Feature request: Inception v3 MetaGraph

[Hvass-Labs]

Would it be possible for the TensorFlow developers to put a tar-ball online with the Inception v3 saved as a MetaGraph? I can't find it anywhere.

I'm currently using the following tar-ball with a frozen graph for Inception v3:

http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz

The problem is that I cannot continue training that graph because it is frozen, so all the variables have been converted to constants before it was saved. I can't find a way to convert the constants back to variables so I don't think that is possible. (There are also some deprecation warnings regarding BatchNormWithGlobalNormalization so it will presumably stop working at some point in the future).

After searching for a solution for days, I found that you have released a newer checkpoint-file for Inception v3:

http://download.tensorflow.org/models/image/imagenet/inception-v3-2016-03-01.tar.gz

I downloaded it but it's only the checkpoint-file, not the graph-definition. So in my Python code I would apparently have to create the Inception graph using this function first:

https://github.com/tensorflow/models/blob/master/inception/inception/slim/inception_model.py#L52

But this apparently requires building TensorFlow from source, as far as I could understand from the README. There's also several options for using the function and it apparently has to be wrapped in arg_scopes and what-not:

https://github.com/tensorflow/models/blob/master/inception/inception/inception_model.py#L76-87

Would it be possible to update the above tar-ball (dated 2016-03-01) so it also contains the MetaGraph-files, so I can load it more easily and use it in my own Python program? I have another data-set so I replace the softmax-layer of the Inception-graph, and I also want to continue optimizing the rest of the variables of the Inception-graph.

Please also consider including a small example program in the tar-ball (or a link to some python-code), as it would make it a lot easier for everyone who wants to use it. Or at least make a list of all the relevant tensor-names (input, output, etc.)

Thanks!

[drpngx]

@martinwicke any reason why we don't ship the graphs?

[martinwicke]

No reason other than finiteness of resources. sguada@ may have more details, but I believe that the code in tensorflow/models produces the proper graph, and could serve as the requested code example. I don't think building from source is a requirement. I believe the (Python) files would work fine with an installed binary version of tensorflow. You may have to check out tensorflow if the files are not contained in the pip package.

[ryszard]

@martinwicke I am trying to do something very similar to what @Hvass-Labs is attempting.

The inception code in tensorflow/models is doing rather advanced stuff (using multiple GPUs, etc), which goes over the head of a beginner (like me), and it seems rather tightly coupled. What is more, it uses a (custom) version of slim, which not only makes the code look different than mainstream TF code, but also forces you to use blaze (not super pleasant if you aren't already using blaze for everything) – you cannot just drop one file into your repo and count it will work.

(Another source of confusion for me is that there's two versions of the slim inception: one in tensorflow/models, and one in https://github.com/tensorflow/models/tree/master/inception/inception/slim, which are very similar, but slightly different).

The inception related code that's much more beginner friendly (and much less Google idiosyncratic) is the one used in https://www.tensorflow.org/versions/r0.9/how_tos/image_retraining/index.html. However, adapting it to training the full network has proven quite daunting (issues mentioned by @Hvass-Labs, like being frozen, etc).

So, getting the Metagraph would actually be very, very helpful. Some non-slim Python code would be even better, but I understand that the rewrite would be far from trivial.

[Hvass-Labs]

Thanks for the input everyone.

I am using the Inception model in my tutorials and I try to keep them fairly simple. This means limiting each tutorial to only one or a few topics. The official tutorials have 10 topics in each tutorial which is an extremely steep learning curve for beginners. I've put the Inception model inside a Python module so it is encapsulated and can easily be loaded from different tutorials:

https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/inception.py

By far the easiest thing would be to replace a few lines in my inception.py with the following:

• Download a newer tar-ball.
• A few lines of code for loading the MetaGraph and Checkpoint.

This would be an elegant solution that I could easily explain in a tutorial.

For the end-user, there really is no reason to build the Inception graph all over again. It is quite messy and you can't make any changes to it, because then you cannot load the checkpoint anymore. Even tiny changes that are logically identical will break the checkpoint.

As far as I understand, the MetaGraph is already saved when you save the checkpoint. So it should be fairly trivial to update the tar-ball. Although I do understand that if you guys are under a lot of stress then there's a million 'fairly trivial' things that needs to be done.

Hopefully you'll see the usefulness of updating the tar-ball as it might help thousands of people who will be watching the tutorials in the following years. I've been stuck for 2 weeks now trying to obtain the MetaGraph so I'm hoping I can get it soon.

Cheers!

[jurastm]

Same about me. I need to use tar-ball in my project

[paolodoors]

Hvass-Labs's tutorials are really great. It would be very helpful to have the MetaGraph available.

[jfhauris]

Hvass-Labs's tutorials are outstanding. Please provide the MetaGraph etc that he is requesting. Thanks, Jon

[chidochipotle]

Hvass-Labs's tutorials are amazing. Would be great if you could implement the MetaGraph so he can make tutorial 10. Thanks, Daniel

[Arora90]

Hvass-Labs's tutorials are very helpful. Would be great if you could implement the MetaGraph.

[souterjk]

Bump for asking for a meta graph. It's become critical for me because the .pb file which is included with the tars of 2015's inception5h model appear to use an API which has since been deprecated

[Caspain]

Hvass-Labs's tutorials are outstanding. Please provide the MetaGraph etc that he is requesting.Everyone deeply appreciates his lessons. Kindly support his request
Thank You.

[prudhvinandan]

Hvass-Labs's tutorials are outstanding. Please provide the MetaGraph etc that he is requesting.

[karl596]

Hvass-Labs的教程非常出色。请提供他要求的MetaGraph等。

[Caspain]

Please support. Give the required files

…

On Feb 16, 2017 7:36 PM, "karl596" ***@***.***> wrote: Hvass-Labs的教程非常出色。请提供他要求的MetaGraph等。 — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#5036 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQAuomMNPiYTBbKMeOWPOc1D7fhk4uX6ks5rdRW_gaJpZM4KZgU8> .

[jurastm]

@karl596 tensorflow 刚出了 v.1.0.
最新版应该有metagraph

[vshkola]

+1 for the really necessary feature.

[Caspain]

Please support this feature request. So the community surrounding tensor can mature further,thank You.

…

On Tue, Mar 14, 2017 at 10:52 AM, Vladyslav Shkola ***@***.*** > wrote: +1 for the really necessary feature. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#5036 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQAuov9ZON-YfH24NLEoBlcuyb3GMCi9ks5rlrfGgaJpZM4KZgU8> .

[Caspain]

Hvass-Labs's tutorials are outstanding. Please provide the MetaGraph etc that he is requesting. Thus extending the usage of this great and best machine learning library. On Tue, Mar 14, 2017 at 10:53 AM, remario richards <remariorich@gmail.com> wrote:

…

Please support this feature request. So the community surrounding tensor can mature further,thank You. On Tue, Mar 14, 2017 at 10:52 AM, Vladyslav Shkola < ***@***.***> wrote: > +1 for the really necessary feature. > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub > <#5036 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AQAuov9ZON-YfH24NLEoBlcuyb3GMCi9ks5rlrfGgaJpZM4KZgU8> > . >

[KoSv]

please support!

[Caspain]

please support this feature request.thank you!

…

On Sat, Mar 18, 2017 at 8:47 AM, van ***@***.***> wrote: please support! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#5036 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQAuogNLGn3cQqMa-wwDm6U-ftNLNy61ks5rm-BngaJpZM4KZgU8> .

[evaliyev]

+1

[ljubisap]

+1

[maximegmd]

+1

[merryHunter]

Please, support! +1

[attakorn1]

+1

[singhranjodh]

+1

[shrikondra]

+1

[mueller91]

+1

[alwynmathew]

+1

[YaroslavSchubert]

+1

[tanayjha]

+1

[gongenhao]

+1

[ram1993]

+1

[bzhong2]

+1

[Vijayabhaskar96]

+1

[trojana]

+1

[hzqjyyx]

+1

[kencano]

+1

[wgong]

Hvass-Labs tensorflow tutorial is awesome, the best I can find, please support his pull request, so we can learn more about tensorflow from Hvass-Labs, Thank tensorflow team

[thorgeirk11]

+1

[gnavihs]

+1

[jematos92]

+1

[sguada]

Please use export_inference_graph to get the graph (see Docs)

[ajinkyaT]

+1

[sguada]

Fixed the links

[grankt]

+1

[AymanASamyM]

+1

[letuantang]

+1

[luongthomas]

+1

[letuantang]

+1

[omarg42]

+1

[nimral]

+1

[AbhiRepository]

+1

[Hvass-Labs]

Thanks to everyone for supporting this issue!

The easiest solution is to just use Keras for doing Transfer Learning and Fine-Tuning in TensorFlow.

• https://keras.io/applications/#usage-examples-for-image-classification-models
• https://github.com/fchollet/deep-learning-with-python-notebooks/blob/master/5.3-using-a-pretrained-convnet.ipynb
• https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/10_Fine-Tuning.ipynb