Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

V10Parser node factory improvements #3824

Merged
merged 4 commits into from
Jan 13, 2021
Merged

V10Parser node factory improvements #3824

merged 4 commits into from
Jan 13, 2021

Conversation

vladislav-volkov
Copy link
Contributor

@vladislav-volkov vladislav-volkov commented Jan 12, 2021 •
edited
Loading

After the analysis some issues were found in V10Parser::createNode function:

  1. Opset is cloned when nGraph node is created
  2. Default opsets operation checking complexity can be reduced from O(N) to O(1)
  3. Supported opsets checking also slightly can be optimized by replacing versions strings comparison inside loop by single condition.

@vladislav-volkov vladislav-volkov changed the title Dev v10parser V10Parser node factory improvements Jan 12, 2021
@openvino-pushbot openvino-pushbot added the category: Core OpenVINO Core (aka ngraph) label Jan 12, 2021
ilyachur
ilyachur reviewed Jan 13, 2021
View reviewed changes
Copy link
Contributor

@ilyachur ilyachur left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just for education, do we have any performance numbers for this improvement?

inference-engine/src/readers/ir_reader/ie_ir_parser.cpp Show resolved Hide resolved
@ilya-lavrenov ilya-lavrenov added category: IR FE OpenVINO IR v10 / v11 FrontEnd and removed category: Core OpenVINO Core (aka ngraph) labels Jan 13, 2021
@ilya-lavrenov ilya-lavrenov added this to the 2021.3 milestone Jan 13, 2021
@vladislav-volkov
Copy link
Contributor Author

vladislav-volkov commented Jan 13, 2021 •
edited
Loading

Just for education, do we have any performance numbers for this improvement?

Model Read(master) Read(optimized) Diff
FP16/face-detection-adas-0001 23.58 13.45 -42.96%
FP16/shufflenet 16.06 9.46 -41.10%
FP32/face-detection-adas-0001 24.87 14.71 -40.85%
FP32/shufflenet 17.88 11.26 -37.02%
FP16/cocosnet 399.77 317.63 -20.55%
FP16/bert-small-uncased-whole-word-masking-squad-0001 180.04 143.27 -20.42%
FP16/brain-tumor-segmentation-0001 88.12 72.29 -17.96%
FP32/bert-small-uncased-whole-word-masking-squad-0001 287.22 249.53 -13.12%
FP32/cocosnet 626.92 547.99 -12.59%
FP16/faster-rcnn-resnet101-coco-sparse-60-0001 101.54 89.8 -11.56%
FP16/resnet-50-pytorch 50.36 44.73 -11.18%
FP16/ssd300 52.09 46.36 -11.00%
FP32/brain-tumor-segmentation-0001 145 133.48 -7.94%
FP16/yolo-v3 128.41 118.56 -7.67%
FP16/u-net 55.98 52.08 -6.97%
FP32/resnet-50-pytorch 86.01 80.29 -6.65%
FP32/faster-rcnn-resnet101-coco-sparse-60-0001 173.4 162.55 -6.26%
FP32/ssh 79.02 74.3 -5.97%
FP16/darknet19 40.9 38.48 -5.92%
FP32/ssd300 89.6 84.37 -5.84%
FP32/yolo-v3 215.23 204.95 -4.78%
FP32/u-net 97.89 93.29 -4.70%
FP32/resnet-101-3d 263.22 252.4 -4.11%
FP16/ssh 49 47.33 -3.41%
FP32/caffenet 177.72 172.36 -3.02%
FP32/darknet19 71.33 69.34 -2.79%
FP16/GNMT 358.16 348.7 -2.64%
FP16/resnet-101-3d 149.73 147.48 -1.50%
FP32/GNMT 680.04 670.53 -1.40%
FP16/caffenet 92.84 91.61 -1.32%

@ilyachur ilyachur merged commit b9447df into openvinotoolkit:master Jan 13, 2021
mryzhov pushed a commit to mryzhov/openvino that referenced this pull request Jan 14, 2021
@vladislav-volkov @mryzhov
V10Parser node factory improvements (openvinotoolkit#3824)
b3b543b 
* Supported opsets check improvement

* Optimization of V10 nodes creation in nGraph

* Redundant type checking in opset removed

* opset2 existance check added
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@ilyachur ilyachur ilyachur approved these changes

@ilya-lavrenov ilya-lavrenov ilya-lavrenov approved these changes

@dmitry-gorokhov dmitry-gorokhov Awaiting requested review from dmitry-gorokhov

Assignees

@ilyachur ilyachur

Labels
category: IR FE OpenVINO IR v10 / v11 FrontEnd optimization
Projects
None yet
Milestone
2021.3
Development

Successfully merging this pull request may close these issues.
4 participants
@vladislav-volkov @ilya-lavrenov @ilyachur @openvino-pushbot