TensorFlow 2.12.0 breaks nightly tests.

[fatcat-z]

self = <test_backend.BackendTests testMethod=test_Conv2DBackpropInput>
tflite_path = '/tmp/tmp9j283eve/test_Conv2DBackpropInput/test_Conv2DBackpropInput.tflite'

def tflite_has_supported_types(self, tflite_path):
    try:
        with open(tflite_path, 'rb') as f:
            buf = f.read()
        buf = bytearray(buf)
        model = Model.GetRootAsModel(buf, 0)
        tensor_cnt = model.Subgraphs(0).TensorsLength()
        interpreter = tf.lite.Interpreter(tflite_path)
        for i in range(tensor_cnt):

          dtype = interpreter._get_tensor_details(i)['dtype']   # pylint: disable=protected-access

E TypeError: Interpreter._get_tensor_details() missing 1 required positional argument: 'subgraph_index'

[pranavladkat]

Is there an ETA for this issue?

[q-ycong-p]

+1 on this issue. Any update on a fix?

[q-ycong-p]

I put up a naive workaround #2204, just to start discussion towards a fix. Welcome more thoughtful fix, especially if folks have deeper knowledge on this tflite's change wrt subgraph at tf-2.12, and tf2onnx's corresponding path forward.

[UPDATE]
Updated PR to directly use TFLite Interpreter's public API instead. I am not seeing an obvious advantage of using the private API. Pls point out if I'm missing anything there! Ty.

[q-ycong-p]

Hi @fatcat-z, all tests passed after re-opening #2204 - which aims to address this issue. Would appreciate a feedback/review from maintainers. Thank you!

[yan12125]

This issue can probably be closed as #2204 is merged?