[Feature] Add mmpose configs for compatibility with Hourglass model in ncnn backend (#1064)

* [FEATURE] add mmpose configs for compatibility with Hourglass model

[Motivation]
- The added configs are aim to support Hourglass pose-detection model in ncnn backend.
- Hourglass model needs explicit input shape with 256x256 or 384x384.
- Hourglass model can be quantitized to int8 successfully in ncnn backend.

[Complementary]
The mmpose config of Hourglass mode is in:
- https://github.com/open-mmlab/mmpose/blob/master/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/hourglass52_coco_256x256.py
- https://github.com/open-mmlab/mmpose/blob/master/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/hourglass52_coco_384x384.py
Quantitization results are shown in https://github.com/Qingrenn/mmdeploy-summer-camp/blob/main/week4/result.md

* Keep only one config (pose-detection_ncnn-int8_static-256x256.py) and remove the others.
pull/1068/head
Qingren 2022-09-19 14:13:55 +08:00 committed by GitHub
parent 18ccb15b7b
commit fd07b26c2e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 3 additions and 0 deletions

View File

@ -0,0 +1,3 @@
_base_ = ['./pose-detection_static.py', '../_base_/backends/ncnn-int8.py']
onnx_config = dict(input_shape=[256, 256])