Translating model from tf_saved_model to mhlo fails about mlir-emitc HOT 4 CLOSED

As a general note, please be aware that the conversion is quite fragile at the moment. We forked from __init__.py of IREE and adjusted it for initial experiments, but cannot guarantee that it will work for arbitrary models.

Regarding the first issue, I assume the pass needs to be adjusted somehow, although I cannot comment how. Regarding the second issue, I so far only tested an MobileNetv2, passed as savemodel. Maybe @simon-camp can spend some time to investigate next week.

from mlir-emitc.

I tried MobileNetv2 conversion from tf model zoo and it met with a similar error as the second case:

2020-09-25 16:02:29.064358: E tensorflow/compiler/mlir/tensorflow/translate/tf_mlir_translate.cc:163] SavedModel import failed: 
Invalid argument: Exported function with exported name(s) '__call__', 'keras_api.__call__' with multiple concrete functions. Add 
@tf.function(input_signature=[...]) on this function, or use a narrower list of exported names that excludes this function.

module {
}

However I was able to overcome this issue for both the models by passing --tf-savedmodel-exported-names=predict flag to mlir-translate.

Furthermore, for tf-opt the addition of following passes caused the first issue: tf-saved-model-optimize-global-tensors, tf-saved-model-freeze-global-tensors. These passes were not present in the IREE pipeline as well and removing them helped. I had to allow partial xla to tf legalization (xla-legalize-tf=allow-partial-conversion) as tf.ReadVariableOp was not converted - I think this is expected. I hope this is of some help, thank you for all your responses.

from mlir-emitc.

I had success wrapping a keras model in a tf.Module, see the conversion script. Then narrow down the exported functions in the translation via --tf-savedmodel-exported-names=predict as you noted.

The conversion script will most likely not work as is on tf.keras.Model subclasses as these are saved in SavedModel format AFAIK.

from mlir-emitc.

Since this seems to be solved, at least as far as we can support, I am closing the issue.

from mlir-emitc.