Porting codebase utilizing tf.slim to TF-2.0

[calledbymountains]

A lot of my TensorFlow codebase depends upon the use of pretrained models as a part of tensorflow/models/research/slim/nets repository.

I routinely utilize the dictionary of end points provided by tf.slim models in my codebase. Furthermore, the model definitions in the aforementioned repository is impressive and general in the sense that for models like resnet_v2, it allows me to directly specify the output_stride which leads to an automatic specification of dilation rates.

As I aim to write my new codes in TF-2.0 ( as well as thinking of porting my existing codebase to TF-2.0), I realize (based on documentation and my knowledge of Keras !!), that there is no concept of end point dictionaries and output_stride (wherever applicable).

Hence, I am confused about how to handle these situations.

I believe that it could be very reasonable for the TensorFlow team to provide documentation on this issue for it could really make transition to TF-2.0, very simple and practical.

[ymodak]

To simplify the migration to TensorFlow 2.0, there is a conversion tool which updates TensorFlow 1.x Python code to use TensorFlow 2.0. Please take a look if haven't already. Thanks!

[calledbymountains]

Thank you. But does it specifically address the case of end_point dictionaries and output_stride ?

[ymodak]

@martinwicke Can you PTAL? Thanks!

[martinwicke]

It does not. In fact, it does not handle anything specific to slim.

You can continue to use slim in 2.0, and if you are so dependent on its specifics, you may want to. slim makes heavy use of variable scopes and collections, and writing code using it will never truly feel like TensorFlow 2.0, but using tf.compat.v1 you will be able to make it work.

We will be publishing more detailed conversion guides. I will close this issue since we are already working on that (and we'll only forget to close it otherwise).

[DenisPolygalov]

@martinwicke To make it clear:

1. The upgrade script does not handle this problem.
2. Using tf.compat.v1 will not make it work.
3. The only “more detailed conversion guide” I was able to find 9 month since this issue was opened is only short paragraph here which basically state “it is not simple and it is your problem”.

Your statement was addressed to the person how opened this issue, but let me inform you how many people in fact this issue affects. If anyone at Google cares of course…

I came across this problem while trying to understand what’s going wrong with DeepLabCut ( hello @AlexEMG ) software. The DeepLabCut is a marker-less pose estimation software for analyzing behavior of various subjects in a video streams. It has a Nature Neuroscience paper and become so popular that it "powers a motion-tracking revolution". So as result, many Labs want to use DeepLabCut now because it, well, “have a Nature paper”!

Looking inside of DeepLabCut’s code unveil several severe technical problems however. These problems (so far) do not affect the main effect (e.g. pose detection) but anyone familiar with long term software support understands where it all goes…

DeepLabCut has been built on top of Human Pose Estimation with TensorFlow project also presented in DeeperCut and ArtTrack papers. The DeeperCut rely on ResNet which was in tf.contrib.slim (TF 1.x) and therefore not available in TF 2.0 anymore...

So what do we have here. Google wrote TensorFlow 1.x and despite disastrous code quality it become popular (because it well… “Made by Google”and it is AI!). Then after TF 1.x become popular more and more people began to have severe problems (e.g. computational graph debugging issues) caused by initially disastrous design decisions. Then someone at Google finally realize that it make sense to try to fix everything, so as result TF 2.x was spitted out. As a consequence of that all code from tf.contrib was wiped out of TF 2.x betraying another Google's own creature TF-Slim library because “it will never truly feel like TensorFlow 2.0” leaving many people dealing with many issues by themselves...

This remind me worst practice used many times by Microsoft with their various “cool APIs” dropped 2 years after rolling out.

[martinwicke]

@pichuan @tomerk How far are we with the slim port?

[pichuan]

Hi @martinwicke , I saw your mention here just now.
Short answer: I do not have an ETA.

Long answer:
After my last update here: google-research/tf-slim#1 (comment) , I did spend some time trying to incorporate @tomerk 's suggestion.

This is taking more time than I expected, and I don't think I fully understand the issue (or the solution) to give an ETA.
Note that even if/when I resolve this, there will still be 2 other pending tasks that I don't plan to take on.
Given my schedule in the next few weeks, I won't be able to look further into this right now until at least February.

[ibarrond]

While fighting with a similar porting, I bumped into this colab notebook containing a very juicy paragraph on the subject:

A note on Slim & contrib.layers

A large amount of older TensorFlow 1.x code uses the Slim library, which was packaged with TensorFlow 1.x as tf.contrib.layers. As a contrib module, this is no longer available in TensorFlow 2.0, even in tf.compat.v1. Converting code using Slim to TF 2.0 is more involved than converting repositories that use v1.layers. In fact, it may make sense to convert your Slim code to v1.layers first, then convert to Keras.

• Remove arg_scopes, all args need to be explicit
• If you use them, split normalizer_fn and activation_fn into their own layers
• Separable conv layers map to one or more different Keras layers (depthwise, pointwise, and separable Keras layers)
• Slim and v1.layers have different arg names & default values
• Some args have different scales
• If you use Slim pre-trained models, try out Keras's pre-traimed models from tf.keras.applications or TF Hub's TF2 SavedModels exported from the original Slim code.

Some tf.contrib layers might not have been moved to core TensorFlow but have instead been moved to the TF add-ons package.

I hope it is useful to other people!

[thesiamak]

Are you kidding me?? Why this issue is closed while there is no solution yet?

I spent days to learn Python lang to be able to work with TF. Then I followed instructions on installing and running TFv2.0 explain here. After facing lots of unexpected issues trying to install TF(illegal instruction and building from source ..) finally I'm done installing it on my Ubuntu server. Now, I have no attribute 'contrib' and interestingly, everyone is suggesting me to downgrade to TFv1.x to pass this stage too. Is't this nasty?
I'm exhausted trying to launch TF.