Output discrepancy between keras.predict and tf_saved_model

[edge7]

Hi,
I am exporting a model like this:

       import keras

        export_archive = keras.export.ExportArchive()
        export_archive.track(model_tra)
        export_archive.add_endpoint(
            name="serve",
            fn=lambda x: model_tra.call(x, training=False),
            input_signature=[
                keras.InputSpec(shape=(None, 3, 320, 320, 1), dtype="float64")
            ],
        )
        export_archive.write_out(
            "/home/edge7/Desktop/projects/ing_edurso/wharfreenet-inference/models/la_ao"
        )

Then I load it like:
tf.saved_model.load(os.path.join(MODEL_DIRECTORY, "la_ao"))

and I use it like:

model.serve(np.expand_dims(np.array(frame), axis=-1))

I get slightly different results than just using:
model.predict or model.predict_on_batch

am I missing something silly in the conversion?

[edge7]

The bug might be linked to #19403
I get different results when using predict_on_batch and then serve (which should use model(x, training=False)) given the way I exported it.
It is probably not a conversion issue.

[SamanehSaadat]

Hi @edge7

Would it be possible to share your model code? Or could you share a repro colab?

[SamanehSaadat]

Another question: Do you get different results calling model.predict or model.predict_on_batch right after training the model, i.e. without going through the export process?

[edge7]

Hi @SamanehSaadat
Please find here a colab, that is self contained.
It downloads the models plus the model definition on the fly.
Then it tests with Keras (Jax), tf_model, and onnx.

The "nice thing" is that in local (with my machine, still GPU) I get even different results than with Colab