Gradient backward of the first filter is much lower than other layers.

[mechanicalsea]

When using PyTorch, I showed that learning the first filter cost much time, e.g.,

1. with a batch (200, 32240), a gradient backward of the first filter took about 1 second,
2. however, other layers took only lower than 0.1 seconds.

Is that happen to you?

I also compared it with ResNetSE34 and showed:

1. For the forward process, SincNet is more efficient than ResNetSE34,
2. however, for the backward, ResNetSE34 took less time.

Thanks for your attention.
Looking forward to your response.