d-smirnov commented on a change in pull request #6018:
URL: https://github.com/apache/incubator-tvm/pull/6018#discussion_r466509216
##########
File path: tests/python/frontend/tflite/test_forward.py
##########
@@ -250,7 +256,7 @@ def compare_tflite_with_tvm(in_data, in_name, input_tensors,
# convert to tflite model
converter = tf.lite.TFLiteConverter.from_session(
sess, input_tensors, output_tensors)
-
+ converter.experimental_new_converter=experimental_new_converter
Review comment:
I understood that it is not an experimental feature any more. However
the name "experimental_new_converter" was preserved. I don't see any harm to
use this feature and have this test especially if we plan to migrate to a newer
version of TFLite.
##########
File path: python/tvm/relay/frontend/tflite.py
##########
@@ -1089,7 +1093,7 @@ def convert_square(self, op):
return out
- def _convert_elemwise(self, relay_op, op):
+ def _convert_elemwise(self, relay_op, op, use_real_qnn=True):
Review comment:
The extraction of "use_real_qnn" functionality to _convert_minimum and
_convert_maximum methods (L1225 and L1229) will lead either: to change to
operation's parameters stripping the qnn_attrs from at least lhs input tensor
or to addition an extra flag forces _convert_elemwise to use non quantized
version of the operation. Alternatively I might not understand your point here.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
