Skip to content

Add High-res FasterRCNN MobileNetV3 and tune Low-res for speed#3265

Merged
datumbox merged 6 commits intopytorch:masterfrom
datumbox:mobilenetv3_highres
Jan 19, 2021
Merged

Add High-res FasterRCNN MobileNetV3 and tune Low-res for speed#3265
datumbox merged 6 commits intopytorch:masterfrom
datumbox:mobilenetv3_highres

Conversation

@datumbox
Copy link
Copy Markdown
Contributor

@datumbox datumbox commented Jan 19, 2021 •
edited
Loading

  • Rename low-res model.
  • Tune low-res model for speed (2x faster) by sacrificing 0.02 mAP.
  • Add High-res model.

Low Resolution: Faster R-CNN MobileNetV3-Large 320 FPN

Heavily optimized for speed. Good for actual mobile usage.

Weight checkpoint:

PR3253/3rd_training/34945231/model_23.pth

Validate:

python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py\
 --dataset coco --model fasterrcnn_mobilenet_v3_large_320_fpn --test-only --pretrained

Accuracy metrics:

IoU metric: bbox
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.228
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.380
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.232
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.026
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.219
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.444
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.215
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.290
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.294
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.037
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.295
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.570

Speed Benchmark: 0.1679 sec per image on CPU

High Resolution: Faster R-CNN MobileNetV3-Large FPN

Offers good balance between speed and accuracy, significantly faster than the model with the resnet backbone without sacrificing too much accuracy.

Weight checkpoint:

PR3253/high_res/34974236/model_25.pth

Validate:

python -m torch.distributed.launch --nproc_per_node=2 --use_env train.py\
 --dataset coco --model fasterrcnn_mobilenet_v3_large_fpn --test-only --pretrained

Accuracy metrics:

0: IoU metric: bbox
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.328
0:  Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.525
0:  Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.343
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.127
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.363
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.502
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.287
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.427
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.445
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.196
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.499
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.648

Speed Benchmark: 0.8409 sec per image on CPU

@datumbox
Tag fasterrcnn mobilenetv3 model with 320, add new inference config t…
16927a2 
…hat makes it 2x faster sacrificing a bit of mAP.
@datumbox datumbox changed the title [WIP] Add High-res FasterRCNN MobileNetV3 and tune Low-res for speed Add High-res FasterRCNN MobileNetV3 and tune Low-res for speed Jan 19, 2021
@datumbox datumbox requested a review from fmassa January 19, 2021 12:26
fmassa
fmassa approved these changes Jan 19, 2021
View reviewed changes
Copy link
Copy Markdown
Member

@fmassa fmassa left a comment

Choose a reason for hiding this comment

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

Looks great, thanks!

@datumbox datumbox mentioned this pull request Jan 19, 2021
Closed
13 tasks
@datumbox datumbox force-pushed the mobilenetv3_highres branch 2 times, most recently from 518c422 to 66d861f Compare January 19, 2021 14:14
@datumbox datumbox merged commit ae0d80b into pytorch:master Jan 19, 2021
@datumbox datumbox deleted the mobilenetv3_highres branch January 19, 2021 14:39
facebook-github-bot pushed a commit that referenced this pull request Jan 21, 2021
@mthrok @facebook-github-bot
Add High-res FasterRCNN MobileNetV3 and tune Low-res for speed (#3265)
ce2b49c 
Summary:
* Tag fasterrcnn mobilenetv3 model with 320, add new inference config that makes it 2x faster sacrificing a bit of mAP.

* Add a high resolution fasterrcnn mobilenetv3 model.

* Update tests and expected values.

Reviewed By: datumbox

Differential Revision: D25954564

fbshipit-source-id: f6b64981d2bc83e3577435481a569df38b427b20
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@fmassa fmassa fmassa approved these changes

Assignees

No one assigned

Labels

cla signed

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@datumbox @fmassa @facebook-github-bot