dnn: add attention layer#24476

Merged

asmorkalov merged 26 commits intoopencv:4.xfrom

fengyuentau:attention_layer

Dec 20, 2023

Merged

dnn: add attention layer#24476
asmorkalov merged 26 commits intoopencv:4.xfrom
fengyuentau:attention_layer

Commits

Commits on Dec 20, 2023

first commit
fengyuentau
committed
resolve comments; add packWeight
fengyuentau
committed
add single-thread fastGemm; implement the first matmul in attention
fengyuentau
committed
fix several bugs to get correct internal results
fengyuentau
committed
quick followup fix
fengyuentau
committed
attention prob impl
fengyuentau
committed
complete impl
fengyuentau
committed
add test case
fengyuentau
committed
add perf
fengyuentau
committed
support v_Slice.end=INT64_MAX; support single-head attention subgraph; add test for attention subgraph fusion
fengyuentau
committed
add acc tests (commented for now); clear qkv_hidden_sizes everytime it is matched; clean comments
fengyuentau
committed
handle optional inputs in graph simplifier
fengyuentau
committed
add perf and acc test for vittrack (comment for now)
fengyuentau
committed
revert graph simplifier changes before rebase
fengyuentau
committed
fix graph simplifier
fengyuentau
committed
clear perf results
fengyuentau
committed
cpu only attention subgraph fusion
fengyuentau
committed
use OPENCV_DNN_BACKEND_DEFAULT
fengyuentau
committed
changes by review
fengyuentau
committed
slice up to 5 inputs
fengyuentau
committed
add acc and perf tests
fengyuentau
committed
empty commit to trigger tests
fengyuentau
committed
empty commit to trigger tests 1
fengyuentau
committed
try to make ci green by fixing output dimension problem
fengyuentau
committed
fix shape; set weight path of vittrack optional
fengyuentau
committed
Skit new test for CUDA FP16 for now.
asmorkalov
committed

Movatterモバイル変換

Navigation Menu

Search code, repositories, users, issues, pull requests...

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

dnn: add attention layer#24476

dnn: add attention layer#24476
asmorkalov merged 26 commits intoopencv:4.xfrom
fengyuentau:attention_layer

Commits

Commits on Dec 20, 2023

first commit

resolve comments; add packWeight

add single-thread fastGemm; implement the first matmul in attention

fix several bugs to get correct internal results

quick followup fix

attention prob impl

complete impl

add test case

add perf

support v_Slice.end=INT64_MAX; support single-head attention subgraph; add test for attention subgraph fusion

add acc tests (commented for now); clear qkv_hidden_sizes everytime it is matched; clean comments

handle optional inputs in graph simplifier

add perf and acc test for vittrack (comment for now)

revert graph simplifier changes before rebase

fix graph simplifier

clear perf results

cpu only attention subgraph fusion

use OPENCV_DNN_BACKEND_DEFAULT

changes by review

slice up to 5 inputs

add acc and perf tests

empty commit to trigger tests

empty commit to trigger tests 1

try to make ci green by fixing output dimension problem

fix shape; set weight path of vittrack optional

Skit new test for CUDA FP16 for now.

Movatterモバイル変換

Uh oh!

dnn: add attention layer#24476asmorkalov merged 26 commits intoopencv:4.xopencv/opencv:4.xfromfengyuentau:attention_layerfengyuentau/opencv:attention_layerCopy head branch name to clipboard

Commits

Commits on Dec 20, 2023

dnn: add attention layer#24476
asmorkalov merged 26 commits intoopencv:4.xfrom
fengyuentau:attention_layer