mirrors/mmdeploy
1
0
Fork 0
mirror of https://github.com/open-mmlab/mmdeploy.git synced 2025-01-14 08:09:43 +08:00
Code Issues Projects Releases Wiki Activity

Add mmseg configs for compatibility with Fast-SCNN model in ncnn backend (#1094)

Browse Source 
* mmdeploy summer camp

* fix lint

* add note

* add note

* add a comment in interpolate.py

Co-authored-by: root <root@LAPTOP-A20M40H8.localdomain>
This commit is contained in:
lansfair 2022-10-27 10:36:26 +08:00 committed by GitHub
parent 2d35c9ab35
commit b6f8c1c50d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 35 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
configs/mmseg/segmentation_ncnn-int8_static-1024x2048.py Normal file
View File

@ -0,0 +1,3 @@
_base_ = ['./segmentation_static.py', '../_base_/backends/ncnn-int8.py']
onnx_config = dict(input_shape=[2048, 1024])

11
docs/en/03-benchmark/quantization.md
View File

@ -47,3 +47,14 @@ Note: MMPose models are tested with `flip_test` explicitly set to `False` in mod
| :-----------------------------------------------------------------------------------------------------------------: | :-----: | :------------: | :------------: |
| [EDSRx2](https://github.com/open-mmlab/mmediting/blob/master/configs/restorers/edsr/edsr_x2c64b16_g1_300k_div2k.py) | Set5 | 35.7733/0.9365 | 35.4266/0.9334 |
| [EDSRx4](https://github.com/open-mmlab/mmediting/blob/master/configs/restorers/edsr/edsr_x4c64b16_g1_300k_div2k.py) | Set5 | 30.2194/0.8498 | 29.9340/0.8409 |
### mmseg
| model | dataset | fp32 mIoU | int8 mIoU |
| :----------------------------------------------------------------------------------------------------------------------------: | :--------: | :-------: | :-------: |
| [Fast-SCNN](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/fastscnn/fast_scnn_lr0.12_8x4_160k_cityscapes.py) | cityscapes | 70.96 | 70.24 |
Note:
- Int8 models of the Fast-SCNN requires ncnnoptimize.
- NCNN will extract 512 images from the train as a calibration dataset

11
docs/zh_cn/03-benchmark/quantization.md
View File

@ -47,3 +47,14 @@
| :-----------------------------------------------------------------------------------------------------------------: | :-----: | :------------: | :------------: |
| [EDSR](https://github.com/open-mmlab/mmediting/blob/master/configs/restorers/edsr/edsr_x2c64b16_g1_300k_div2k.py) | Set5 | 35.7733/0.9365 | 35.4266/0.9334 |
| [EDSRx4](https://github.com/open-mmlab/mmediting/blob/master/configs/restorers/edsr/edsr_x4c64b16_g1_300k_div2k.py) | Set5 | 30.2194/0.8498 | 29.9340/0.8409 |
### 语义分割任务
| model | dataset | fp32 mIoU | int8 mIoU |
| :----------------------------------------------------------------------------------------------------------------------------: | :--------: | :-------: | :-------: |
| [Fast-SCNN](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/fastscnn/fast_scnn_lr0.12_8x4_160k_cityscapes.py) | cityscapes | 70.96 | 70.24 |
备注：
- Fast-SCNN 的int8模型需要使用ncnnoptimize优化。
- NCNN将会从train中抽取512张图片作为校准集。

7
mmdeploy/pytorch/functions/interpolate.py
View File

@ -22,13 +22,16 @@ def interpolate__ncnn(ctx,
"""Rewrite `interpolate` for ncnn backend.
ncnn require `size` should be constant in ONNX Node. We use `scale_factor`
instead of `size` to avoid dynamic size.
instead of `size` to avoid dynamic size. To avoid rounding errors, add a
small number when `scale_factor` is not an integer
"""
input_size = input.shape
if scale_factor is None:
scale_factor = [
s_out / s_in for s_out, s_in in zip(size, input_size[2:])
s_out / s_in if int(s_out / s_in) == s_out / s_in else
(s_out / s_in + 0.00001)
for s_out, s_in in zip(size, input_size[2:])
]
return ctx.origin_func(

6
mmdeploy/pytorch/ops/adaptive_pool.py
View File

@ -10,4 +10,8 @@ def adaptive_avg_pool2d__ncnn(ctx, g, x, output_size):
Align symbolic of adaptive_avg_pool2d in ncnn.
"""
return g.op('mmdeploy::AdaptiveAvgPool2d', x, output_size)
size = getattr(output_size.node(), 't')('value').tolist()
if size != [1, 1]:
return g.op('mmdeploy::AdaptiveAvgPool2d', x, output_size)
else:
return g.op('GlobalAveragePool', x)