tensorflow2.0 - Unable to save tensorflow ConvNeXtTiny model in .tf format (TF 2.13.0) - Stack Overflow

admin管理员组

文章数量:1400566

I trained a tensorflow ConvNeXtTiny model (tf.keras.applications.convnext.ConvNeXtTiny). I am able to train the model without any issues. But when I try to save the model with model.save(my_model, save_format="tf") I get the following error:

Traceback (most recent call last):
tensorflow-1  |   File "training_classifier.py", line 12, in <module>
tensorflow-1  |     model.save("my_model", save_format="tf")
tensorflow-1  |   File "/usr/local/lib/python3.8/dist-packages/keras/src/utils/traceback_utils.py", line 70, in error_handler
tensorflow-1  |     raise e.with_traceback(filtered_tb) from None
tensorflow-1  |   File "/usr/local/lib/python3.8/dist-packages/tensorflow/python/trackable/trackable_utils.py", line 126, in escape_local_name
tensorflow-1  |     return (name.replace(_ESCAPE_CHAR, _ESCAPE_CHAR + _ESCAPE_CHAR).replace(
tensorflow-1  | AttributeError: 'NoneType' object has no attribute 'replace'

I did not make any modifications to the model.

This error occurs while trying to save the model both before and after training.

I don't get the error when I save in the h5 format. I don't get this error with other models of tensorflow like Resnet, EfficientNet, etc, but I get it with all sizes of ConvNeXt.

From what I have understood so far, people have only faced this issue when training custom models.

Is there any modification I can do to avoid this error and save my model in the tf format? This is the only format acceptable for me.

本文标签： tensorflow20Unable to save tensorflow ConvNeXtTiny model in tf format (TF 2130)Stack Overflow