ksize in max_pool

[ayushchd]

I am trying to build a convolution (followed by max_pool) for variable length input size. Since the length of the input in a batch should be the same, I set the batch size to 1. However, for some reason ksize in tf.nn.max_pool is a list attribute required at run-time. This would prevent any neural network that uses variable input size to apply max pool. Is there a workaround for this? Should ksize be an input tensor otherwise?

[tatatodd]

FYI this was also asked about in #1957

We welcome contributions!

[ayushchd]

#1957 seems completely unrelated. Am I missing something?

[tatatodd]

I'm sorry, I meant #1967. Indeed 1957 is completely unrelated.

[guotong1988]

I want to work on this issue

[guotong1988]

found something useful for the development.
http://stackoverflow.com/questions/33919948/how-to-set-adaptive-learning-rate-for-gradientdescentoptimizer

[girving]

It may be too late for @guoguo12's assistance, but if anyone else wants to jump on this I'm happy to answer questions. The solution is to make a MaxPoolV2 C++ op similar to TopKV2, and do the same for all the other convolutional ops. We'd then deprecate the old op with .Deprecated and make the Python routines use the new version. Here's TopK and TopKV2 for comparison:

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/ops/nn_ops.cc#L2013
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/ops/nn_ops.cc#L2048

[guoguo12]

(You mean @guotong1988.)

[girving]

Whoops, sorry about that.

[yongtang]

Created a PR #9514 to try to address this issue. Would appreciate any review or comments.

[colaskirschoff]

@ayushchd I guess it's a bit late but there is a work around if you need to do your max_pool on a unique dimension. In this case you do not need any kernel size nor strides so you can use reduce_max directly.

[nkszjx]

pool = tf.reduce_max(activation, axis=1, keep_dims=True)