tensorflow: Group convolutions UnimplementedError for nightly build

I am getting the following error message while running the following code using the nightly version of tensorflow 1.14.1-dev20190524. When will this feature be available?

Code: weights = tf.ones([3,3,1,3]) X = tf.ones([3,1024,1024,3]) Y = tf.nn.conv2d(X, weights, [1, 1, 1, 1], "VALID")

Error: UnimplementedError: Generic conv implementation does not support grouped convolutions for now. [[node Conv2D_4 (defined at <ipython-input-13-3298cb87493e>:4) ]]

About this issue

Original URL
State: closed
Created 5 years ago
Comments: 19 (15 by maintainers)

Commits related to this issue

Implement grouped convolution on CPU To get better compute resources utilization group-compute loop has to be parallelized, but it involves a lot of changes in Conv2D primitives. Will address that la... — committed to tensorflow/tensorflow by ezhulenev 3 years ago

Most upvoted comments

I have a pending PR for that, should be submitted soon, although currently it has few unresolved TODOs that are needed to get to top performance.

ezhulenev on Mar 19, 2021

Anyone who is facing the same issue even after using tf version 2.6.0rc1, you need to pass the batch size of exactly 1. If you pass more than 1, then it won’t work on CPU.

ashutosh1919 on Jul 24, 2021

Grouped convolutions are currently only supported on GPUs, not CPUs. You are using a grouped convolution because the input channel dimension of 3 (X.shape[3]) does not match the filter in-depth dimension of 1 (weights.shape[2]). @chsigg, any idea when support is coming for CPUs?

Also the error message is a bit unclear. I’ll fix the error message.

reedwm on May 30, 2019