Universal Sentence Encoder module loading stall

[Aaron Tuor]

Loading the ELMo model is pretty fast: 

elmo = hub.Module("https://tfhub.dev/google/elmo/1", trainable=True)

but the universal sentence encoder module doesn't load for me:

embed = hub.Module("https://tfhub.dev/google/universal-sentence-encoder/1")

 Can someone check to make sure this is working properly?

Thanks,

Aaron

[George Sterpu]

Yes, it works here, now. This is a large 1GB model however, is there any progress in /tmp/tfhub_modules/* ?

--
You received this message because you are subscribed to the Google Groups "TensorFlow Hub" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hub+uns...@tensorflow.org.
Visit this group at https://groups.google.com/a/tensorflow.org/group/hub/.

[Aaron Tuor]

Ok, I will check. 

[s...@carlabs.com]

When I run

embed = hub.Module("https://tfhub.dev/google/universal-sentence-encoder/1")

I get the following error, at the bottom of a 10 deep stack

KeyError: "The name 'global_step:0' refers to a Tensor which does not exist. The operation, 'global_step', does not exist in the graph."

Sorry if this isn't the right place for this, but it is happening directly from the example colab notebook, so I don't see how it could be user error. Though it probably still is.

[Abhishek Singh]

Got the same error too. If anything works for you do post.

[André Susano Pinto]

Thanks for reporting... this is related with tensorflow pre-release "1.8.0rc0".. I will investigate further and update.

For now you can manually fallback to tensorflow 1.7.0 (You might need to restart your runtime for it to work).

[Abhishek Singh]

Thanks a lot.

[André Susano Pinto]

Hi all,

This turned out to be a issue with tensorflow-1.8.0rc0 and not with tensorflow-hub package.

Stick with "pip install tensorflow==1.7.0" until there is a new release candidate.

See https://github.com/tensorflow/tensorflow/issues/18598 for details.

--

André

[Abhishek Singh]

Hi all, 

Are there any licensing terms involved. I mean can I use all those pre trained weights commercially.

[André Susano Pinto]

That is off-topic from this thread about the module failing to load. See https://github.com/tensorflow/hub/issues/12 instead.

[rajeev gupta]

Whenever I run this module it takes very long to run even a test program. Why doesn't it cache the module rather than loading every time? Is there any way to download the module and refer the local file system location instead of the url?

-Rajeev

[Andrew Gasparovic]

Hi Rajeev,

In fact, TF Hub does cache modules! Check out the section "Caching Modules" here:

https://www.tensorflow.org/hub/basics

The issue is that it will use a temp location, which may not be retained between runs. You can control the cache location using the TFHUB_CACHE_DIR environment variable. Thanks!

--

/A

[rajeev gupta]

Thanks for replying Andrew. I am using some other location for caching. Seems most of the time is going in calculating embedding for each sentence. The test run has 10 sentences and it takes around 5 minutes to get their embeddings.

-Rajeev

--

Thanks and Regards,
Rajeev Gupta