ImportError: Cannot import name 'batching' about benchmarks

[ppwwyyxx]

Something is in the codebase but not in your machine -- which means your tensorflow needs to be updated.

[amrawadk]

Thank you for the quick response. The version I have installed with pip3 is 'tensorflow-gpu1.4.0' which seems to be the latest. I ran the 'pip3 install --upgrade tensorflow-gpu' and it says all requirements are up to date.

[ppwwyyxx]

You'll need tf-nightly-gpu as mentioned in #80 .

I personally found the nightly sometimes unstable. (tried it yesterday and got some segfaults). Alternatively you can just use an old version of this repo with a stable version of tensorflow.

[amrawadk]

I see, I'll try this out and let you know if my issue is solved.

Thank you.

[amrawadk]

I tried installing 'tf-nightly-gpu' but I still get the same error.

Is there a certain version of this repo. I should use ? and how do I clone a certain version ?

[ppwwyyxx]

Works for me. tf.VERSION='1.5.0-dev20171113'.
You may need to uninstall tensorflow or tensorflow-gpu before installation. You can check your tf.VERSION to confirm.

[amrawadk]

Thanks for your help :D

I managed to get the testbench marks working with CUDA V8 & CUDNN V6 & tf-nightly-gpu. I apologize for the late reply but I had to test a lot of variations before it worked for me.

[nasica88]

Hello, I have the same error with TF 1.3. Do I have to use TF1.5 to address this error ? Isn't there any way to run this benchmark with TF 1.3 ?

[amrawadk]

Hello @nasica88, I'm not sure, I didn't test the benchmark with TF 1.3.

You can see from this issue #80 that people have had the same trouble with different versions of TF, so I'd suggest switching to a version that has proven to work. tf-nightly-gpu worked for me (which gets the latest, not sure which version it was though) , you can try ppwwyyxx's suggestion as well.

[reedwm]

It's currently impossible to run the benchmark with TF 1.3. As others have stated, you need to use tf-nightly-gpu, which requires Cuda 8 and cuDNN 6.

Alternatively, you can compile from source, which allows you to use later versions of Cuda and cuDNN.

[bzamecnik]

In my case (tf 1.8.0) it was possibly caused by %load_ext line_profiler in a Jupyter notebook.

[nitinmnsn]

for me the same issue was created by dask being inconsistent. try pip install -U dask