TensorFlow max_pool2d wrong output size

Question

I'm trying to use a max pool layer with filter size 2x2, so I expect the output size to be roughly half the input size. The input size is 9x14x64, but for some reason the output size is 7x12x64 (see the attached TensorBoard graph).

Here is the TensorFlow code:

layer = conv2d(layer, 64, 3, 2, activation_fn=None, scope="conv_2", name=name)
layer = max_pool2d(layer, 2, 1, scope="max_1", name=name)
layer = conv2d(layer, 32, 3, 1, activation_fn=None, scope="conv_3", name=name)

Am I missing something?

Ishant Mrinal · Answer 1 · 2017-08-23T13:01:41.183

0

It seems you are tensorflow default data_format NHWC; but your input format is NCHW. So you need to change your input format to NHWC

N -batch_size, H-height, W-width, C-num_channels

Note: Max-pool only changes height and width of the input feature maps. There will be no effect on num_channels (it will be same for both input and output).

If you want to change the number of channels then you can use a 1x1 convolution layer as follows.

# 9 x 14 x 64
layer = max_pool2d(layer, 2, 2, scope="max_1", name=name, padding='SAME')
# output size 5x7x64
# to reduce num_channels use 1x1 conv
layer = conv2d(layer, 32, 1, 1, activation_fn=None, scope="conv_pool", name=name, padding='SAME')
# output size 5x7x32

edited Aug 23 '17 at 13:01

answered Aug 23 '17 at 12:36

Ishant Mrinal

4,898
3
29
47

then the output is expected. `max_pool` only change the `height` and `width`, not `num_channels` – Ishant Mrinal Aug 23 '17 at 12:46
It actually changed height and width. Take a look on the tensorboard graph screenshot. – doetensor Aug 23 '17 at 12:49
see my above comment. – Ishant Mrinal Aug 23 '17 at 12:54
The number of channels is not changed. There are 64 channels in the input and output. The only dimensions changed are H and W. – doetensor Aug 23 '17 at 12:56

TensorFlow max_pool2d wrong output size

1 Answers1