Save TextVectorization Model to load it later
Question:
I’m not used to the TextVectorization Encoder Layer. I created my vocabulary manually before. I was wondering how one can save a Keras Model which uses the TextVectorization layer. When I tried to do it with simply model.save() and later models.load_model() I was prompted with this error:
AssertionError: Found 1 Python objects that were not bound to checkpointed values, likely due to changes in the Python program. Showing 1 of 1 unmatched objects: [<tensorflow.python.ops.lookup_ops.MutableHashTable object at 0x7fb9602df7c0>]
Answers:
I’ve solved my problem by using another version of Keras. If someone faces a similar issue I can recommend to use a different (most of the time newer) version of Keras.
As I already said in my comment. I can’t really recommend Keras and or Tensorflow right now. I’ve started a big NLP project some time ago (half a year). And since then Keras had multiple updates. Their documents changed like 2 times. And the old examples are not there anymore. The new way to create Text Tokens is quite nice but their example uses Masking_zero=True. Which basically means that It will pad the sequences for you and following layers will ignore the zero. That sounds nice but masking is not compatible with Cuda which makes training larger models a time consuming job because it’s not hardware accelerated with the GPU. And most NLP models are quite large.
I’m not used to the TextVectorization Encoder Layer. I created my vocabulary manually before. I was wondering how one can save a Keras Model which uses the TextVectorization layer. When I tried to do it with simply model.save() and later models.load_model() I was prompted with this error:
AssertionError: Found 1 Python objects that were not bound to checkpointed values, likely due to changes in the Python program. Showing 1 of 1 unmatched objects: [<tensorflow.python.ops.lookup_ops.MutableHashTable object at 0x7fb9602df7c0>]
I’ve solved my problem by using another version of Keras. If someone faces a similar issue I can recommend to use a different (most of the time newer) version of Keras.
As I already said in my comment. I can’t really recommend Keras and or Tensorflow right now. I’ve started a big NLP project some time ago (half a year). And since then Keras had multiple updates. Their documents changed like 2 times. And the old examples are not there anymore. The new way to create Text Tokens is quite nice but their example uses Masking_zero=True. Which basically means that It will pad the sequences for you and following layers will ignore the zero. That sounds nice but masking is not compatible with Cuda which makes training larger models a time consuming job because it’s not hardware accelerated with the GPU. And most NLP models are quite large.