[QNN] Channel wise quantization - Quantize & Requantize #4629
Conversation
anijain2305
force-pushed
the
channelwise_qnn
branch
3 times, most recently
from
cf1fba5
to
d692126
Compare
anijain2305
changed the title
anijain2305
force-pushed
the
channelwise_qnn
branch
from
d692126
to
bfa50a7
Compare
|
|
||
| auto expanded_output_scale = output_scale; | ||
| if (!IsConstScalar(output_scale)) { | ||
| expanded_output_scale = ExpandBiasToMatchAxis(output_scale, n_dim, {axis}); |
There was a problem hiding this comment.
what about renaming ExpandBiasToMatchAxis to something more generic?
There was a problem hiding this comment.
This helper is used in many other passes. We can keep this name for now. And then send a separate PR for better naming across all those files.
anijain2305
force-pushed
the
channelwise_qnn
branch
from
bfa50a7
to
5a94896
Compare
Thanks for the helpful comments @jackwish All the comments were very good and have been addressed :) |
|
@tmoreau89 Could you please manage the PR? |
|
@zhiics on it, thanks |
anijain2305
force-pushed
the
channelwise_qnn
branch
from
5a94896
to
602965f
Compare
|
@shoubhik Thanks for the comments. Addressed. |
|
Thank you @anijain2305 , @shoubhik , @jackwish ; the PR has been merged |
| .describe("The output channel axis for channel wise quantization. Default value is -1," | ||
| "which corresponds to the last axis.") |
There was a problem hiding this comment.
I don't have an opinion on whether this is the right approach or not but the question that popped in my mind was :
Why not be more explicit and say that axis == -1 actually means per-tensor quantization ?
Thanks,
Ramana
There was a problem hiding this comment.
Never mind , later in the thread things are a bit clearer. The axis of -1 is fine - it is part of the per-channel quantization scheme .
In some sense I would have preferred to see a IsPerChannelQuantized predicate, rather than reusing a IsConstScalar predicate further below but that's a code hygiene thing.
And , yes sorry about the slow response, I'm a bit snowed under.
RFC - https://discuss.tvm.ai/t/qnn-channel-wise-quantization/5300