[Solved] Errors in converting Tensorflow Model to Tensorflow Lite Model

[aparico]

I wanted to use yolov4-tiny in the Tensorflow lite framework.

I converted my darknet weights trained from @AlexeyAB using these commands:

python save_model.py --weights yolov4-tiny.weights --output ./checkpoints/yolov4-tiny-608-tf --input_size 608 --model yolov4 --tiny --framework tflite
python convert_tflite.py --weights ./checkpoints/yolov4-tiny-608-tf --output ./checkpoints/yolov4-tiny-608.tflite

The first command is successful using numpy==1.19.0. However, the second one shows these errors.

loc("batch_normalization/moving_mean"): error: is not immutable, try running tf-saved-model-optimize-global-tensors to prove tensors are immutable
Traceback (most recent call last):
  File "C:\Python37\lib\site-packages\tensorflow\lite\python\convert.py", line 213, in toco_convert_protos
    enable_mlir_converter)
  File "C:\Python37\lib\site-packages\tensorflow\lite\python\wrap_toco.py", line 38, in wrapped_toco_convert
    enable_mlir_converter)
Exception: <unknown>:0: error: loc("batch_normalization/moving_mean"): is not immutable, try running tf-saved-model-optimize-global-tensors to prove tensors are immutable


During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "convert_tflite.py", line 76, in <module>
    app.run(main)
  File "C:\Python37\lib\site-packages\absl\app.py", line 303, in run
    _run_main(main, args)
  File "C:\Python37\lib\site-packages\absl\app.py", line 251, in _run_main
    sys.exit(main(argv))
  File "convert_tflite.py", line 71, in main
    save_tflite()
  File "convert_tflite.py", line 45, in save_tflite
    tflite_model = converter.convert()
  File "C:\Python37\lib\site-packages\tensorflow\lite\python\lite.py", line 762, in convert
    result = _convert_saved_model(**converter_kwargs)
  File "C:\Python37\lib\site-packages\tensorflow\lite\python\convert.py", line 648, in convert_saved_model
    enable_mlir_converter=True)
  File "C:\Python37\lib\site-packages\tensorflow\lite\python\convert.py", line 216, in toco_convert_protos
    raise ConverterError(str(e))
tensorflow.lite.python.convert.ConverterError: <unknown>:0: error: loc("batch_normalization/moving_mean"): is not immutable, try running tf-saved-model-optimize-global-tensors to prove tensors are immutable

I have tried other versions of Tensorflow (2.2, 2.3, 2.4 with numpy==1.19.0) but I had no luck. What should I do?

There is a similar issue raised here: Tensorflow Issue 44790

System information:
Windows 10, x64
GeForce GTX 1060
NVIDIA Driver 460.89
CUDA 11.0.3
CuDNN 8.0.5.39
Python 3.7.2

[Adithia99]

yeah same issue , i work on google colab and get same error

loc("batch_normalization/moving_mean"): is not immutable, try running tf-saved-model-optimize-global-tensors to prove tensors are immutable

[Marouaneghoulami]

Have anyone found the solution for this yet ?

[Marouaneghoulami]

Maybe try training again your model with :

tensorflow-gpu==2.3.0rc0 opencv-python==4.1.1.26

[rahulzag]

same issue, please fix

[rahulzag]

well, I had done this before on a dataset and it had worked but now I used another dataset and tweaked the batches and subdivisions does it have something to do with this error?

[Adithia99]

Can u show me the collab and the dataset all od your data?

…

On Wed, 30 Dec 2020, 21:56 Marouaneghoulami, ***@***.***> wrote: Maybe try training again your model with : tensorflow-gpu==2.3.0rc0 opencv-python==4.1.1.26 — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#285 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANKLHMUBCYBQRSBCFXAVKU3SXNERNANCNFSM4VFHFU3A> .

[Marouaneghoulami]

Can u show me the collab and the dataset all od your data?
…
On Wed, 30 Dec 2020, 21:56 Marouaneghoulami, @.***> wrote: Maybe try training again your model with : tensorflow-gpu==2.3.0rc0 opencv-python==4.1.1.26 — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#285 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ANKLHMUBCYBQRSBCFXAVKU3SXNERNANCNFSM4VFHFU3A .

I'm afraid its confidential, but i faced too the same problem, it was because tensorflow has a new release 2.4.0
but when i uninstalled it and re installed the version as mention above, that seem to solve the problem for me.

[rahulzag]

can someone please convert the model to tflite and send me it would be a great help this is the model file -https://drive.google.com/drive/folders/1-1jAHa_N2YoA28sLCa2a0xJFkEfJu4Td?usp=sharing

[gouliang1997]

try tf-nightly

[mikeDonahoe]

In case it helps. In google colab I had this issue if I used the default TF version which was 2.4.0. Running !pip install tensorflow==2.3.0 and restarting the runtime, then converting corrected the issue.

[leosapucaia]

In case it helps. In google colab I had this issue if I used the default TF version which was 2.4.0. Running !pip install tensorflow==2.3.0 and restarting the runtime, then converting corrected the issue.

Thanks! This solution works nicely!

[aparico]

In case it helps. In google colab I had this issue if I used the default TF version which was 2.4.0. Running !pip install tensorflow==2.3.0 and restarting the runtime, then converting corrected the issue.

Thanks for this! I will close this issue now.

[fmigone]

It works for me. Thanks a lot.

[khg2478]

@leosapucaia @mikeDonahoe @fmigone @Marouaneghoulami @aparico

Hello all, I've changed the package as mentioned above, tensorflow-gpu==2.3.0rc0 opencv-python==4.1.1.26, but they don't help resolving the issue still. I have uploaded the results. I have my custom model that slightly changes configurations such as resolution. Please help.

[Marouaneghoulami]

@khg2478 The solution mentioned above is related to conversion from SavedModel to Tflite.

As for your problem, you should check the code you used for rescaling.

[khg2478]

@khg2478 The solution mentioned above is related to conversion from SavedModel to Tflite.

As for your problem, you should check the code you used for rescaling.

Thanks for your reply. Perhaps, it's because of the rescaling problem. Could you tell me how I can fix this then? I did change the size parameter in detect.py as my trained model, 960x960 but still the problem remains. I thought it was caused by the model conversion issue.

[Marouaneghoulami]

@khg2478 i think u should just check you rescaling code, perhaps you missed up width and height when drawing boxes.

to check whether you converted the model successfully or not please use this site : https://netron.app/ and see output and input sizes of the model

[khg2478]

@khg2478 i think u should just check you rescaling code, perhaps you missed up width and height when drawing boxes.

to check whether you converted the model successfully or not please use this site : https://netron.app/ and see output and input sizes of the model

@Marouaneghoulami Thanks for your reply. I tried my own scaling to fix the problem. After obtaining boxes, I ran np.amax function to check the largest value and found out that it is 1.
So, I resized the value as below since I figured out that it multiplies image's shape in utils.draw_bbox function.
boxes[:,:,0::2] = boxes[:,:,0::2] / 960 (model input shape)
boxes[:,:,1::2] = boxes[:,:,1::2] / 960
However, the problem still remains.. The model works fine in darknet but the converted tflite doesn't work properly.. I think I need guidance to fix the issue. Any ideas?