mirrors / paddlepaddle--paddleocr

项目文件夹

paddlepaddle--paddleocr
文件
main
T
wehub-resource-sync e904b667c6
Build/Publish Develop Docs / deploy (push) Failing after 1s
详情
PaddleOCR Code Style Check / check-code-style (push) Failing after 1s
详情
PaddleOCR PR Tests GPU / detect-changes (push) Failing after 1s
详情
PaddleOCR PR Tests / detect-changes (push) Failing after 1s
详情
PaddleOCR PR Tests GPU / test-pr-gpu (push) Has been cancelled
详情
PaddleOCR PR Tests / test-pr (push) Has been cancelled
详情
PaddleOCR PR Tests GPU / test-pr-gpu-impl (push) Has been cancelled
详情
PaddleOCR PR Tests / test-pr-python (3.13) (push) Has been cancelled
详情
PaddleOCR PR Tests / test-pr-python (3.8) (push) Has been cancelled
详情
PaddleOCR PR Tests / test-pr-python (3.9) (push) Has been cancelled
详情
chore: import upstream snapshot with attribution
2026-07-13 11:59:26 +08:00

1019 行
37 KiB
Python
原始文件 永久链接 Blame 文件历史

此文件含有模棱两可的 Unicode 字符
此文件含有可能会与其他字符混淆的 Unicode 字符。 如果您是想特意这样的，可以安全地忽略该警告。 使用 Escape 按钮显示他们。
# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
from paddle import nn
import paddle.nn.functional as F
from paddle import ParamAttr
import os
import sys
from ppocr.modeling.necks.intracl import IntraCLBlock
__dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.append(__dir__)
sys.path.insert(0, os.path.abspath(os.path.join(__dir__, "../../..")))
from ppocr.modeling.backbones.det_mobilenet_v3 import SEModule
class DSConv(nn.Layer):
def __init__(
self,
in_channels,
out_channels,
kernel_size,
padding,
stride=1,
groups=None,
if_act=True,
act="relu",
**kwargs,
):
super(DSConv, self).__init__()
if groups == None:
groups = in_channels
self.if_act = if_act
self.act = act
self.conv1 = nn.Conv2D(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias_attr=False,
)
self.bn1 = nn.BatchNorm(num_channels=in_channels, act=None)
self.conv2 = nn.Conv2D(
in_channels=in_channels,
out_channels=int(in_channels * 4),
kernel_size=1,
stride=1,
bias_attr=False,
)
self.bn2 = nn.BatchNorm(num_channels=int(in_channels * 4), act=None)
self.conv3 = nn.Conv2D(
in_channels=int(in_channels * 4),
out_channels=out_channels,
kernel_size=1,
stride=1,
bias_attr=False,
)
self._c = [in_channels, out_channels]
if in_channels != out_channels:
self.conv_end = nn.Conv2D(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
stride=1,
bias_attr=False,
)
def forward(self, inputs):
x = self.conv1(inputs)
x = self.bn1(x)
x = self.conv2(x)
x = self.bn2(x)
if self.if_act:
if self.act == "relu":
x = F.relu(x)
elif self.act == "hardswish":
x = F.hardswish(x)
else:
print(
"The activation function({}) is selected incorrectly.".format(
self.act
)
)
exit()
x = self.conv3(x)
if self._c[0] != self._c[1]:
x = x + self.conv_end(inputs)
return x
class DBFPN(nn.Layer):
def __init__(self, in_channels, out_channels, use_asf=False, **kwargs):
super(DBFPN, self).__init__()
self.out_channels = out_channels
self.use_asf = use_asf
weight_attr = paddle.nn.initializer.KaimingUniform()
self.in2_conv = nn.Conv2D(
in_channels=in_channels[0],
out_channels=self.out_channels,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.in3_conv = nn.Conv2D(
in_channels=in_channels[1],
out_channels=self.out_channels,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.in4_conv = nn.Conv2D(
in_channels=in_channels[2],
out_channels=self.out_channels,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.in5_conv = nn.Conv2D(
in_channels=in_channels[3],
out_channels=self.out_channels,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.p5_conv = nn.Conv2D(
in_channels=self.out_channels,
out_channels=self.out_channels // 4,
kernel_size=3,
padding=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.p4_conv = nn.Conv2D(
in_channels=self.out_channels,
out_channels=self.out_channels // 4,
kernel_size=3,
padding=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.p3_conv = nn.Conv2D(
in_channels=self.out_channels,
out_channels=self.out_channels // 4,
kernel_size=3,
padding=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.p2_conv = nn.Conv2D(
in_channels=self.out_channels,
out_channels=self.out_channels // 4,
kernel_size=3,
padding=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
if self.use_asf is True:
self.asf = ASFBlock(self.out_channels, self.out_channels // 4)
def forward(self, x):
c2, c3, c4, c5 = x
in5 = self.in5_conv(c5)
in4 = self.in4_conv(c4)
in3 = self.in3_conv(c3)
in2 = self.in2_conv(c2)
out4 = in4 + F.upsample(
in5, scale_factor=2, mode="nearest", align_mode=1
) # 1/16
out3 = in3 + F.upsample(
out4, scale_factor=2, mode="nearest", align_mode=1
) # 1/8
out2 = in2 + F.upsample(
out3, scale_factor=2, mode="nearest", align_mode=1
) # 1/4
p5 = self.p5_conv(in5)
p4 = self.p4_conv(out4)
p3 = self.p3_conv(out3)
p2 = self.p2_conv(out2)
p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
fuse = paddle.concat([p5, p4, p3, p2], axis=1)
if self.use_asf is True:
fuse = self.asf(fuse, [p5, p4, p3, p2])
return fuse
class RSELayer(nn.Layer):
def __init__(self, in_channels, out_channels, kernel_size, shortcut=True):
super(RSELayer, self).__init__()
weight_attr = paddle.nn.initializer.KaimingUniform()
self.out_channels = out_channels
self.in_conv = nn.Conv2D(
in_channels=in_channels,
out_channels=self.out_channels,
kernel_size=kernel_size,
padding=int(kernel_size // 2),
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.se_block = SEModule(self.out_channels)
self.shortcut = shortcut
def forward(self, ins):
x = self.in_conv(ins)
if self.shortcut:
out = x + self.se_block(x)
else:
out = self.se_block(x)
return out
class RSEFPN(nn.Layer):
def __init__(self, in_channels, out_channels, shortcut=True, **kwargs):
super(RSEFPN, self).__init__()
self.out_channels = out_channels
self.ins_conv = nn.LayerList()
self.inp_conv = nn.LayerList()
self.intracl = False
if "intracl" in kwargs.keys() and kwargs["intracl"] is True:
self.intracl = kwargs["intracl"]
self.incl1 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl2 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl3 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl4 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
for i in range(len(in_channels)):
self.ins_conv.append(
RSELayer(in_channels[i], out_channels, kernel_size=1, shortcut=shortcut)
)
self.inp_conv.append(
RSELayer(
out_channels, out_channels // 4, kernel_size=3, shortcut=shortcut
)
)
def forward(self, x):
c2, c3, c4, c5 = x
in5 = self.ins_conv[3](c5)
in4 = self.ins_conv[2](c4)
in3 = self.ins_conv[1](c3)
in2 = self.ins_conv[0](c2)
out4 = in4 + F.upsample(
in5, scale_factor=2, mode="nearest", align_mode=1
) # 1/16
out3 = in3 + F.upsample(
out4, scale_factor=2, mode="nearest", align_mode=1
) # 1/8
out2 = in2 + F.upsample(
out3, scale_factor=2, mode="nearest", align_mode=1
) # 1/4
p5 = self.inp_conv[3](in5)
p4 = self.inp_conv[2](out4)
p3 = self.inp_conv[1](out3)
p2 = self.inp_conv[0](out2)
if self.intracl is True:
p5 = self.incl4(p5)
p4 = self.incl3(p4)
p3 = self.incl2(p3)
p2 = self.incl1(p2)
p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
fuse = paddle.concat([p5, p4, p3, p2], axis=1)
return fuse
class RepLKFPN(nn.Layer):
"""Optimized RSEFPN: replaces 3x3 standard Conv in inp_conv with
DilatedReparamBlock (DW, 5x5) + PWConv 1x1 + SE.
Uses the existing DilatedReparamBlock from UniRepLKNet to provide
multi-branch dilated training with single-conv inference.
Changes vs RSEFPN:
- inp_conv: RSELayer(3x3 std Conv + SE)
→ DilatedReparamBlock(5x5 DW) + PWConv(1x1) + SE
- ins_conv: unchanged (1x1 Conv, no benefit from DW decomposition)
Parameter comparison (out_channels=96, 4 levels):
RSEFPN inp_conv: 4 × (96×24×9 + SE) = 4 × 21,054 = 84,216
RepLKFPN inp_conv: 4 × (DilReparam96 + 96×24 + SE)
= 4 × (96×25 + 96×2 + 2×(96×9+96×2) + 96×24 + SE)
= 4 × (2,400 + 192 + 2×(864+192) + 2,304 + 318)
= 4 × 7,326 = 29,304
inp_conv reduction: 65.2%
Receptive field: 3×3 → 5×5 (with multi-dilation 3,5 patterns in training)
Inference: DilatedReparamBlock merges to single 5x5 DWConv, zero extra cost.
"""
def __init__(
self, in_channels, out_channels, shortcut=True, dilated_kernel_size=7, **kwargs
):
super(RepLKFPN, self).__init__()
self.out_channels = out_channels
self.is_repped = False
self.ins_conv = nn.LayerList()
self.inp_conv_dw = nn.LayerList()
self.inp_conv_pw = nn.LayerList()
self.inp_conv_se = nn.LayerList()
self.shortcut = shortcut
self.intracl = False
if "intracl" in kwargs.keys() and kwargs["intracl"] is True:
self.intracl = kwargs["intracl"]
self.incl1 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl2 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl3 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl4 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
weight_attr = paddle.nn.initializer.KaimingUniform()
for i in range(len(in_channels)):
self.ins_conv.append(
RSELayer(in_channels[i], out_channels, kernel_size=1, shortcut=shortcut)
)
self.inp_conv_dw.append(
DilatedReparamBlock(
channels=out_channels, kernel_size=dilated_kernel_size
)
)
self.inp_conv_pw.append(
nn.Conv2D(
in_channels=out_channels,
out_channels=out_channels // 4,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
)
self.inp_conv_se.append(SEModule(out_channels // 4))
def _inp_forward(self, x, idx):
x = self.inp_conv_dw[idx](x)
x = self.inp_conv_pw[idx](x)
if self.shortcut:
x = x + self.inp_conv_se[idx](x)
else:
x = self.inp_conv_se[idx](x)
return x
def forward(self, x):
c2, c3, c4, c5 = x
in5 = self.ins_conv[3](c5)
in4 = self.ins_conv[2](c4)
in3 = self.ins_conv[1](c3)
in2 = self.ins_conv[0](c2)
out4 = in4 + F.upsample(
in5, scale_factor=2, mode="nearest", align_mode=1
) # 1/16
out3 = in3 + F.upsample(
out4, scale_factor=2, mode="nearest", align_mode=1
) # 1/8
out2 = in2 + F.upsample(
out3, scale_factor=2, mode="nearest", align_mode=1
) # 1/4
p5 = self._inp_forward(in5, 3)
p4 = self._inp_forward(out4, 2)
p3 = self._inp_forward(out3, 1)
p2 = self._inp_forward(out2, 0)
if self.intracl is True:
p5 = self.incl4(p5)
p4 = self.incl3(p4)
p3 = self.incl2(p3)
p2 = self.incl1(p2)
p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
fuse = paddle.concat([p5, p4, p3, p2], axis=1)
if self.training:
return {"fuse": fuse, "aux_p4": out4, "aux_p3": out3, "aux_p2": out2}
return fuse
def rep(self):
"""Merge DilatedReparamBlock branches for inference deployment."""
if self.is_repped:
return
for i in range(len(self.inp_conv_dw)):
self.inp_conv_dw[i].rep()
self.is_repped = True
class LKPAN(nn.Layer):
def __init__(self, in_channels, out_channels, mode="large", **kwargs):
super(LKPAN, self).__init__()
self.out_channels = out_channels
weight_attr = paddle.nn.initializer.KaimingUniform()
self.ins_conv = nn.LayerList()
self.inp_conv = nn.LayerList()
# pan head
self.pan_head_conv = nn.LayerList()
self.pan_lat_conv = nn.LayerList()
if mode.lower() == "lite":
p_layer = DSConv
elif mode.lower() == "large":
p_layer = nn.Conv2D
else:
raise ValueError(
"mode can only be one of ['lite', 'large'], but received {}".format(
mode
)
)
for i in range(len(in_channels)):
self.ins_conv.append(
nn.Conv2D(
in_channels=in_channels[i],
out_channels=self.out_channels,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
)
self.inp_conv.append(
p_layer(
in_channels=self.out_channels,
out_channels=self.out_channels // 4,
kernel_size=9,
padding=4,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
)
if i > 0:
self.pan_head_conv.append(
nn.Conv2D(
in_channels=self.out_channels // 4,
out_channels=self.out_channels // 4,
kernel_size=3,
padding=1,
stride=2,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
)
self.pan_lat_conv.append(
p_layer(
in_channels=self.out_channels // 4,
out_channels=self.out_channels // 4,
kernel_size=9,
padding=4,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
)
self.intracl = False
if "intracl" in kwargs.keys() and kwargs["intracl"] is True:
self.intracl = kwargs["intracl"]
self.incl1 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl2 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl3 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl4 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
def forward(self, x):
c2, c3, c4, c5 = x
in5 = self.ins_conv[3](c5)
in4 = self.ins_conv[2](c4)
in3 = self.ins_conv[1](c3)
in2 = self.ins_conv[0](c2)
out4 = in4 + F.upsample(
in5, scale_factor=2, mode="nearest", align_mode=1
) # 1/16
out3 = in3 + F.upsample(
out4, scale_factor=2, mode="nearest", align_mode=1
) # 1/8
out2 = in2 + F.upsample(
out3, scale_factor=2, mode="nearest", align_mode=1
) # 1/4
f5 = self.inp_conv[3](in5)
f4 = self.inp_conv[2](out4)
f3 = self.inp_conv[1](out3)
f2 = self.inp_conv[0](out2)
pan3 = f3 + self.pan_head_conv[0](f2)
pan4 = f4 + self.pan_head_conv[1](pan3)
pan5 = f5 + self.pan_head_conv[2](pan4)
p2 = self.pan_lat_conv[0](f2)
p3 = self.pan_lat_conv[1](pan3)
p4 = self.pan_lat_conv[2](pan4)
p5 = self.pan_lat_conv[3](pan5)
if self.intracl is True:
p5 = self.incl4(p5)
p4 = self.incl3(p4)
p3 = self.incl2(p3)
p2 = self.incl1(p2)
p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
fuse = paddle.concat([p5, p4, p3, p2], axis=1)
return fuse
class DilatedReparamBlock(nn.Layer):
"""
Dilated Reparam Block from UniRepLKNet.
Reference: https://github.com/AILab-CVC/UniRepLKNet
Training: uses multiple parallel dilated depthwise convolutions + BN
Inference: all branches merge into a single large-kernel depthwise conv
For kernel_size=9, the branches are:
- origin: 9x9 DW Conv (dil=1)
- branch1: 5x5 DW Conv (dil=1, equiv RF=5)
- branch2: 5x5 DW Conv (dil=2, equiv RF=9)
- branch3: 3x3 DW Conv (dil=3, equiv RF=7)
- branch4: 3x3 DW Conv (dil=4, equiv RF=9)
"""
def __init__(self, channels, kernel_size=9, deploy=False):
super(DilatedReparamBlock, self).__init__()
self.channels = channels
self.kernel_size = kernel_size
self.is_repped = deploy
if kernel_size == 9:
self.kernel_sizes = [5, 5, 3, 3]
self.dilates = [1, 2, 3, 4]
elif kernel_size == 7:
self.kernel_sizes = [5, 3, 3]
self.dilates = [1, 2, 3]
elif kernel_size == 5:
self.kernel_sizes = [3, 3]
self.dilates = [1, 2]
elif kernel_size == 11:
self.kernel_sizes = [5, 5, 3, 3, 3]
self.dilates = [1, 2, 3, 4, 5]
elif kernel_size == 13:
self.kernel_sizes = [5, 7, 3, 3, 3]
self.dilates = [1, 2, 3, 4, 5]
else:
raise ValueError(
"DilatedReparamBlock requires kernel_size in [5,7,9,11,13], "
"but got {}".format(kernel_size)
)
if not self.is_repped:
self.lk_origin = nn.Conv2D(
in_channels=channels,
out_channels=channels,
kernel_size=kernel_size,
stride=1,
padding=kernel_size // 2,
groups=channels,
bias_attr=False,
)
self.origin_bn = nn.BatchNorm2D(channels)
for k, r in zip(self.kernel_sizes, self.dilates):
equiv_ks = r * (k - 1) + 1
p = equiv_ks // 2
conv = nn.Conv2D(
in_channels=channels,
out_channels=channels,
kernel_size=k,
stride=1,
padding=p,
dilation=r,
groups=channels,
bias_attr=False,
)
bn = nn.BatchNorm2D(channels)
setattr(self, "dil_conv_k{}_{}".format(k, r), conv)
setattr(self, "dil_bn_k{}_{}".format(k, r), bn)
else:
self.lk_origin = nn.Conv2D(
in_channels=channels,
out_channels=channels,
kernel_size=kernel_size,
stride=1,
padding=kernel_size // 2,
groups=channels,
bias_attr=True,
)
def forward(self, x):
if self.is_repped:
return self.lk_origin(x)
out = self.origin_bn(self.lk_origin(x))
for k, r in zip(self.kernel_sizes, self.dilates):
conv = getattr(self, "dil_conv_k{}_{}".format(k, r))
bn = getattr(self, "dil_bn_k{}_{}".format(k, r))
out = out + bn(conv(x))
return out
@staticmethod
def _fuse_bn(conv, bn):
"""Fuse Conv2D + BatchNorm2D into a single Conv2D (weight, bias)."""
kernel = conv.weight
gamma = bn.weight
beta = bn.bias
running_mean = bn._mean
running_var = bn._variance
eps = bn._epsilon
std = paddle.sqrt(running_var + eps)
fused_weight = kernel * (gamma / std).reshape([-1, 1, 1, 1])
fused_bias = beta - running_mean * gamma / std
return fused_weight, fused_bias
@staticmethod
def _convert_dilated_to_nondilated(kernel, dilate_rate):
"""Convert dilated conv kernel to equivalent non-dilated (sparse) kernel
by inserting zeros between kernel elements using transposed convolution."""
if dilate_rate == 1:
return kernel
identity = paddle.ones(shape=[1, 1, 1, 1], dtype=kernel.dtype)
# F.conv2d_transpose with stride=dilate_rate inserts zeros
# Process each channel independently
C = kernel.shape[0]
result_list = []
for i in range(C):
k_i = kernel[i : i + 1] # (1, 1, kH, kW)
dilated = F.conv2d_transpose(k_i, identity, stride=dilate_rate)
result_list.append(dilated)
return paddle.concat(result_list, axis=0)
@staticmethod
def _merge_dilated_into_large_kernel(large_kernel, dilated_kernel, dilated_r):
"""Pad dilated equivalent kernel to large kernel size and add."""
large_k = large_kernel.shape[2]
dilated_k = dilated_kernel.shape[2]
equiv_ks = dilated_r * (dilated_k - 1) + 1
equiv_kernel = DilatedReparamBlock._convert_dilated_to_nondilated(
dilated_kernel, dilated_r
)
rows_to_pad = large_k // 2 - equiv_ks // 2
if rows_to_pad > 0:
merged = large_kernel + F.pad(
equiv_kernel, [rows_to_pad, rows_to_pad, rows_to_pad, rows_to_pad]
)
else:
merged = large_kernel + equiv_kernel
return merged
@paddle.no_grad()
def rep(self):
"""Merge all parallel branches into a single large-kernel DW conv.
Call this before switching to deploy/inference mode."""
if self.is_repped:
return
origin_k, origin_b = self._fuse_bn(self.lk_origin, self.origin_bn)
for k, r in zip(self.kernel_sizes, self.dilates):
conv = getattr(self, "dil_conv_k{}_{}".format(k, r))
bn = getattr(self, "dil_bn_k{}_{}".format(k, r))
branch_k, branch_b = self._fuse_bn(conv, bn)
origin_k = self._merge_dilated_into_large_kernel(origin_k, branch_k, r)
origin_b = origin_b + branch_b
merged_conv = nn.Conv2D(
in_channels=self.channels,
out_channels=self.channels,
kernel_size=self.kernel_size,
stride=1,
padding=self.kernel_size // 2,
groups=self.channels,
bias_attr=True,
)
merged_conv.weight.set_value(origin_k)
merged_conv.bias.set_value(origin_b)
self.lk_origin = merged_conv
self.is_repped = True
delattr(self, "origin_bn")
for k, r in zip(self.kernel_sizes, self.dilates):
delattr(self, "dil_conv_k{}_{}".format(k, r))
delattr(self, "dil_bn_k{}_{}".format(k, r))
class DilatedReparamConv(nn.Layer):
"""
A drop-in replacement for standard Conv2D (in_ch → out_ch, large kernel)
using DilatedReparamBlock (depthwise) + 1x1 pointwise convolution.
Architecture:
input(in_ch) → DilatedReparamBlock(in_ch, DW, kernel_size) → 1x1 Conv(in_ch→out_ch) → BN
This decomposition replaces a single large standard conv with DW + PW,
drastically reducing parameters while maintaining the large receptive field.
"""
def __init__(
self, in_channels, out_channels, kernel_size=9, deploy=False, **kwargs
):
super(DilatedReparamConv, self).__init__()
self.is_repped = False
weight_attr = paddle.nn.initializer.KaimingUniform()
self.dw = DilatedReparamBlock(
channels=in_channels, kernel_size=kernel_size, deploy=deploy
)
self.pw = nn.Conv2D(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
self.bn = nn.BatchNorm2D(out_channels)
def forward(self, x):
x = self.dw(x)
x = self.pw(x)
if not self.is_repped:
x = self.bn(x)
return x
@paddle.no_grad()
def rep(self):
"""Fuse DW branches + PW Conv + BN for deployment."""
if self.is_repped:
return
self.dw.rep()
# Fuse pw(Conv2D, no bias) + bn(BatchNorm2D) into single Conv2D with bias
conv, bn = self.pw, self.bn
gamma = bn.weight
std = paddle.sqrt(bn._variance + bn._epsilon)
scale = gamma / std
w = conv.weight * scale[:, None, None, None]
b = bn.bias - bn._mean * scale
fused = nn.Conv2D(
conv._in_channels,
conv._out_channels,
conv._kernel_size,
stride=conv._stride,
padding=conv._padding,
dilation=conv._dilation,
groups=conv._groups,
)
fused.weight.set_value(w)
fused.bias.set_value(b)
self.pw = fused
del self.bn
self.is_repped = True
class RepLKPAN(nn.Layer):
"""
Optimized LKPAN using UniRepLKNet's DilatedReparamBlock.
Replaces the 8 standard 9x9 Conv2D in LKPAN (4 inp_conv + 4 pan_lat_conv)
with DilatedReparamConv (DW large-kernel reparam + 1x1 pointwise).
Parameter comparison (out_channels=256, i.e. inner_ch=64):
┌──────────────────────────────────────────────────────────────────────────┐
│ Component │ Original (Std 9x9 Conv) │ UniRepLK (DW+PW) │
├──────────────────────────────────────────────────────────────────────────┤
│ inp_conv[i]: │ 256×64×9×9 = 1,327,104 │ DW: 256×1×9×9 = 20,736 │
│ (256→64, ×4) │ │ +4 dilated DW branches │
│ │ │ ≈ 256×(25+25+9+9) │
│ │ │ = 17,408 (training) │
│ │ │ PW: 256×64×1×1 = 16,384 │
│ │ │ BN: 64×2 = 128 │
│ │ │ Subtotal/layer ≈ 54,656 │
│ │ │ (vs 1,327,104 original) │
├──────────────────────────────────────────────────────────────────────────┤
│ pan_lat_conv[i]: │ 64×64×9×9 = 331,776 │ DW: 64×1×9×9 = 5,184 │
│ (64→64, ×4) │ │ +4 dilated DW branches │
│ │ │ ≈ 64×(25+25+9+9) │
│ │ │ = 4,352 (training) │
│ │ │ PW: 64×64×1×1 = 4,096 │
│ │ │ BN: 64×2 = 128 │
│ │ │ Subtotal/layer ≈ 13,760 │
│ │ │ (vs 331,776 original) │
├──────────────────────────────────────────────────────────────────────────┤
│ Total 9x9 params │ 4×1,327,104 + 4×331,776 │ 4×54,656 + 4×13,760 │
│ (8 layers) │ = 6,635,520 │ = 273,664 │
│ │ │ 95.9% reduction │
├──────────────────────────────────────────────────────────────────────────┤
│ Inference(reparam)│ Same as above │ DW 9x9 merged (no extra │
│ │ │ branch overhead)+PW 1x1 │
│ │ │ FLOPs also greatly reduced│
└──────────────────────────────────────────────────────────────────────────┘
Note: BN params in dilated branches are only present during training and
merged into the DW conv weights at inference. The training param count
includes these BN params; inference param count is slightly lower.
"""
def __init__(self, in_channels, out_channels, mode="large", **kwargs):
super(RepLKPAN, self).__init__()
self.out_channels = out_channels
self.is_repped = False
weight_attr = paddle.nn.initializer.KaimingUniform()
self.ins_conv = nn.LayerList()
self.inp_conv = nn.LayerList()
# pan head
self.pan_head_conv = nn.LayerList()
self.pan_lat_conv = nn.LayerList()
for i in range(len(in_channels)):
self.ins_conv.append(
nn.Conv2D(
in_channels=in_channels[i],
out_channels=self.out_channels,
kernel_size=1,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
)
self.inp_conv.append(
DilatedReparamConv(
in_channels=self.out_channels,
out_channels=self.out_channels // 4,
kernel_size=9,
)
)
if i > 0:
self.pan_head_conv.append(
nn.Conv2D(
in_channels=self.out_channels // 4,
out_channels=self.out_channels // 4,
kernel_size=3,
padding=1,
stride=2,
weight_attr=ParamAttr(initializer=weight_attr),
bias_attr=False,
)
)
self.pan_lat_conv.append(
DilatedReparamConv(
in_channels=self.out_channels // 4,
out_channels=self.out_channels // 4,
kernel_size=9,
)
)
self.intracl = False
if "intracl" in kwargs.keys() and kwargs["intracl"] is True:
self.intracl = kwargs["intracl"]
self.incl1 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl2 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl3 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
self.incl4 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
def forward(self, x):
c2, c3, c4, c5 = x
in5 = self.ins_conv[3](c5)
in4 = self.ins_conv[2](c4)
in3 = self.ins_conv[1](c3)
in2 = self.ins_conv[0](c2)
out4 = in4 + F.upsample(
in5, scale_factor=2, mode="nearest", align_mode=1
) # 1/16
out3 = in3 + F.upsample(
out4, scale_factor=2, mode="nearest", align_mode=1
) # 1/8
out2 = in2 + F.upsample(
out3, scale_factor=2, mode="nearest", align_mode=1
) # 1/4
f5 = self.inp_conv[3](in5)
f4 = self.inp_conv[2](out4)
f3 = self.inp_conv[1](out3)
f2 = self.inp_conv[0](out2)
pan3 = f3 + self.pan_head_conv[0](f2)
pan4 = f4 + self.pan_head_conv[1](pan3)
pan5 = f5 + self.pan_head_conv[2](pan4)
p2 = self.pan_lat_conv[0](f2)
p3 = self.pan_lat_conv[1](pan3)
p4 = self.pan_lat_conv[2](pan4)
p5 = self.pan_lat_conv[3](pan5)
if self.intracl is True:
p5 = self.incl4(p5)
p4 = self.incl3(p4)
p3 = self.incl2(p3)
p2 = self.incl1(p2)
p5 = F.upsample(p5, scale_factor=8, mode="nearest", align_mode=1)
p4 = F.upsample(p4, scale_factor=4, mode="nearest", align_mode=1)
p3 = F.upsample(p3, scale_factor=2, mode="nearest", align_mode=1)
fuse = paddle.concat([p5, p4, p3, p2], axis=1)
if self.training:
return {"fuse": fuse, "aux_p4": out4, "aux_p3": out3, "aux_p2": out2}
return fuse
def rep(self):
"""Merge all DilatedReparamBlock branches and fuse PW+BN for deployment."""
if self.is_repped:
return
for i in range(len(self.inp_conv)):
self.inp_conv[i].rep()
for i in range(len(self.pan_lat_conv)):
self.pan_lat_conv[i].rep()
self.is_repped = True
class ASFBlock(nn.Layer):
"""
This code is referred from:
https://github.com/MhLiao/DB/blob/master/decoders/feature_attention.py
"""
def __init__(self, in_channels, inter_channels, out_features_num=4):
"""
Adaptive Scale Fusion (ASF) block of DBNet++
Args:
in_channels: the number of channels in the input data
inter_channels: the number of middle channels
out_features_num: the number of fused stages
"""
super(ASFBlock, self).__init__()
weight_attr = paddle.nn.initializer.KaimingUniform()
self.in_channels = in_channels
self.inter_channels = inter_channels
self.out_features_num = out_features_num
self.conv = nn.Conv2D(in_channels, inter_channels, 3, padding=1)
self.spatial_scale = nn.Sequential(
# Nx1xHxW
nn.Conv2D(
in_channels=1,
out_channels=1,
kernel_size=3,
bias_attr=False,
padding=1,
weight_attr=ParamAttr(initializer=weight_attr),
),
nn.ReLU(),
nn.Conv2D(
in_channels=1,
out_channels=1,
kernel_size=1,
bias_attr=False,
weight_attr=ParamAttr(initializer=weight_attr),
),
nn.Sigmoid(),
)
self.channel_scale = nn.Sequential(
nn.Conv2D(
in_channels=inter_channels,
out_channels=out_features_num,
kernel_size=1,
bias_attr=False,
weight_attr=ParamAttr(initializer=weight_attr),
),
nn.Sigmoid(),
)
def forward(self, fuse_features, features_list):
fuse_features = self.conv(fuse_features)
spatial_x = paddle.mean(fuse_features, axis=1, keepdim=True)
attention_scores = self.spatial_scale(spatial_x) + fuse_features
attention_scores = self.channel_scale(attention_scores)
assert len(features_list) == self.out_features_num
out_list = []
for i in range(self.out_features_num):
out_list.append(attention_scores[:, i : i + 1] * features_list[i])
return paddle.concat(out_list, axis=1)
在新工单中引用 查看 Git Blame 复制永久链接