Export not working

Hi, I’m using Perceptilabs version 0.12.21.

The problem I have is that while trying to export a model, both in TensorFlow Model and Jupyter Notebook format, I receive the message “Exported with success.”, but no file or folder is created. (Also, apparently, there are also no warnings or errors in the JS console related to the click on the Export button).

The model I’m trying to export has an Input Text and Target Categorical.
In the Model Hub, the model is marked as “Training complete”, no test run (also because apparently I can’t run any of the available tests on this model), and Not Exported.

Dunno if I’m missing something relevant here.

Thanks.

It’s a known bug and the devs are working on it :slight_smile:

1 Like

Hi, I updated to version 0.12.25. Now it “correctly” says “Model export failed”, using TensorFlow Model format, with any combination of Compressed and Quantized toggles, but it gives me no info on what the actual error is, or what can I do to resolve it.

Nothing also appears in the cmd console, so, I’m kinda stuck.

Hi @lc555,
I’m sorry to hear that you still are having an issue!
All TensorFlow exports should be working now, I was exporting a few models just earlier this week.

When you say that nothing appears in the cmd console, are you running the tool as “perceptilabs -v=3” to enable more logging?

Also, can you try to train one of the tutorial datasets to completion (mnist for 1 epoch with the recommended model for example) and see if that can be exported?

export error.txt (10.6 KB)

Hi, sorry, didn’t know about the verbosity option. I run it again with the verbosity 3, and here you found attached the stack trace related to the export operation.

The “real” error seems to be “The last dimension of the inputs to Dense should be defined. Found None.”, but this means little to me (again, I’m quite the newbie when it comes to AI). I searched around the various options of my basic pipeline (the 4 basic pre-defined building blocks, and the input configuration), and no option that could be set to “None” is actually set to “None”, at least in the UI, but, again, I may have missed it, because I don’t really know where to search.

About the trying of the tutorial datasets, I managed to make them work in a previous Perceptilabs version, but didn’t actually run the export. I’ll try to do it as soon as I can, and report back to you.

Hmm strange, I haven’t seen that error before.
Let me know how it goes with the tutorial dataset (I’m very curious), and from there we can dig into the issue :slight_smile:

Hi, I confirm that using a newly-created model using one of the tutorial datasets (Covid19, in my case) the export indeed works.

To be extra sure, I also created a new model using my dataset, with the same configuration as the one failing (input text - target categorical, and no other options touched), and that still fails with the same error ( The last dimension of the inputs to Dense should be defined. Found None.)

Thanks.

In the error.txt file: if there is no code for DeepLearningFC_Dense_1Keras, the function would probably “produce” nothing… hence dimension = None???

A stab in the dark, is all :wink:

Hmm, it could also be the text data doing something strange.
Would you be able to share a small part of your dataset @lc555 so we can test it internally? :slight_smile:

@JulianSMoore, very true, the question is why it’s not able to fetch the code though :thinking:

Ja, thanks @robertl, I’ll send you a DM.

Ah, damn, I see there is a messaging system here, but I can’t seem able to send new private messages (or maybe I just don’t see the button, I feel dumb).
@robertl, there is any other way I can send you the file?

@lc555 if you hop on our slack you can send me a DM there :slight_smile:

Done, thanks!

Text to reach 20 characters.