【Pytorch】Conv1d

conv1d 先看看官方文档

再来个简单的例子

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

data = np.arange(1, 13).reshape([1, 4, 3])
data = torch.tensor(data, dtype=torch.float)
print("[data]:\n", data)
conv = nn.Conv1d(in_channels=4, out_channels=1, kernel_size=3, stride=1, padding=0, bias=False)
print("[weight]:\n", conv.weight)
print("[element-wise multiply]:", (data*conv.weight).sum().item())
output = conv(data)
print("[output size]:", output.size())
print("[output]:", output)

"""
[data]:
 tensor([[[ 1.,  2.,  3.],
         [ 4.,  5.,  6.],
         [ 7.,  8.,  9.],
         [10., 11., 12.]]])
[weight]:
 Parameter containing:
tensor([[[ 0.2599, -0.1309, -0.2319],
         [-0.1974, -0.0371, -0.1319],
         [-0.2190, -0.1151,  0.0644],
         [-0.0862, -0.2313,  0.1159]]], requires_grad=True)
[element-wise multiply]: -6.354348182678223
[output size]: torch.Size([1, 1, 1])
[output]: tensor([[[-6.3543]]], grad_fn=<SqueezeBackward1>)
"""

2D卷积是以 n 个 3D tensor 在二维平面滑动，所以叫 2D，标准2D卷积的卷积核 $C$ 和输入特征图的 $C_{in}$ 应该保持一致

忽略 C 的 2D conv 演示，金字塔，哈哈哈

同理

1D 卷积是以 n 个 2D tensor 在一维平面滑动，标准 1D 卷积核的 h 和 输入的 h 应该保持一致，别潜意识的理解为 h 只能等于 1，像下图这样，h = 1

再看一个稍微复杂的例子

eg:

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

data = np.arange(1, 13).reshape([1, 4, 3])
data = torch.tensor(data, dtype=torch.float)
print("[data]:\n", data)
conv = nn.Conv1d(in_channels=4, out_channels=2, kernel_size=3, stride=1, padding=0, bias=False)
print("[weight]:\n", conv.weight)
print("[element-wise multiply]:", (data*conv.weight[0]).sum().item())
print("[element-wise multiply]:", (data*conv.weight[1]).sum().item())
output = conv(data)
print("[output size]:", output.size())
print("[output]:", output)

"""
[data]:
 tensor([[[ 1.,  2.,  3.],
         [ 4.,  5.,  6.],
         [ 7.,  8.,  9.],
         [10., 11., 12.]]])
[weight]:
 Parameter containing:
tensor([[[-0.1565,  0.2598, -0.2152],
         [-0.0130, -0.1495, -0.0799],
         [-0.2842, -0.1508, -0.1939],
         [-0.1133, -0.2627,  0.1949]],

        [[ 0.0576,  0.1712, -0.1465],
         [-0.2060,  0.1648,  0.2039],
         [ 0.2221, -0.1940,  0.1126],
         [-0.2098, -0.0749,  0.1407]]], requires_grad=True)
[element-wise multiply]: -8.187724113464355
[element-wise multiply]: 0.9679145812988281
[output size]: torch.Size([1, 2, 1])
[output]: tensor([[[-8.1877],
         [ 0.9679]]], grad_fn=<SqueezeBackward1>)
"""

结合可视化看看 1D 卷积是怎么滑动的（来自 添加链接描述）

eg

m = nn.Conv1d(4, 2, 3, stride=2)
input = torch.randn(1, 4, 9)
print(input)
output = m(input)
print(output)
print(output.size())

output

tensor([[[-0.2105, -1.0958,  0.7299,  1.1003,  2.3175,  0.8186, -1.7510,  -0.1925,  0.8591],
         [ 1.0991, -0.3016,  1.5633,  0.6162,  0.3150,  1.0413,  1.0571,  -0.7014,  0.2239],
         [-0.0658,  0.4755, -0.6653, -0.0696,  0.3483, -0.0360, -0.4665,   1.2606,  1.3365],
         [-0.0186, -1.1802, -0.8835, -1.1813, -0.5145, -0.0534, -1.2568,   0.3211, -2.4793]]])
         
tensor([[[-0.8012,  0.0589,  0.1576, -0.8222],
         [-0.8231, -0.4233,  0.7178, -0.6621]]], grad_fn=<SqueezeBackward1>)

torch.Size([1, 2, 4])

第一个卷积核

第二个卷积核