Added tflite frontend support for quantized mean #4339
Conversation
|
cc @FrozenGene @srkreddy1238 @anijain2305 can you take a look? |
python/tvm/relay/frontend/tflite.py
Outdated
| @@ -659,7 +659,23 @@ def _convert_reduce(self, relay_op, op): | |||
| reduce_options.Init(op_options.Bytes, op_options.Pos) | |||
| keep_dims = reduce_options.KeepDims() | |||
|
|
|||
| if input_tensor.qnn_params: | |||
| in_expr = _qnn.op.dequantize(data=in_expr, | |||
There was a problem hiding this comment.
Hmm...do we have to dequantize -> compute fp32 -> requantize? could we do only integer computation?
There was a problem hiding this comment.
Similar concern as that of @FrozenGene
From maths size, it looks safe to me, but let me know if I am wrong
Initial Quantized tensor - QA - shape (N, dtype='int8')
Mean would be
scale_output * (Qout - zp_out) = scale_in * [(QA[0] + QA[1] .... + QA[N-1]) - N* zp_in]/N
scale_output * (Qout - zp_out) = scale_in * (Mean(QA) - zp_in)
So, basically, upcast the quantized tensor to int32, call Int Mean, and then requantize if the output scale/zp are different.
There was a problem hiding this comment.
Updated as suggested above
| @@ -710,16 +710,23 @@ def test_forward_zeros_like(): | |||
| # Reduce | |||
| # ------ | |||
|
|
|||
| def _test_reduce(math_op, data, keep_dims=None): | |||
| def _test_reduce(math_op, data, keep_dims=None, quantized=False): | |||
There was a problem hiding this comment.
Suggest breaking into two functions. Another is _test_reduce_quantize
There was a problem hiding this comment.
Suggest breaking into two functions. Another is _test_reduce_quantize
Done
tristan-arm
force-pushed
the
master
branch
3 times, most recently
from
3beaf12 to
c05fd08
Compare
|
CI failure is unrelated to this patch - I'll kick it off again once the issue is resolved |
|
@tristan-arm CI is consistently failing. Shouldn't relate to your change. |
|
Thanks @tristan-arm @anijain2305 @FrozenGene |
No description provided.