ValueError upon model initialization

[np-n]

I am always getting the exception ValueError: Tensor("Const_1:0", shape=(3,), dtype=float32) must be from the same graph as Tensor("stack:0", shape=(1, 1024, 1024, 3), dtype=float32) (graphs are <tensorflow.python.framework.ops.Graph object at 0x7f7a8cd8ee50> and <tensorflow.python.framework.ops.Graph object at 0x7f7a180613d0>). while initializing the model using model = ocr_predictor(pretrained=True).

How to handle the above exception? When I started to debug, I found that it is unable to get repr. i.e. Unable to get repr for <class 'doctr.models.predictor.tensorflow.OCRPredictor>. I have described more at :https://stackoverflow.com/questions/72210187/unable-to-get-repr-for-class-doctr-models-predictor-tensorflow-ocrpredictor

[felixdittrich92]

Hi @np-n 👋 ,
Can you please post your code and the output from: https://github.com/mindee/doctr/blob/main/scripts/collect_env.py ? :)
Edit: and the shape from your algned_image please

[np-n]

yes, I have done everything!

[felixdittrich92]

please try:

import tensorflow as tf
tf.keras.backend.clear_session()
model = ocr_predictor(pretrained=True)
x = model([aligned_image])

This is really strange 😅

[np-n]

Yes, it works. Could you help me by explaining how it works after clearing the backend session? Thank you

[felixdittrich92]

(graphs are <tensorflow.python.framework.ops.Graph object at 0x7f7a8cd8ee50> and <tensorflow.python.framework.ops.Graph object at 0x7f7a180613d0>) indicates, that there is already a Graph object into memory after clearing the session this cannot be the problem :)
I cannot explain why this happens (has nothing to do with docTR maybe a good issue for the TF repo) but i have noticed this quite often, especially when writing test cases for docTR 😅

[np-n]

Thank you so much for your help!. Now, I am wondering to know, how to train DoCTR for the different languages. I am looking for an OCR for Devanagari scripts. Could you please guide me through the approach that I should follow in order to train it for the Devanagari Language?

[frgfm]

For people looking for solutions, as suggested by Felix here, the problem was coming from keras graph referencing in a loop. Resetting the session index of keras should do the trick :)