self.weight_logstd = Parameter(torch.Tensor(out_channels, in_channels // groups, kernel_size))

        # 定义偏置项的均值和对数标准差，这些参数是需要学习的
        if bias:
            self.bias_mean = Parameter(torch.Tensor(out_channels))
            self.bias_logstd = Parameter(torch.Tensor(out_channels))
        else:
            self.register_parameter('bias_mean', None)
            self.register_parameter('bias_logstd', None)

        # 重置各个参数的初始值
        self.reset_parameters()

    # 重置各个参数的初始值
    def reset_parameters(self):
        nn.init.kaiming_uniform_(self.weight_mean, a=math.sqrt(5))
        nn.init.constant_(self.weight_logstd, -10)

        if self.bias is not None:
            fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight_mean)
            bound = 1 / math.sqrt(fan_in)
            nn.init.uniform_(self.bias_mean, -bound, bound)
            nn.init.constant_(self.bias_logstd, -10)

    # 定义前向传播函数
    def forward(self, x):
        weight_epsilon = torch.randn(self.weight_mean.shape).to(x.device)
        weight = self.weight_mean + weight_epsilon * torch.exp(self.weight_logstd)

        if self.bias is not None:
            bias_epsilon = torch.randn(self.bias_mean.shape).to(x.device)
            bias = self.bias_mean + bias_epsilon * torch.exp(self.bias_logstd)
        else:
            bias = None

        # 获取输入和权重的维度
        batch_size, input_channels, input_length = x.size()
        output_channels, _, kernel_size = weight.size()

        # 计算输出的维度
        output_length = (input_length + 2 * self.padding - dilation * (kernel_size - 1) - 1) // self.stride + 1

           # 将输入和权重张量重塑为卷积运算所需的形状
        x = x.view(batch_size, input_channels // self.groups, self.groups, input_length)
        weight = weight.view(output_channels, input_channels // self.groups, self.groups, kernel_size)


        # 使用groups进行卷积运算
         output = F.conv1d(x, weight, bias, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups)

        # 将输出张量重塑为卷积运算输出的形状
        output = output.view(batch_size, output_channels, output_length)

    return output

       

一个新的卷积结构就要重定义一个新的卷积层，这还是其中一个示例。所以说，既然这么麻烦，算了还不如直接调用tensorflow现成的