InvalidArgumentError (see above for traceback): No OpKernel was registered to support Op 'ComputeflowGrad'

[kevinkit]

When trying to run the shapenet experiments I get the following erros:

InvalidArgumentError (see above for traceback): No OpKernel was registered to support Op 'ComputeflowGrad' with these attrs. Registreed devices: [CPU], Registered kernels: device = 'GPU'; T in [DT_FLOAT].

Then an output for the node and then ...:

CUDA Runtime Error: no CUDA-capable device is detected

I am using the same versiosn as you suggested for tensorflow, CUDA and all other dependencies. If I execute small test script like:

import tensorflow as tf
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))

I can see my 2 Titan 1080 GPUs mapped as devices.

I used any kind of index for $GPU_ID, but the error is still the same.

We really appreciate your help and thank you for your help

[yuxng]

I am not sure what causes the problem here.

[kevinkit]

We found out that the error comes from the Script:

shapenet_scene_mutli_normal.sh

However, the code looks fine - which kind of GPU are you using? Does it provide INT8-Precision? Is it a GPU specially designed for Deep-Learning?

The same error occurs on

shapenet_scene_single_normal.sh

Furthermore, the tensorflow error is gone, however the

CUDA Runtime Error: no CUDA-capable device is detected

stays.

Thank you really much for your help

[yuxng]

I have tested the code on Titan X (Pascal) and GeForce 1080. I can run both the scripts. Maybe check your cuda.

[kevinkit]

Changing to Ubuntu 16.04 resolvied this issue. Also the $GPU_ID must be given as an input, when running with two cards it crashes when this parameter is left out.

Thank you really much for your help.

[yuxng]

Good to know!