Errors when running test scripts about compression

[nmjohn]

What version of TensorFlow are you running? Due to features being relied on by the Keras layer, we require the current Master build (or 1.9.0 when it's released).

[mtisz]

Oh wow, okay. I'm using 1.8.

I didn't know 1.9 is released already. Where can I get it?

[mtisz]

Okay getting the nightly build now - will write back

[nmjohn]

I think 1.9 should be coming out soon. Unfortunately, you'll need to grab either a nightly build or compile from source. We apologize for this extra step as an early adopter but it should not be necessary soon.

[mtisz]

Thanks @nmjohn!

Okay got the nightly build via pip, now I get a different error. The first script is still good.

`*** Received signal 11 ***
*** BEGIN MANGLED STACK TRACE ***
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(+0x6646eb)[0x7f454efcf6eb]
/lib/x86_64-linux-gnu/libpthread.so.0(+0x11390)[0x7f459745c390]
/local/lib/python2.7/site-packages/compression/python/ops/../../_coder_ops.so(+0x725a)[0x7f451b22725a]
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(_ZN10tensorflow6thread10ThreadPool4Impl11ParallelForExxSt8functionIFvxxEE+0x444)[0x7f454efa4b74]
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(_ZN10tensorflow6thread10ThreadPool11ParallelForExxSt8functionIFvxxEE+0x5f)[0x7f454efa4d4f]
/local/lib/python2.7/site-packages/compression/python/ops/../../_coder_ops.so(+0x7545)[0x7f451b227545]
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(_ZN10tensorflow16ThreadPoolDevice7ComputeEPNS_8OpKernelEPNS_15OpKernelContextE+0x9a)[0x7f454ef8a7ba]
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(+0x5d55f9)[0x7f454ef405f9]
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(+0x5d59af)[0x7f454ef409af]
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(_ZN5Eigen26NonBlockingThreadPoolTemplIN10tensorflow6thread16EigenEnvironmentEE10WorkerLoopEi+0x220)[0x7f454efa32a0]
/local/lib/python2.7/site-packages/tensorflow/python/../libtensorflow_framework.so(_ZNSt17_Function_handlerIFvvEZN10tensorflow6thread16EigenEnvironment12CreateThreadESt8functionIS0_EEUlvE_E9_M_invokeERKSt9_Any_data+0x32)[0x7f454efa2002]
/usr/lib/x86_64-linux-gnu/libstdc++.so.6(+0xb8c80)[0x7f4534833c80]
/lib/x86_64-linux-gnu/libpthread.so.0(+0x76ba)[0x7f45974526ba]
/lib/x86_64-linux-gnu/libc.so.6(clone+0x6d)[0x7f459718841d]
*** END MANGLED STACK TRACE ***

*** Begin stack trace ***
tensorflow::CurrentStackTrace()

tensorflow::thread::ThreadPool::Impl::ParallelFor(long long, long long, std::function<void (long long, long long)>)
tensorflow::thread::ThreadPool::ParallelFor(long long, long long, std::function<void (long long, long long)>)

tensorflow::ThreadPoolDevice::Compute(tensorflow::OpKernel*, tensorflow::OpKernelContext*)


Eigen::NonBlockingThreadPoolTempl<tensorflow::thread::EigenEnvironment>::WorkerLoop(int)
std::_Function_handler<void (), tensorflow::thread::EigenEnvironment::CreateThread(std::function<void ()>)::{lambda()#1}>::_M_invoke(std::_Any_data const&)


clone

*** End stack trace ***
Aborted (core dumped)`

[jonycgn]

Hi @migtissera, just to cover the simple things: Have you recompiled the library against the nightly build?

[mtisz]

Hi @jonycgn - Yes I have. Also am trying all nightly builds of 1.9 - Still can't get it to work. The first script runs fine.

[mtisz]

Also, @jonycgn - Good work on the paper! I haven't been able to get the GAN adversarial training to converge well with a VAE. Keen to try your paper!

[jonycgn]

Thanks! Glad to hear :)

Could you try running the following command:
python -c 'import tensorflow as tf; print(tf.version)'

If you're seeing version 1.8, you might still be compiling against the 1.8 version. It might help to make sure that the nightly is on your python path.

[mtisz]

It says AttributeError: 'module' object has no attribute 'version'?

I definitely have the nightly build - even did a clean install on a virtual environment. My pip list is following.

`Package Version

---

absl-py 0.2.1
astor 0.6.2
backports.weakref 1.0.post1
enum34 1.1.6
funcsigs 1.0.2
futures 3.2.0
gast 0.2.0
grpcio 1.12.0
Markdown 2.6.11
mock 2.0.0
numpy 1.14.3
pbr 4.0.3
pip 10.0.1
protobuf 3.5.2.post1
setuptools 39.1.0
six 1.11.0
tb-nightly 1.9.0a20180516
termcolor 1.1.0
tf-nightly-gpu 1.9.0.dev20180516
Werkzeug 0.14.1
wheel 0.31.1 `

[mtisz]

Okay I just noticed this when tensorflow session starts up.

2018-05-16 15:25:06.854167: W tensorflow/core/framework/op_kernel.cc:1290] CtxFailure at constant_op.cc:418: Invalid argument: You must feed a value for placeholder tensor 'Placeholder_1' with dtype int32 and shape [3]

[jonycgn]

Hey,

I can reproduce the error. We're looking into it, but due to the NIPS deadline on Friday we might not be able to get back to you until next week. Meanwhile, please post any other helpful information here.

Thanks!

[mtisz]

@jonycgn Okay! Good luck on NIPS - you'll be good! :)

If I figure out a workaround will post.

[nmjohn]

@migtissera , I apologize for making you try additional things. But would you try compiling the TensorFlow library on your machine? I'm finding the issue occurs when linking the op library with the prebuild binary wheels, but not compiling from source.

I'm leaving the issue open as I determine how best to fix this issue, but if you're eager to try the library / code, I'd recommend installing from source until I can resolve: https://www.tensorflow.org/install/install_linux

[mtisz]

Hey @nmjohn and @jonycgn, thanks for the quick responses!
I figured out a workaround. The nightly build has these layers in the coder in tf.contrib, so I'm calling them directly now.
I do have a few questions regarding implementing the hyper prior now though! :) Is there a good way to connect with you guys? Specifically, are you implementing two bottleneck layers? One for the main autoencoder and the other for the hyper network? If so, I'm curios to know more about the training setup (whether you minimize both bottlenecks and both squared errors).
I can introduce myself directly to you guys too (if you're wondering). :)
Thanks heaps for being really quick in responses and proactive!

[jonycgn]

Hi,

yes, you can use this workaround for now. The code should be identical up to changes in documentation. In the long run, tf.contrib.coder is going to go away though, so we still need to figure out why compiling from scratch is no problem and using the nightly builds is!

Regarding the hyperprior, please feel free to shoot me an email. We are also in the process of setting up a Google group for library-related discussions at https://groups.google.com/forum/#!forum/tensorflow-compression. Feel free to post there, but be aware that we're still setting things up!

[jonycgn]

Hi @migtissera,

the problem is in different compiler settings used for the nightlies compared to when building from scratch, which makes the binary we compile here incompatible with the TF binaries. We're still looking into it (it's complicated).

Meanwhile, I've added a workaround such that the entropy bottleneck class uses the existing range coder implementation in tf.contrib for now. Can you confirm that the entropy bottleneck test now succeeds (the range coder test might fail, but it shouldn't matter if you don't want to use it directly)?

[mtisz]

Hi @jonycgn,

Both tests succeeded when I called the layers in tf.contrib.coder, so it was all good. Is the range coder implementation here different than the one in tf.contrib.coder?

I'll email you today regarding the hyperprior.

[mtisz]

Also, I can confirm the bottleneck layer is working. I've gotten this to work with your instructions on the readme. I had to figure out how to get the cdf variable when saving graphs but apart from that the instructions are pretty clear. Well done!

[nmjohn]

It seems like this is currently working, I'm going to mark as closed unless there's other issues.