[TFLite] Enable reference int8 fully connected with output shape with dimensions > 2 #54223
Conversation
haozha111
requested review from
jianlijianli
and removed request for
haozha111
…o toupstream/fully_connected_dims_2
…QuantizedOutputShape3DInt8 tests
google-ml-butler
bot
added
kokoro:force-run
|
Could you add a e2e test case under the lite_v2_test.py? |
gbaned
removed
awaiting review
SaoirseARM
force-pushed
the
toupstream/fully_connected_dims_2
branch
from
ea13c59
to
4badb5f
Compare
SaoirseARM
force-pushed
the
toupstream/fully_connected_dims_2
branch
from
4badb5f
to
a8ddaaf
Compare
|
Hi, Thanks for review. I have added end to end test to lite_v2_test.py. Please let me know if there is anything I need to add/change. Best regards, |
| @@ -34,12 +34,13 @@ inline void FullyConnected( | |||
| const int32_t output_activation_min = params.quantized_activation_min; | |||
| const int32_t output_activation_max = params.quantized_activation_max; | |||
| TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2); | |||
| TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2); | |||
| TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1); | |||
There was a problem hiding this comment.
One more thing: shouldn't this be TFLITE_DCHECK_GE(.., 2) ?
There was a problem hiding this comment.
Hi,
I think this is ok, as the output shape can have 1 dimension? This is similar implementation to: 1e0716e#diff-2cb25dbe2523a0754af5b5057b800add3642578751a4083dd094b5b99ecdecb2
Best regards,
Saoirse
| interpreter.allocate_tensors() | ||
| output_details = interpreter.get_output_details() | ||
|
|
||
| self.assertLen(output_details[0]['shape_signature'], 3) |
There was a problem hiding this comment.
could you add value comparisons between TF and TFLite models?
There was a problem hiding this comment.
Thanks for review. I have added some extra comparisons. Please let me know if any more needs to be updated.
Best regards,
Saoirse
…dims_2 PiperOrigin-RevId: 438793489
|
Seems auto-merge is not happening but the changes are merged into master now, so we can close this. Thank you for the PR. |
3D output shape support for Fully_Connected layer under xtensa. The fix to the reference kernels was made with tensorflow/tensorflow#54223 (and the corresponding import from TF Lite to TFLM). TFLM currently does not have a test case for this (corresponding to what was added with tensorflow/tensorflow#54223), BUG=tensorflow/tensorflow#53501
Hi,
This PR enables reference int8 support for fully connected with output shape with dimensions > 2 in TensorFlow Lite.
Relevant issue which describes this is here: #53501
This PR also contains an update to the SimpleTestQuantizedOutputShape3DInt16 to ensure that output shape with dimensions >2 is verified.
Thanks,
Saoirse