Why does the number of channels (C) become 1 in the output of the 1D convolutional layer for SCBT data?

I have a SCBT dataset with dimensions of 100×49×1×1500, and I set up a 1D convolutional layer as：convolution1dLayer(3, 1, "Stride", 1). Why does the output data become 100×1×1×1500? How did the dimensions of the output data change to this result?

Answer 1

[Hitesh]

Hi 家俊,

convolution1dLayer(3, 1, "Stride", 1) creates a 1D convolutional layer with

• Filter size = 3
• Number of filters (output channels) = 1
• Stride = 1

This means:

• The convolution slides over the sequence dimension (length 49) with a window of 3, step size 1.
• The output will have only 1 channel.

Input: 100 × 49 × 1 × 1500

Layer: convolution1dLayer(3, 1, "Stride", 1)

MATLAB expects 1D convolutions to apply over the first spatial dimension, not the second one. The convolution is applied over the second dimension (W = 49).Output width is calculated as:

Output width = ⌊Input width−Filter sizeStride⌋+1=47,so output shape becomes: 100 × 47 × 1 × 1500, but output needs to be 100 × 1 × 1 × 1500.

This means that globalAveragePooling1dLayer or averagePooling1dLayer(47) has been applied which reduces the width from 47 to 1. The reason the number of channels becomes 1 is because you set the number of filters in convolution1dLayer to 1:

• This sets the output channels to 1.
• The reduction in spatial dimensions (from 49 to 1) is due to a pooling or flattening operation after the convolution.

[家俊]

Thank you for your answer.

I'm still a bit confused about how to ensure that MATLAB's convolution1Dlayer correctly operates along the first spatial dimension and produces an output of shape 100 × 47 × 1 × 1500. Below is my overall network architecture. Could you please provide guidance?

layers=[sequenceInputLayer(1,'MiniLength',1024);

stfLayer('Window',hanning(1024,'periodic'),'TransformMode','realimag');

functionLayer(@(X) X(1:100,1:2:end,:,:));

convolution1dLayer(100,1,"Stride",1)]

[家俊]

For additional clarification.the functionLayer' output format is "SCBT" dlarray, seems correct.

[Hitesh]

Hi 家俊,

MATLAB uses this dimension order for 1D data in Deep Learning Toolbox:[H × W × C × N] = [Height × Width × Channels × Batch Size]

• 49 is the "time" or "sequence" length → this is what convolution1dLayer slides over
• 1 is the number of channels
• 100 is treated as "height" — meaning each of the 100 rows is convolved independently, like 100 separate 1D sequences per observation
• 1500 is batch size.

You want to convolve over the W = 49 dimension with filter size 3, stride 1, and produce 1 output channel.

layer = convolution1dLayer(3, 1, 'Stride', 1);

This will reduce the width from 49 to 47, using this formula:Output width=⌊49−31⌋+1=47

So the output shape will be: 100 × 47 × 1 × 1500

[家俊]

Actually, “layer = convolution1dLayer(3, 1, 'Stride', 1)”，the output shape is not100x47x1x1500. As you see, I haven't added averagePooling1dLayer(47) in my net, but the output result is indeed 100x1x1x1500.This is where I get confused.

Do you mean I have to change the channel order of the input format "SCBT" for the convolution1D layer?

Thank you for your generous help.