What is meant by " Number of channels in the input image" and " Number of channels in the output image" in paramaters of torch.nn.Conv2d here:http://pytorch.org/docs/master/nn.html
I’m sorry I’m new, but can’t find asnwer…some people say it’s colour channels, but then why it’s 1,6 and 6,16 in beginner tutorial(http://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html)? Or are these channels number of different filters in one bundle of convolution layers?
In general, a “2d” tensor in CNNs is of size “Batch x Channels x Height x Width.” For the actual input to the network, channels is usually 3 for RGB or 1 if it’s greyscale. For the outputs of layers in the network, “output channels” is analagous to the number of neurons, or the number of hidden units, of a layer. So, the latter–output channels are the number of filters in one layer, while input channels are the number of filters in the incoming layer.
Where would you see this and be able to change in a typical pytorch CNN?
Usually you specify the number of channels when defining your architecture (see, e.g. this example: https://github.com/AghdamAmir/3D-UNet/blob/main/unet3d.py#L132), if you are working with pre-defined architectures there is not much you can do in terms of changing the number of input channels. If you have a greyscale image, and you wish to use it in a 3-channel RGB-indended network, you can always duplicate the dimension on run it like that.
See also, for instance, this link on defining a neural network layer, with specific number of input channels: Conv3d — PyTorch 2.0 documentation