What's the best practice to save and load trained/fine-tuned models e.g. BertClassifier?

[alvarobartt]

Hi @jbischof! I was wondering what's the best way to save and load any ...Classifier model e.g. BertClassifier that has been trained/fine-tuned via .fit on custom data.

I assume that it should be .save and then tf.keras.models.load_model, but the latter is complaining due to missing inputs and outputs (TypeError: Functional.__init__() missing 2 required positional arguments: 'inputs' and 'outputs').

So I was wondering whether there's a clear approach on how to properly save and load models. I saw some discussions about some proposals of HuggingFace-like interfaces, but not sure if there's been any clear decision on that.

Thanks!

[alvarobartt]

P.S. for more context, I was experimenting with BertClassifier.from_preset("bert_base_en", num_classes=2) to fine-tune it for GLUE SST-2 for experimentation purposes. And looking not just for a solution on how to load any fine-tuned model, but also on how to do that easily (user-friendly speaking).

[alvarobartt]

Since for the moment you have from_preset which does that but only for the available models, there's not an easy way to e.g. register a new preset, or load a fine-tuned model using the same preset schema, etc.

[jbischof]

model.save() should serialize the weights along with the architecture. What's the feature gap?

presets are an implementation detail of the library....unless we go from a curated library to full 🤗

[alvarobartt]

Yes, it does, but one's not able to keras.models.load_model("<MODEL_PATH>"), as it's expecting both inputs and outputs too, so there's no way to load a fine-tuned model with the path only, you need to know more details in advance, so there's not a clear way on how to do that in an abstract way, so that anyone can load any model just with the model path (to the SavedModel files).

[jbischof]

That sounds like a bug. Could you please file an issue?

See https://keras.io/guides/serialization_and_saving/ as a general guide for how saving should work.

[alvarobartt]

Sure @jbischof, creating it rn!