Tensor2tensor: AttributeError: module 'tensorflow.contrib.data' has no attribute 'parallel_interleave'

[wachaong]

File "/home/work/wangchao/python/lib/python3.6/site-packages/tensor2tensor/data_generators/problem.py", line 521, in dataset
tf.contrib.data.parallel_interleave(
AttributeError: module 'tensorflow.contrib.data' has no attribute 'parallel_interleave'

t2t version : 1.4.3
tf version : 1.4.1

[martinpopel]

T2T 1.4.3 needs TF 1.5 (as stated in setup.py).

[wachaong]

3q

[Alexyitx]

i also meet this question ,can you tell me how to solve it?

[martinpopel]

@lihongwei521: update to the newest T2T and if possibly also TF>=1.5.
There is a check so maybe it works with older TF as well.

[RunshengZhu]

@martinpopel the check you provided works~ my problem solved , training process well in tf1.4, thx!

[Yangandtom]

i meet this issue too in tf1.4, could you tell me how to fix it,please?

[martinpopel]

@Yangandtom: If you cannot update to newer TF, you must downgrade T2T to a version which supports TF1.4 (this info is in setup.py).

[Yangandtom]

the setup,py is mean? I find the setup.py in the check which you provided, the version='1.5.6'

[martinpopel]

Here you can see that T2T 1.6.0 requires tensorflow>=1.5.0.
And here you can see that T2T 1.5.7 required tensorflow>=1.4.1.
Similarly, e.g. T2T 1.2.9 supported tensorflow>=1.3.0

[Yangandtom]

I tried to install T2T 1.2.9 in tf 1.4.0. But this issue still exists.

[samanthawyf]

@Yangandtom @martinpopel I met the same problem, module has no attribute 'parallel_interleave'
My tensorflow is 1.4.0 and I have not installed T2T. From your reply, I get that T2T 1.2.9 supported tensorflow>=1.3.0, I am wondering could it solve the problem by installing T2T 1.2.9. And I also want to ask how to install T2T 1.2.9, as command pip install tensor2tensorinstalls newest version.

[martinpopel]

pip install tensor2tensor==1.2.9

[samanthawyf]

@martinpopel Thanks for your reply. After I installed T2T 1.2.9, I still meet the problem as @Yangandtom . I doubt tensorflow1.4 doesn't support tensorflow.contrib.data.parallel_interleave?

[simo23]

If you are using TF1.4, a non-optimal solution is to call the function sloppy_interleave TF1.4Docs, which I believe is the older version of the function.

This worked for me.

[levinwil]

This is because you are using an old version of tensorflow.

The easiest fix is to install the latest version of tensorflow.

If you need to be using your current version of tensorflow, perhaps for compatability with CUDA 8, you can change the line to the following:
"records_dataset = filename_dataset.interleave(file_read_func, cycle_length=config.num_readers, block_length=config.read_block_length)"

[paullinzipeng]

@levinwil This worked like magic to me! Could you tell me where to find this code?

[levinwil]

I wrote it myself. Glad I could help!

[paullinzipeng]

Thank you! I've been trying to fix this problem for days!

[guilhermeh2m]

@ levinwil Thanks a lot man! Its work for me in Tensorflow 1.4
I comment the "records_dataset=..." in the file "...tensorflow/models/research/object_detection/builders/dataset_builder.py"

and add your line instead:
"records_dataset = filename_dataset.interleave(file_read_func, cycle_length=config.num_readers, block_length=config.read_block_length)"

and works to train mobilenet with tensorflow API