Giters

vlawhern / arl-eegmodels

This is the Army Research Laboratory (ARL) EEGModels Project: A Collection of Convolutional Neural Network (CNN) models for EEG signal classification, using Keras and Tensorflow

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Run EEGNet on CPU

ShobiPr opened this issue · comments

commented

Hi,

I'm getting this error message when testing the code:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Default AvgPoolingOp only supports NHWC on device type CPU
	 [[{{node average_pooling2d/AvgPool}}]]

Is it possible to run EEGNet without GPU?

commented

So that error comes from using older versions of tensorflow-cpu (<2.0.0). I did a quick check to see if newer versions of tensorflow are affected and it appears this is no longer an issue if you're using tensorflow >=2.0.0.

If you're using Anaconda's python distribution (https://www.anaconda.com/distribution/) then you can install the CPU version of tensorflow (currently 2.0.0) with conda:

conda install tensorflow

I don't know if this is true if you install tensorflow with pip; I know that the version installed with conda is compiled against an optimized BLAS so perhaps there could be other differences with the pip package versus the conda package. Which version of tensorflow are you using and how was it installed?

commented

I used tensorflow version 1.15.0 and installed it using pip3.
I now tried with tensorflow version 2.1.0 (installed using pip3) and getting this error message:

tensorflow.python.framework.errors_impl.InvalidArgumentError: Default AvgPoolingOp only supports NHWC on device type CPU
[[node model/average_pooling2d/AvgPool (defined at /Users/****/Desktop/cnn/test.py:124) ]] [Op:__inference_distributed_function_1921]
Function call stack:
distributed_function

commented

OK I can recreate the error message you reported for tensorflow 2.0.0 as well when installed through pip so this points to differences in the conda package vs the pip package and how they are compiled and configured.

So the situation is:

  1. If you install tensorflow through pip you'll unfortunately need the GPU version as the AveragePool op doesn't support CPU
  2. If you install tensorflow through Anaconda, either the CPU or GPU version works. I'm guessing this is because tensorflow through conda is configured differently (i.e. compiles against the Intel MKL BLAS, perhaps there are other differences)
  3. You can try to compile tensorflow yourself pointing to an optimized BLAS (Intel MKL or maybe even OpenBLAS, although I don't know if this will work). I've never tried to do this so I wouldn't be able to help much here.

If none of these options are available to you, a very hacky last-resort, and something I do not recommend doing for numerous reasons, is to replace the AveragePooling2D layer with a DepthwiseConv2D layer with no bias and with kernel size and stride equal to the size of the pooling layer. Meaning for example:

AveragePooling2D(pool_size=(1,4))

would be something like

DepthwiseConv2D(kernel_size=(1,4), strides=(1,4), bias=False)

then set that layer's weights using .set_weights() equal to (1/K) where K is the length of the kernel. This should in theory work but I've never tested this and it may have weird edge cases (for example if you're using padding = 'same' vs 'causal' vs 'valid'). So I definitely wouldn't pursue this unless you're willing to validate it.

commented

Another alternative and maybe the easiest, is to install intel-tensorflow with pip.

commented

I uninstalled tensorflow and reinstalled it through Anaconda and it worked! Thanks for the help!

commented

good to hear! I'll leave this open so that other users can find it easier. Let me know if there are other issues.