第一层 1D CNN 的最佳输入大小

Question

我正在尝试构建具有时间序列的一维 CNN。输入的长度为 500。（只有）2 个标签。到目前为止，我构建的架构如下：每个有 3 个卷积层，其中有一个激活层。第一个卷积层以 50 个通道作为输入。

import torch
import torch.nn as nn
import numpy as np
import random

class Simple1DCNN3(torch.nn.Module):
    def __init__(self):
        super(Simple1DCNN5, self).__init__()
        self.sequence = nn.Sequential(
            torch.nn.Conv1d(in_channels=50, 
                                          out_channels=64, 
                                          kernel_size=5, 
                                          stride=2),
            torch.nn.ReLU(),
            torch.nn.Conv1d(in_channels=64, 
                                          out_channels=128, 
                                          kernel_size=3),
            torch.nn.ReLU(),
            torch.nn.Conv1d(in_channels=128, 
                                          out_channels=256, 
                                          kernel_size=1),
            torch.nn.ReLU(),
        )
        self.fc1 = nn.Linear(256, 2)

        

    def forward(self, x):
        x = x.view(1, 50,-1)
        
        for layer in self.sequence:
            x = layer(x)
            print(x.size())
        x = x.view(1,-1)
        #print(x.size())
        x = self.fc1(x)
        #print(x.size())
        return x

net = Simple1DCNN3()

input_try = np.random.uniform(-10, 10, 500)
input_try = torch.from_numpy(input_try).float()
net(input_try)
print("input successfull passed to net")
input_try_modif = input_try.view(1, 50,-1)
print(input_try.shape)
print(input_try_modif.shape)

据我了解，这迫使我将输入分成 50 个时间点的 10 个片段。我理解错了吗？用 500 个通道作为输入构建第一层并有一个滑动窗口内核不是更明智吗？我在以下其他脚本中尝试过，但收到以下错误消息

import torch
import torch.nn as nn
import numpy as np
import random

class Simple1DCNN4(torch.nn.Module):
    def __init__(self):
        super(Simple1DCNN5, self).__init__()
        self.sequence = nn.Sequential(
            torch.nn.Conv1d(in_channels=500, 
                                          out_channels=64, 
                                          kernel_size=5, 
                                          stride=2),
            torch.nn.ReLU(),
            torch.nn.Conv1d(in_channels=64, 
                                          out_channels=128, 
                                          kernel_size=3),
            torch.nn.ReLU(),
            torch.nn.Conv1d(in_channels=128, 
                                          out_channels=256, 
                                          kernel_size=1),
            torch.nn.ReLU(),
        )
        self.fc1 = nn.Linear(256, 2)

        

    def forward(self, x):
        x = x.view(1, 50,-1)
        
        for layer in self.sequence:
            x = layer(x)
            print(x.size())
        x = x.view(1,-1)
        #print(x.size())
        x = self.fc1(x)
        #print(x.size())
        return x

net = Simple1DCNN4()

input_try = np.random.uniform(-10, 10, 500)
input_try = torch.from_numpy(input_try).float()
net(input_try)
print("input successfull passed to net")
input_try_modif = input_try.view(1, 50,-1)
print(input_try.shape)
print(input_try_modif.shape)

错误信息：

RuntimeError: Given groups=1, weight of size [64, 500, 5], expected input[1, 50, 10] to have 500 channels, but got 50 channels instead

编辑

感谢@ghchoi 的回答，这是工作内核的代码。为此，我还必须将所有卷积层的内核大小更改为 1。

class Simple1DCNN5(torch.nn.Module):
    def __init__(self):
        super(Simple1DCNN5, self).__init__()
        self.sequence = nn.Sequential(
            torch.nn.Conv1d(in_channels=500, 
                                          out_channels=64, 
                                          kernel_size=1, 
                                          stride=2),
            torch.nn.ReLU(),
            torch.nn.Conv1d(in_channels=64, 
                                          out_channels=128, 
                                          kernel_size=1),
            torch.nn.ReLU(),
            torch.nn.Conv1d(in_channels=128, 
                                          out_channels=256, 
                                          kernel_size=1),
            torch.nn.ReLU(),
        )
        self.fc1 = nn.Linear(256, 2)

        

    def forward(self, x):
        x = x.view(1, 500,-1)
        
        for layer in self.sequence:
            x = layer(x)
            #print(x.size())
        x = x.view(1,-1)
        #print(x.size())
        x = self.fc1(x)
        #print(x.size())
        return x

我拥有的数据类型是 2 秒的单导 ECG（心电图）信号。这是对听觉电信号的记录。这是样本的外观（绘制在 2D 图表上）的想法，其中 x 轴上有时间，y 轴上有电压/幅度

Answer 1

试试

    def forward(self, x):
        x = x.view(1, 500, -1) 

        ...

net = Simple1DCNN4()

input_try = np.random.uniform(-10, 10, 5000)

这样，第一个 Conv1d 的输入将有 500 个通道。