Skip to content

remove roi_align plugin for ORT #258

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

Merged
AllentDan merged 6 commits into from
Mar 24, 2022
Merged

remove roi_align plugin for ORT #258

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
92e177a
remove roi_align plugin
AllentDan Mar 22, 2022
8f8be9a
remove ut
AllentDan Mar 22, 2022
4848201
skip single_roi_extractor UT for ORT in CI
AllentDan Mar 22, 2022
f4cafa8
move align to symbolic and update docs
AllentDan Mar 23, 2022
9e87e13
recover UT
AllentDan Mar 23, 2022
3009d98
resolve comments
AllentDan Mar 23, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
255 changes: 0 additions & 255 deletions csrc/backend_ops/onnxruntime/roi_align/roi_align.cpp
View file
Open in desktop

This file was deleted.

59 changes: 0 additions & 59 deletions csrc/backend_ops/onnxruntime/roi_align/roi_align.h
View file
Open in desktop

This file was deleted.

22 changes: 15 additions & 7 deletions mmdeploy/codebase/mmdet/models/roi_heads/single_level_roi_extractor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@

from mmdeploy.core.optimizers import mark
from mmdeploy.core.rewriters import FUNCTION_REWRITER
from mmdeploy.utils import get_backend
from mmdeploy.utils import get_backend, get_ir_config
from mmdeploy.utils.constants import Backend


Expand Down Expand Up @@ -137,14 +137,22 @@ def single_roi_extractor__forward(ctx,
device=target_lvls.device)
target_lvls = torch.cat((_tmp, _tmp, target_lvls))
for i in range(num_levels):
# use the roi align in torhcvision to accelerate the inference
# roi_align in MMCV is same as torchvision when pool mode is 'avg'
if backend == Backend.TORCHSCRIPT or self.roi_layers[
i].pool_mode == 'avg':
self.roi_layers[i].use_torchvision = True
mask = target_lvls == i
inds = mask.nonzero(as_tuple=False).squeeze(1)
roi_feats_t = self.roi_layers[i](feats[i], rois[inds])
rois_t = rois[inds]
# use the roi align in torhcvision
if backend == Backend.TORCHSCRIPT:
self.roi_layers[i].use_torchvision = True
if backend == Backend.ONNXRUNTIME:
ir_cfg = get_ir_config(ctx.cfg)
opset_version = ir_cfg.get('opset_version', 11)
offset = 0.5 if self.roi_layers[i].aligned is True else 0.0
if opset_version <= 16:
# use RoiAlign in onnxruntime and preprocess rois to
# make compatible with op_set 10
rois_t[:, 1:] = rois_t[:, 1:] - offset / self.roi_layers[
i].spatial_scale
roi_feats_t = self.roi_layers[i](feats[i], rois_t)
roi_feats[inds] = roi_feats_t
# slice to recover original size
roi_feats = roi_feats[num_levels * 2:]
Expand Down
Loading