In CNN, can I turn on/off the filters during test?

I want on/off the some filters in the layers on CNN.
what I want is below.
during test, i set the batch size to 1. then, the first batch enters network.
the network outputs the softmax values.
using this softmax values, I want to turn off the filters in network, which is strongly related to the low softmax values.
then next batch, off filter do not perform computation, and the number of softmax results is less than first batch.

can I do like above with pytorch?

I think what you want is to depress some values in the feature maps but not filter, is that right?

yes, maybe…
can I increase or decrease the size of feature maps during test?
in other words, while test, every iteration, the size of feature maps is not same.

I am not sure what you want, but I think you can hook out features what you want to edit
and edit them anyway.