Skip to content

add postprocessing_masks gpu version #276

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
grimoire merged 4 commits into from
Apr 6, 2022
Merged

add postprocessing_masks gpu version #276

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
42f61ae
add postprocessing_masks gpu version
Mar 25, 2022
f9587b1
default device cpu
Mar 30, 2022
0458f32
pre-commit fix
Mar 31, 2022
6f05691
Merge branch 'master' of github.com:open-mmlab/mmdeploy
Apr 6, 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
35 changes: 21 additions & 14 deletions mmdeploy/codebase/mmdet/deploy/object_detection_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,7 @@ def __init__(self, backend: Backend, backend_files: Sequence[str],
super().__init__(deploy_cfg=deploy_cfg)
self.CLASSES = class_names
self.deploy_cfg = deploy_cfg
self.device = device
self._init_wrapper(
backend=backend, backend_files=backend_files, device=device)

Expand Down Expand Up @@ -114,6 +115,7 @@ def postprocessing_masks(det_bboxes: np.ndarray,
det_masks: np.ndarray,
img_w: int,
img_h: int,
device: str = 'cpu',
mask_thr_binary: float = 0.5) -> np.ndarray:
"""Additional processing of masks. Resizes masks from [num_det, 28, 28]
to [num_det, img_w, img_h]. Analog of the 'mmdeploy.codebase.mmdet.
Expand All @@ -138,17 +140,25 @@ def postprocessing_masks(det_bboxes: np.ndarray,
return np.zeros((0, img_h, img_w))

if isinstance(masks, np.ndarray):
masks = torch.tensor(masks)
bboxes = torch.tensor(bboxes)
masks = torch.tensor(masks, device=torch.device(device))
bboxes = torch.tensor(bboxes, device=torch.device(device))

result_masks = []
for bbox, mask in zip(bboxes, masks):

x0_int, y0_int = 0, 0
x1_int, y1_int = img_w, img_h

img_y = torch.arange(y0_int, y1_int, dtype=torch.float32) + 0.5
img_x = torch.arange(x0_int, x1_int, dtype=torch.float32) + 0.5
img_y = torch.arange(
y0_int,
y1_int,
dtype=torch.float32,
device=torch.device(device)) + 0.5
img_x = torch.arange(
x0_int,
x1_int,
dtype=torch.float32,
device=torch.device(device)) + 0.5
x0, y0, x1, y1 = bbox

img_y = (img_y - y0) / (y1 - y0) * 2 - 1
Expand All @@ -169,10 +179,8 @@ def postprocessing_masks(det_bboxes: np.ndarray,
grid[None, :, :, :],
align_corners=False)

mask = img_masks
mask = (mask >= mask_thr_binary).to(dtype=torch.bool)
result_masks.append(mask.numpy())
result_masks = np.concatenate(result_masks, axis=1)
result_masks.append(img_masks)
result_masks = torch.cat(result_masks, 1)
return result_masks.squeeze(0)

def forward(self, img: Sequence[torch.Tensor], img_metas: Sequence[dict],
Expand Down Expand Up @@ -206,6 +214,8 @@ def forward(self, img: Sequence[torch.Tensor], img_metas: Sequence[dict],
if isinstance(scale_factor, (list, tuple, np.ndarray)):
assert len(scale_factor) == 4
scale_factor = np.array(scale_factor)[None, :] # [1,4]
scale_factor = torch.from_numpy(scale_factor).to(
device=torch.device(self.device))
dets[:, :4] /= scale_factor

if 'border' in img_metas[i]:
Expand All @@ -216,7 +226,7 @@ def forward(self, img: Sequence[torch.Tensor], img_metas: Sequence[dict],
y_off = img_metas[i]['border'][0]
dets[:, [0, 2]] -= x_off
dets[:, [1, 3]] -= y_off
dets[:, :4] *= (dets[:, :4] > 0).astype(dets.dtype)
Copy link
Member

@grimoire grimoire Apr 1, 2022

Choose a reason for hiding this comment

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

Why do you remove astype?

Copy link
Contributor Author

@xizi xizi Apr 1, 2022

Choose a reason for hiding this comment

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

dets的数据类型是torch.tensor，dtype是float32，不需要astype

Copy link
Member

@grimoire grimoire Apr 1, 2022

Choose a reason for hiding this comment

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

(dets[:, :4] > 0) will give you a bool tensor.

Copy link
Contributor Author

@xizi xizi Apr 1, 2022

Choose a reason for hiding this comment

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

torch.tensor dtype float32与bool 相乘会自动转换为float32

Copy link
Member

@grimoire grimoire Apr 1, 2022 •
edited
Loading

Choose a reason for hiding this comment

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

I see. A better way is using torch.nn.functional.threshold, which might be faster.

import torch
from torch.profiler import profile, ProfilerActivity


def func_test(func,
              data,
              threshold,
              num_test=10,
              activate=ProfilerActivity.CPU):

    # warmup
    for _ in range(num_test):
        out = func(data, threshold)

    with profile(activities=[activate], record_shapes=True) as prof:
        for _ in range(num_test):
            out = func(data, threshold)

    print(prof.key_averages().table(
        sort_by="cpu_time_total", top_level_events_only=True, row_limit=30))

    return out


def main():
    data = torch.rand(1, 3, 224, 224)
    threshold = 0.5
    activate = ProfilerActivity.CPU

    out0 = func_test(
        lambda data, threshold: data * (data > threshold),
        data,
        threshold,
        activate=activate)
    out1 = func_test(
        lambda data, threshold: torch.nn.functional.threshold(
            data, threshold, 0),
        data,
        threshold,
        activate=activate)
    torch.testing.assert_allclose(out0, out1)


if __name__ == '__main__':
    main()

Copy link
Contributor Author

@xizi xizi Apr 1, 2022

Choose a reason for hiding this comment

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

牛皮。

dets[:, :4] *= (dets[:, :4] > 0)

dets_results = bbox2result(dets, labels, len(self.CLASSES))

Expand All @@ -234,16 +244,14 @@ def forward(self, img: Sequence[torch.Tensor], img_metas: Sequence[dict],
'export_postprocess_mask', True)
if not export_postprocess_mask:
masks = End2EndModel.postprocessing_masks(
dets[:, :4], masks, ori_w, ori_h)
dets[:, :4], masks, ori_w, ori_h, self.device)
else:
masks = masks[:, :img_h, :img_w]
# avoid to resize masks with zero dim
if rescale and masks.shape[0] != 0:
masks = masks.astype(np.float32)
masks = torch.from_numpy(masks)
masks = torch.nn.functional.interpolate(
masks.unsqueeze(0), size=(ori_h, ori_w))
masks = masks.squeeze(0).detach().numpy()
masks = masks.squeeze(0)
if masks.dtype != bool:
masks = masks >= 0.5
segms_results = [[] for _ in range(len(self.CLASSES))]
Expand All @@ -267,7 +275,6 @@ def forward_test(self, imgs: torch.Tensor, *args, **kwargs) -> \
"""
outputs = self.wrapper({self.input_name: imgs})
outputs = self.wrapper.output_to_list(outputs)
outputs = [out.detach().cpu().numpy() for out in outputs]
return outputs

def show_result(self,
Expand Down