mirrors
/
mmdeploy
mirror of https://github.com/open-mmlab/mmdeploy.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

[Enhancement] optimize delta2bboxes (#152) 

* optimize delta2bboxes

* ncnn update
pull/187/head
q.yao 2022-02-23 10:59:05 +08:00 committed by GitHub
parent 141d956636
commit 2c25eff32c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 53 additions and 85 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

138
mmdeploy/codebase/mmdet/core/bbox/delta_xywh_bbox_coder.py
View File

@ -92,48 +92,42 @@ def delta2bbox(ctx,
bboxes (Tensor): Boxes with shape (N, num_classes * 4) or (N, 4),
where 4 represent tl_x, tl_y, br_x, br_y.
"""
means = deltas.new_tensor(means).view(1,
-1).repeat(1,
deltas.size(-1) // 4)
stds = deltas.new_tensor(stds).view(1, -1).repeat(1, deltas.size(-1) // 4)
denorm_deltas = deltas * stds + means
dx = denorm_deltas[..., 0::4]
dy = denorm_deltas[..., 1::4]
dw = denorm_deltas[..., 2::4]
dh = denorm_deltas[..., 3::4]
means = deltas.new_tensor(means).view(1, -1)
stds = deltas.new_tensor(stds).view(1, -1)
delta_shape = deltas.shape
reshaped_deltas = deltas.view(delta_shape[:-1] + (-1, 4))
denorm_deltas = reshaped_deltas * stds + means
x1, y1 = rois[..., 0], rois[..., 1]
x2, y2 = rois[..., 2], rois[..., 3]
# Compute center of each roi
px = ((x1 + x2) * 0.5).unsqueeze(-1).expand_as(dx)
py = ((y1 + y2) * 0.5).unsqueeze(-1).expand_as(dy)
# Compute width/height of each roi
pw = (x2 - x1).unsqueeze(-1).expand_as(dw)
ph = (y2 - y1).unsqueeze(-1).expand_as(dh)
dxy = denorm_deltas[..., :2]
dwh = denorm_deltas[..., 2:]
dx_width = pw * dx
dy_height = ph * dy
xy1 = rois[..., None, :2]
xy2 = rois[..., None, 2:]
pxy = (xy1 + xy2) * 0.5
pwh = xy2 - xy1
dxy_wh = pwh * dxy
max_ratio = np.abs(np.log(wh_ratio_clip))
if add_ctr_clamp:
dx_width = torch.clamp(dx_width, max=ctr_clamp, min=-ctr_clamp)
dy_height = torch.clamp(dy_height, max=ctr_clamp, min=-ctr_clamp)
dw = torch.clamp(dw, max=max_ratio)
dh = torch.clamp(dh, max=max_ratio)
dxy_wh = torch.clamp(dxy_wh, max=ctr_clamp, min=-ctr_clamp)
dwh = torch.clamp(dwh, max=max_ratio)
else:
dw = dw.clamp(min=-max_ratio, max=max_ratio)
dh = dh.clamp(min=-max_ratio, max=max_ratio)
dwh = dwh.clamp(min=-max_ratio, max=max_ratio)
# Use exp(network energy) to enlarge/shrink each roi
gw = pw * dw.exp()
gh = ph * dh.exp()
half_gwh = pwh * dwh.exp() * 0.5
# Use network energy to shift the center of each roi
gx = px + dx_width
gy = py + dy_height
gxy = pxy + dxy_wh
# Convert center-xy/width/height to top-left, bottom-right
x1 = gx - gw * 0.5
y1 = gy - gh * 0.5
x2 = gx + gw * 0.5
y2 = gy + gh * 0.5
xy1 = gxy - half_gwh
xy2 = gxy + half_gwh
x1 = xy1[..., 0]
y1 = xy1[..., 1]
x2 = xy2[..., 0]
y2 = xy2[..., 1]
if clip_border and max_shape is not None:
from mmdeploy.codebase.mmdet.deploy import clip_bboxes
@ -190,68 +184,42 @@ def delta2bbox__ncnn(ctx,
or (N, num_classes * 4) or (N, 4), where 4 represent tl_x, tl_y,
br_x, br_y.
"""
means = deltas.new_tensor(means).view(1, 1,
-1).repeat(1, deltas.size(-2),
deltas.size(-1) // 4).data
stds = deltas.new_tensor(stds).view(1, 1,
-1).repeat(1, deltas.size(-2),
deltas.size(-1) // 4).data
denorm_deltas = deltas * stds + means
if denorm_deltas.shape[-1] == 4:
dx = denorm_deltas[..., 0:1]
dy = denorm_deltas[..., 1:2]
dw = denorm_deltas[..., 2:3]
dh = denorm_deltas[..., 3:4]
else:
dx = denorm_deltas[..., 0::4]
dy = denorm_deltas[..., 1::4]
dw = denorm_deltas[..., 2::4]
dh = denorm_deltas[..., 3::4]
means = deltas.new_tensor(means).view(1, 1, 1, -1).data
stds = deltas.new_tensor(stds).view(1, 1, 1, -1).data
delta_shape = deltas.shape
reshaped_deltas = deltas.view(delta_shape[:-1] + (-1, 4))
denorm_deltas = reshaped_deltas * stds + means
x1, y1 = rois[..., 0:1], rois[..., 1:2]
x2, y2 = rois[..., 2:3], rois[..., 3:4]
dxy = denorm_deltas[..., :2]
dwh = denorm_deltas[..., 2:]
# Compute center of each roi
px = (x1 + x2) * 0.5
py = (y1 + y2) * 0.5
# Compute width/height of each roi
pw = x2 - x1
ph = y2 - y1
xy1 = rois[..., None, :2]
xy2 = rois[..., None, 2:]
# do not use expand unless necessary
# since expand is a custom ops
if px.shape[-1] != 4:
px = px.expand_as(dx)
if py.shape[-1] != 4:
py = py.expand_as(dy)
if pw.shape[-1] != 4:
pw = pw.expand_as(dw)
if px.shape[-1] != 4:
ph = ph.expand_as(dh)
dx_width = pw * dx
dy_height = ph * dy
pxy = (xy1 + xy2) * 0.5
pwh = xy2 - xy1
dxy_wh = pwh * dxy
max_ratio = np.abs(np.log(wh_ratio_clip))
if add_ctr_clamp:
dx_width = torch.clamp(dx_width, max=ctr_clamp, min=-ctr_clamp)
dy_height = torch.clamp(dy_height, max=ctr_clamp, min=-ctr_clamp)
dw = torch.clamp(dw, max=max_ratio)
dh = torch.clamp(dh, max=max_ratio)
dxy_wh = torch.clamp(dxy_wh, max=ctr_clamp, min=-ctr_clamp)
dwh = torch.clamp(dwh, max=max_ratio)
else:
dw = dw.clamp(min=-max_ratio, max=max_ratio)
dh = dh.clamp(min=-max_ratio, max=max_ratio)
dwh = dwh.clamp(min=-max_ratio, max=max_ratio)
# Use exp(network energy) to enlarge/shrink each roi
gw = pw * dw.exp()
gh = ph * dh.exp()
half_gwh = pwh * dwh.exp() * 0.5
# Use network energy to shift the center of each roi
gx = px + dx_width
gy = py + dy_height
gxy = pxy + dxy_wh
# Convert center-xy/width/height to top-left, bottom-right
x1 = gx - gw * 0.5
y1 = gy - gh * 0.5
x2 = gx + gw * 0.5
y2 = gy + gh * 0.5
xy1 = gxy - half_gwh
xy2 = gxy + half_gwh
x1 = xy1[..., 0]
y1 = xy1[..., 1]
x2 = xy2[..., 0]
y2 = xy2[..., 1]
if clip_border and max_shape is not None:
from mmdeploy.codebase.mmdet.deploy import clip_bboxes