[DRAFT][luci] Introduce MinimumMSE quantization algorithm #12854
Conversation
SlavikMIPT
force-pushed
the
minimum-mse-weights-only
branch
from
e836ee6 to
b2da3dc
Compare
SlavikMIPT
force-pushed
the
minimum-mse-weights-only
branch
from
b2da3dc to
7684e7a
Compare
SlavikMIPT
changed the title
SlavikMIPT
force-pushed
the
minimum-mse-weights-only
branch
from
7684e7a to
2dfc832
Compare
|
Please fix build errors. @SlavikMIPT Can you make the new logic as a new function? How a user can control the algorithm? |
SlavikMIPT
force-pushed
the
minimum-mse-weights-only
branch
2 times, most recently
from
5accef2 to
19f0e71
Compare
SlavikMIPT
force-pushed
the
minimum-mse-weights-only
branch
from
19f0e71 to
9b17a98
Compare
|
@jinevening PTAL |
| @@ -37,6 +37,12 @@ struct LayerInfo | |||
| QuantizationGranularity granularity; | |||
| }; | |||
|
|
|||
| enum struct QuantizationAlgorithm | |||
| { | |||
| Common = 0, | |||
There was a problem hiding this comment.
Q) what is the meaning of Common ?
There was a problem hiding this comment.
Q) what is the meaning of
Common?
General algorithm when scale = max / qmax
There was a problem hiding this comment.
I don't know much about the math itself but it would be better to use a word that can express this.
| const auto kSearchIterations = 100; | ||
| const auto kPhi = 1.618033988749894848204586834365638118; | ||
| const auto kRangeCoefficient = 0.1; |
There was a problem hiding this comment.
How are these parameters determined? Do you have any reference?
There was a problem hiding this comment.
How are these parameters determined? Do you have any reference?
kPhi is golden ratio constant, other are user controlled parameters of algorithm, default values are determined empirically(optimal value of scale is near the max / qmax base scale)
Could you answer this question? Can you make a full draft? This PR is just internal change. |
|
close as inactive for some time, feel free to reopen if needed |
SlavikMIPT
changed the title
SlavikMIPT
added
the
DRAFT
SlavikMIPT
force-pushed
the
minimum-mse-weights-only
branch
from
05b2e7a to
28f6b16
Compare
This commit introduces MinimumMSE quantization algorithm. ONE-DCO-1.0-Signed-off-by: Vyacheslav Bazhenov <[email protected]>
SlavikMIPT
force-pushed
the
minimum-mse-weights-only
branch
from
28f6b16 to
40713f7
Compare
This commit introduces MinimumMSE quantization algorithm.
For #13677
ONE-DCO-1.0-Signed-off-by: Vyacheslav Bazhenov [email protected]