Tensorflow: how could tf.nn.max_pool work with dynamic kernel size?

[ahmed-touati]

[zheng-xq]

Not with the current op interface. You would have to write a new op that takes the kernel size as an input tensor. The kernel implementation can be refactored to use most of the current max_pool implementation.

[vincentvanhoucke]

Closing. Feel free to reopen if this is a feature request worth tracking.

[apple2373]

+1
I would like to request that feature.