lhutton1 commented on a change in pull request #9910:
URL: https://github.com/apache/tvm/pull/9910#discussion_r786768031
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -949,7 +949,10 @@ def concat_func(*inputs):
op = tf.concat(list(inputs), axis)
return op
- _compare_tvm_with_tflite(concat_func, shapes, accel_type)
+ # TODO(lhutton1) For now output is not bit exact with TFLite.
+ # This is because TFLite reference kernels are not being used.
+ # For this, TFLite will need upgrading to 2.6.
+ _compare_tvm_with_tflite(concat_func, shapes, accel_type,
output_tolerance=1)
Review comment:
`qnn.concatenate` is lowered to a concatenate+requantize. Previously the
requantize would not be offloaded, but now it is. There is a numerical
inaccuracy in this case using the current version of TFLite, but we should see
this issue resolved when we upgrade to TFLite 2.6 and use the reference kernels
that are made available.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
