TensorFlow Lite tests failing with TensorFlow 2.7+

[guillaumekln]

The following test cases are failing with TensorFlow 2.7 and are currently disabled:

FAILED opennmt/tests/tflite_test.py::TFLiteTest::testTFLiteInterpreter_0
FAILED opennmt/tests/tflite_test.py::TFLiteTest::testTFLiteInterpreter_1
FAILED opennmt/tests/tflite_test.py::TFLiteTest::testTFLiteInterpreter_3

cc @gcervantes8

[gcervantes8]

It seems that with TensorFlow 2.7, the models that are no longer exporting with TensorFlow Lite are:

• RNN-based models with Luong Attention decoder
• Relative Transformer models

I tried fixing the error with the RNN models by changing the output shape, and attention shape in the decoding while loop back to tf.TensorShape(()), but that didn't seem to fix it.

I don't really know how to fix the error with Relative Transformer models, it's a strange error.

The good thing about TensorFlow 2.7 is that quantization now works for Transformer models.