Has Conv3D fully supported on Apple Silicon yet?

General Discussion
, , ,
#1

Hi Community,

I am using an Apple Silicon device try to run a 3D-Unet learning network. I have installed the Tensorflow correctly and tried the MINST dataset training successfully (Using Conv2D). But when I changed the layer with 3D there are always error messages as follow:

InvalidArgumentError: Conv3DBackpropInputOpV2 only supports NDHWC on the CPU.
[[{{node training/Adam/gradients/gradients/conv3d_29/Conv3D_grad/Conv3DBackpropInputV2}}]]

The codes/network is executable on Google Colab, but I really want to run it on my local device. If anyone knows what could the issue be or if the conv3D is not yet fully supported on Apple silicon please let me know.

Any suggestion and comment is welcome
Many many thanks

#3

Hi @MW_Shay

Welcome to the TensorFlow Forum!

Conv3D layer accepts input_shape in 5+D tensor with shape: - NDHWC (Batches, Dimensions, Height, Width, Channels_last). Please ensure you are providing the inputs in 5D shape.

Also please check if you are providing the channel_last (default) in the input_shape for Conv3D layer because channels_first format is currently not supported by TensorFlow on CPU which could be the reason of the above error. You may need to enable the GPU with tensorflow to be compatible for supporting Channel_first input_shape. Thank you.