Here is what you may find
torch.nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode=”zeros”)
Parameters
- in_channels (int) – Number of channels in the input image
- out_channels (int) – Number of channels produced by the convolution
- kernel_size (int or tuple) – Size of the convolving kernel
- stride (int or tuple, optional) – Stride of the convolution. (Default: 1)
- padding (int or tuple, optional) – Zero-padding added to both sides of the input (Default: 0)
- padding_mode (string, optional) – zeros
- dilation (int or tuple, optional) – Spacing between kernel elements. (Default: 1)
- groups (int, optional) – Number of blocked connections from input to output channels. (Default: 1)
- bias (bool, optional) – If True, adds a learnable bias to the output. (Default: True)
And this URL has helpful visualization of the process.
So the in_channels
in the beginning is 3 for images with 3 channels (colored images).
For images black and white it should be 1.
Some satellite images should have 4.
The out_channels
is the number of filters and you can set this arbitrary.
Let’s create an example to “prove” that.
import torch
import torch.nn as nn
c = nn.Conv2d(1,3, stride = 1, kernel_size=(4,5))
print(c.weight.shape)
print(c.weight)
Out
torch.Size([3, 1, 4, 5])
Parameter containing:
tensor([[[[ 0.1571, 0.0723, 0.0900, 0.1573, 0.0537],
[-0.1213, 0.0579, 0.0009, -0.1750, 0.1616],
[-0.0427, 0.1968, 0.1861, -0.1787, -0.2035],
[-0.0796, 0.1741, -0.2231, 0.2020, -0.1762]]],
[[[ 0.1811, 0.0660, 0.1653, 0.0605, 0.0417],
[ 0.1885, -0.0440, -0.1638, 0.1429, -0.0606],
[-0.1395, -0.1202, 0.0498, 0.0432, -0.1132],
[-0.2073, 0.1480, -0.1296, -0.1661, -0.0633]]],
[[[ 0.0435, -0.2017, 0.0676, -0.0711, -0.1972],
[ 0.0968, -0.1157, 0.1012, 0.0863, -0.1844],
[-0.2080, -0.1355, -0.1842, -0.0017, -0.2123],
[-0.1495, -0.2196, 0.1811, 0.1672, -0.1817]]]], requires_grad=True)
If we would alter the number of out_channels,
c = nn.Conv2d(1,5, stride = 1, kernel_size=(4,5))
print(c.weight.shape) # torch.Size([5, 1, 4, 5])
We will get 5 filters each filter 4×5 as this is our kernel size.
If we would set 2 channels, (some images may have 2 channels only)
c = nn.Conv2d(2,5, stride = 1, kernel_size=(4,5))
print(c.weight.shape) # torch.Size([5, 2, 4, 5])
our filter will have 2 channels.
I think they have terms from this book and since they haven’t called it filters, they haven’t used that term.
So you are right; filters are what conv layer is learning and the number of filters is the number of out channels. They are set randomly at the start.
Number of activations is calculated based on bs
and image dimension:
bs=16
x = torch.randn(bs, 3, 28, 28)
c = nn.Conv2d(3,10,kernel_size=5,stride=1,padding=2)
out = c(x)
print(out.nelement()) #125440 number of activations