Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 1 年 前同步成功

通知 695
Star 11112
Fork 2696
P
PaddleDetection
未验证 提交 5b20aa32 编写于 作者: F Feng Ni 提交者： GitHub
浏览文件
操作

remove name and norm_name in head and fpn (#2649) 

* fix name of ConvNormLayer for fcos solov2

* fix name in other head convbn

* fix yolo name, fix ttf_head name, test=document_fix
上级 5aafa1b6
隐藏空白更改
内联 并排
Showing with 85 addition and 177 deletion
ppdet/modeling/backbones/darknet.py
浏览文件 @ 5b20aa32
...@@ -35,8 +35,8 @@ class ConvBNLayer(nn.Layer): ...@@ -35,8 +35,8 @@ class ConvBNLayer(nn.Layer):
norm_type='bn', norm_type='bn',
norm_decay=0., norm_decay=0.,
act="leaky", act="leaky",
name=None, data_format='NCHW',
data_format='NCHW'): name=''):
""" """
conv + bn + activation layer conv + bn + activation layer
...@@ -50,7 +50,6 @@ class ConvBNLayer(nn.Layer): ...@@ -50,7 +50,6 @@ class ConvBNLayer(nn.Layer):
norm_type (str): batch norm type, default bn norm_type (str): batch norm type, default bn
norm_decay (str): decay for weight and bias of batch norm layer, default 0. norm_decay (str): decay for weight and bias of batch norm layer, default 0.
act (str): activation function type, default 'leaky', which means leaky_relu act (str): activation function type, default 'leaky', which means leaky_relu
name (str): layer name
data_format (str): data format, NCHW or NHWC data_format (str): data format, NCHW or NHWC
""" """
super(ConvBNLayer, self).__init__() super(ConvBNLayer, self).__init__()
...@@ -68,7 +67,6 @@ class ConvBNLayer(nn.Layer): ...@@ -68,7 +67,6 @@ class ConvBNLayer(nn.Layer):
ch_out, ch_out,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
name=name,
data_format=data_format) data_format=data_format)
self.act = act self.act = act
...@@ -91,7 +89,6 @@ class DownSample(nn.Layer): ...@@ -91,7 +89,6 @@ class DownSample(nn.Layer):
padding=1, padding=1,
norm_type='bn', norm_type='bn',
norm_decay=0., norm_decay=0.,
name=None,
data_format='NCHW'): data_format='NCHW'):
""" """
downsample layer downsample layer
...@@ -104,7 +101,6 @@ class DownSample(nn.Layer): ...@@ -104,7 +101,6 @@ class DownSample(nn.Layer):
padding (int): padding size, default 1 padding (int): padding size, default 1
norm_type (str): batch norm type, default bn norm_type (str): batch norm type, default bn
norm_decay (str): decay for weight and bias of batch norm layer, default 0. norm_decay (str): decay for weight and bias of batch norm layer, default 0.
name (str): layer name
data_format (str): data format, NCHW or NHWC data_format (str): data format, NCHW or NHWC
""" """
...@@ -118,8 +114,7 @@ class DownSample(nn.Layer): ...@@ -118,8 +114,7 @@ class DownSample(nn.Layer):
padding=padding, padding=padding,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format)
name=name)
self.ch_out = ch_out self.ch_out = ch_out
def forward(self, inputs): def forward(self, inputs):
...@@ -133,7 +128,6 @@ class BasicBlock(nn.Layer): ...@@ -133,7 +128,6 @@ class BasicBlock(nn.Layer):
ch_out, ch_out,
norm_type='bn', norm_type='bn',
norm_decay=0., norm_decay=0.,
name=None,
data_format='NCHW'): data_format='NCHW'):
""" """
BasicBlock layer of DarkNet BasicBlock layer of DarkNet
...@@ -143,7 +137,6 @@ class BasicBlock(nn.Layer): ...@@ -143,7 +137,6 @@ class BasicBlock(nn.Layer):
ch_out (int): output channel ch_out (int): output channel
norm_type (str): batch norm type, default bn norm_type (str): batch norm type, default bn
norm_decay (str): decay for weight and bias of batch norm layer, default 0. norm_decay (str): decay for weight and bias of batch norm layer, default 0.
name (str): layer name
data_format (str): data format, NCHW or NHWC data_format (str): data format, NCHW or NHWC
""" """
...@@ -157,8 +150,7 @@ class BasicBlock(nn.Layer): ...@@ -157,8 +150,7 @@ class BasicBlock(nn.Layer):
padding=0, padding=0,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format)
name=name + '.0')
self.conv2 = ConvBNLayer( self.conv2 = ConvBNLayer(
ch_in=ch_out, ch_in=ch_out,
ch_out=ch_out * 2, ch_out=ch_out * 2,
...@@ -167,8 +159,7 @@ class BasicBlock(nn.Layer): ...@@ -167,8 +159,7 @@ class BasicBlock(nn.Layer):
padding=1, padding=1,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format)
name=name + '.1')
def forward(self, inputs): def forward(self, inputs):
conv1 = self.conv1(inputs) conv1 = self.conv1(inputs)
...@@ -205,8 +196,7 @@ class Blocks(nn.Layer): ...@@ -205,8 +196,7 @@ class Blocks(nn.Layer):
ch_out, ch_out,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format)
name=name + '.0')
self.res_out_list = [] self.res_out_list = []
for i in range(1, count): for i in range(1, count):
block_name = '{}.{}'.format(name, i) block_name = '{}.{}'.format(name, i)
...@@ -217,8 +207,7 @@ class Blocks(nn.Layer): ...@@ -217,8 +207,7 @@ class Blocks(nn.Layer):
ch_out, ch_out,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format))
name=block_name))
self.res_out_list.append(res_out) self.res_out_list.append(res_out)
self.ch_out = ch_out self.ch_out = ch_out
...@@ -272,16 +261,14 @@ class DarkNet(nn.Layer): ...@@ -272,16 +261,14 @@ class DarkNet(nn.Layer):
padding=1, padding=1,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format)
name='yolo_input')
self.downsample0 = DownSample( self.downsample0 = DownSample(
ch_in=32, ch_in=32,
ch_out=32 * 2, ch_out=32 * 2,
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format)
name='yolo_input.downsample')
self._out_channels = [] self._out_channels = []
self.darknet_conv_block_list = [] self.darknet_conv_block_list = []
...@@ -311,8 +298,7 @@ class DarkNet(nn.Layer): ...@@ -311,8 +298,7 @@ class DarkNet(nn.Layer):
ch_out=32 * (2**(i + 2)), ch_out=32 * (2**(i + 2)),
norm_type=norm_type, norm_type=norm_type,
norm_decay=norm_decay, norm_decay=norm_decay,
data_format=data_format, data_format=data_format))
name=down_name))
self.downsample_list.append(downsample) self.downsample_list.append(downsample)
def forward(self, inputs): def forward(self, inputs):
......
ppdet/modeling/heads/bbox_head.py
浏览文件 @ 5b20aa32
...@@ -128,10 +128,8 @@ class XConvNormHead(nn.Layer): ...@@ -128,10 +128,8 @@ class XConvNormHead(nn.Layer):
filter_size=3, filter_size=3,
stride=1, stride=1,
norm_type=self.norm_type, norm_type=self.norm_type,
norm_name=head_conv_name + '_norm',
freeze_norm=self.freeze_norm, freeze_norm=self.freeze_norm,
initializer=initializer, initializer=initializer))
name=head_conv_name))
self.bbox_head_convs.append(head_conv) self.bbox_head_convs.append(head_conv)
fan = conv_dim * resolution * resolution fan = conv_dim * resolution * resolution
......
ppdet/modeling/heads/fcos_head.py
浏览文件 @ 5b20aa32
...@@ -81,10 +81,8 @@ class FCOSFeat(nn.Layer): ...@@ -81,10 +81,8 @@ class FCOSFeat(nn.Layer):
stride=1, stride=1,
norm_type=norm_type, norm_type=norm_type,
use_dcn=use_dcn, use_dcn=use_dcn,
norm_name=cls_conv_name + '_norm',
bias_on=True, bias_on=True,
lr_scale=2., lr_scale=2.))
name=cls_conv_name))
self.cls_subnet_convs.append(cls_conv) self.cls_subnet_convs.append(cls_conv)
reg_conv_name = 'fcos_head_reg_tower_conv_{}'.format(i) reg_conv_name = 'fcos_head_reg_tower_conv_{}'.format(i)
...@@ -97,10 +95,8 @@ class FCOSFeat(nn.Layer): ...@@ -97,10 +95,8 @@ class FCOSFeat(nn.Layer):
stride=1, stride=1,
norm_type=norm_type, norm_type=norm_type,
use_dcn=use_dcn, use_dcn=use_dcn,
norm_name=reg_conv_name + '_norm',
bias_on=True, bias_on=True,
lr_scale=2., lr_scale=2.))
name=reg_conv_name))
self.reg_subnet_convs.append(reg_conv) self.reg_subnet_convs.append(reg_conv)
def forward(self, fpn_feat): def forward(self, fpn_feat):
......
ppdet/modeling/heads/mask_head.py
浏览文件 @ 5b20aa32
...@@ -63,10 +63,8 @@ class MaskFeat(nn.Layer): ...@@ -63,10 +63,8 @@ class MaskFeat(nn.Layer):
filter_size=3, filter_size=3,
stride=1, stride=1,
norm_type=self.norm_type, norm_type=self.norm_type,
norm_name=conv_name + '_norm',
initializer=KaimingNormal(fan_in=fan_conv), initializer=KaimingNormal(fan_in=fan_conv),
skip_quant=True, skip_quant=True))
name=conv_name))
mask_conv.add_sublayer(conv_name + 'act', nn.ReLU()) mask_conv.add_sublayer(conv_name + 'act', nn.ReLU())
else: else:
for i in range(self.num_convs): for i in range(self.num_convs):
......
ppdet/modeling/heads/solov2_head.py
浏览文件 @ 5b20aa32
...@@ -75,9 +75,7 @@ class SOLOv2MaskHead(nn.Layer): ...@@ -75,9 +75,7 @@ class SOLOv2MaskHead(nn.Layer):
ch_out=self.mid_channels, ch_out=self.mid_channels,
filter_size=3, filter_size=3,
stride=1, stride=1,
norm_type='gn', norm_type='gn'))
norm_name=conv_feat_name + '.conv' + str(i) + '.gn',
name=conv_feat_name + '.conv' + str(i)))
self.add_sublayer('conv_pre_feat' + str(i), conv_pre_feat) self.add_sublayer('conv_pre_feat' + str(i), conv_pre_feat)
self.convs_all_levels.append(conv_pre_feat) self.convs_all_levels.append(conv_pre_feat)
else: else:
...@@ -94,9 +92,7 @@ class SOLOv2MaskHead(nn.Layer): ...@@ -94,9 +92,7 @@ class SOLOv2MaskHead(nn.Layer):
ch_out=self.mid_channels, ch_out=self.mid_channels,
filter_size=3, filter_size=3,
stride=1, stride=1,
norm_type='gn', norm_type='gn'))
norm_name=conv_feat_name + '.conv' + str(j) + '.gn',
name=conv_feat_name + '.conv' + str(j)))
conv_pre_feat.add_sublayer( conv_pre_feat.add_sublayer(
conv_feat_name + '.conv' + str(j) + 'act', nn.ReLU()) conv_feat_name + '.conv' + str(j) + 'act', nn.ReLU())
conv_pre_feat.add_sublayer( conv_pre_feat.add_sublayer(
...@@ -114,9 +110,7 @@ class SOLOv2MaskHead(nn.Layer): ...@@ -114,9 +110,7 @@ class SOLOv2MaskHead(nn.Layer):
ch_out=self.out_channels, ch_out=self.out_channels,
filter_size=1, filter_size=1,
stride=1, stride=1,
norm_type='gn', norm_type='gn'))
norm_name=conv_pred_name + '.gn',
name=conv_pred_name))
def forward(self, inputs): def forward(self, inputs):
""" """
...@@ -216,9 +210,7 @@ class SOLOv2Head(nn.Layer): ...@@ -216,9 +210,7 @@ class SOLOv2Head(nn.Layer):
ch_out=self.seg_feat_channels, ch_out=self.seg_feat_channels,
filter_size=3, filter_size=3,
stride=1, stride=1,
norm_type='gn', norm_type='gn'))
norm_name='bbox_head.kernel_convs.{}.gn'.format(i),
name='bbox_head.kernel_convs.{}'.format(i)))
self.kernel_pred_convs.append(kernel_conv) self.kernel_pred_convs.append(kernel_conv)
ch_in = self.in_channels if i == 0 else self.seg_feat_channels ch_in = self.in_channels if i == 0 else self.seg_feat_channels
cate_conv = self.add_sublayer( cate_conv = self.add_sublayer(
...@@ -228,9 +220,7 @@ class SOLOv2Head(nn.Layer): ...@@ -228,9 +220,7 @@ class SOLOv2Head(nn.Layer):
ch_out=self.seg_feat_channels, ch_out=self.seg_feat_channels,
filter_size=3, filter_size=3,
stride=1, stride=1,
norm_type='gn', norm_type='gn'))
norm_name='bbox_head.cate_convs.{}.gn'.format(i),
name='bbox_head.cate_convs.{}'.format(i)))
self.cate_pred_convs.append(cate_conv) self.cate_pred_convs.append(cate_conv)
self.solo_kernel = self.add_sublayer( self.solo_kernel = self.add_sublayer(
...@@ -241,11 +231,9 @@ class SOLOv2Head(nn.Layer): ...@@ -241,11 +231,9 @@ class SOLOv2Head(nn.Layer):
kernel_size=3, kernel_size=3,
stride=1, stride=1,
padding=1, padding=1,
weight_attr=ParamAttr( weight_attr=ParamAttr(initializer=Normal(
name="bbox_head.solo_kernel.weight", mean=0., std=0.01)),
initializer=Normal( bias_attr=True))
mean=0., std=0.01)),
bias_attr=ParamAttr(name="bbox_head.solo_kernel.bias")))
self.solo_cate = self.add_sublayer( self.solo_cate = self.add_sublayer(
'bbox_head.solo_cate', 'bbox_head.solo_cate',
nn.Conv2D( nn.Conv2D(
...@@ -254,14 +242,10 @@ class SOLOv2Head(nn.Layer): ...@@ -254,14 +242,10 @@ class SOLOv2Head(nn.Layer):
kernel_size=3, kernel_size=3,
stride=1, stride=1,
padding=1, padding=1,
weight_attr=ParamAttr( weight_attr=ParamAttr(initializer=Normal(
name="bbox_head.solo_cate.weight", mean=0., std=0.01)),
initializer=Normal( bias_attr=ParamAttr(initializer=Constant(
mean=0., std=0.01)), value=float(-np.log((1 - 0.01) / 0.01))))))
bias_attr=ParamAttr(
name="bbox_head.solo_cate.bias",
initializer=Constant(
value=float(-np.log((1 - 0.01) / 0.01))))))
def _points_nms(self, heat, kernel_size=2): def _points_nms(self, heat, kernel_size=2):
hmax = F.max_pool2d(heat, kernel_size=kernel_size, stride=1, padding=1) hmax = F.max_pool2d(heat, kernel_size=kernel_size, stride=1, padding=1)
......
ppdet/modeling/heads/ssd_head.py
浏览文件 @ 5b20aa32
...@@ -28,8 +28,7 @@ class SepConvLayer(nn.Layer): ...@@ -28,8 +28,7 @@ class SepConvLayer(nn.Layer):
out_channels, out_channels,
kernel_size=3, kernel_size=3,
padding=1, padding=1,
conv_decay=0, conv_decay=0):
name=None):
super(SepConvLayer, self).__init__() super(SepConvLayer, self).__init__()
self.dw_conv = nn.Conv2D( self.dw_conv = nn.Conv2D(
in_channels=in_channels, in_channels=in_channels,
...@@ -38,16 +37,13 @@ class SepConvLayer(nn.Layer): ...@@ -38,16 +37,13 @@ class SepConvLayer(nn.Layer):
stride=1, stride=1,
padding=padding, padding=padding,
groups=in_channels, groups=in_channels,
weight_attr=ParamAttr( weight_attr=ParamAttr(regularizer=L2Decay(conv_decay)),
name=name + "_dw_weights", regularizer=L2Decay(conv_decay)),
bias_attr=False) bias_attr=False)
self.bn = nn.BatchNorm2D( self.bn = nn.BatchNorm2D(
in_channels, in_channels,
weight_attr=ParamAttr( weight_attr=ParamAttr(regularizer=L2Decay(0.)),
name=name + "_bn_scale", regularizer=L2Decay(0.)), bias_attr=ParamAttr(regularizer=L2Decay(0.)))
bias_attr=ParamAttr(
name=name + "_bn_offset", regularizer=L2Decay(0.)))
self.pw_conv = nn.Conv2D( self.pw_conv = nn.Conv2D(
in_channels=in_channels, in_channels=in_channels,
...@@ -55,8 +51,7 @@ class SepConvLayer(nn.Layer): ...@@ -55,8 +51,7 @@ class SepConvLayer(nn.Layer):
kernel_size=1, kernel_size=1,
stride=1, stride=1,
padding=0, padding=0,
weight_attr=ParamAttr( weight_attr=ParamAttr(regularizer=L2Decay(conv_decay)),
name=name + "_pw_weights", regularizer=L2Decay(conv_decay)),
bias_attr=False) bias_attr=False)
def forward(self, x): def forward(self, x):
...@@ -125,8 +120,7 @@ class SSDHead(nn.Layer): ...@@ -125,8 +120,7 @@ class SSDHead(nn.Layer):
out_channels=num_prior * 4, out_channels=num_prior * 4,
kernel_size=kernel_size, kernel_size=kernel_size,
padding=padding, padding=padding,
conv_decay=conv_decay, conv_decay=conv_decay))
name=box_conv_name))
self.box_convs.append(box_conv) self.box_convs.append(box_conv)
score_conv_name = "scores{}".format(i) score_conv_name = "scores{}".format(i)
...@@ -146,8 +140,7 @@ class SSDHead(nn.Layer): ...@@ -146,8 +140,7 @@ class SSDHead(nn.Layer):
out_channels=num_prior * self.num_classes, out_channels=num_prior * self.num_classes,
kernel_size=kernel_size, kernel_size=kernel_size,
padding=padding, padding=padding,
conv_decay=conv_decay, conv_decay=conv_decay))
name=score_conv_name))
self.score_convs.append(score_conv) self.score_convs.append(score_conv)
@classmethod @classmethod
......
ppdet/modeling/heads/ttf_head.py
浏览文件 @ 5b20aa32
...@@ -61,8 +61,7 @@ class HMHead(nn.Layer): ...@@ -61,8 +61,7 @@ class HMHead(nn.Layer):
LiteConv( LiteConv(
in_channels=ch_in if i == 0 else ch_out, in_channels=ch_in if i == 0 else ch_out,
out_channels=ch_out, out_channels=ch_out,
norm_type=norm_type, norm_type=norm_type))
name=lite_name))
head_conv.add_sublayer(lite_name + '.act', nn.ReLU6()) head_conv.add_sublayer(lite_name + '.act', nn.ReLU6())
else: else:
if dcn_head: if dcn_head:
...@@ -85,19 +84,17 @@ class HMHead(nn.Layer): ...@@ -85,19 +84,17 @@ class HMHead(nn.Layer):
bias_attr=ParamAttr( bias_attr=ParamAttr(
learning_rate=2., regularizer=L2Decay(0.)))) learning_rate=2., regularizer=L2Decay(0.))))
head_conv.add_sublayer(name + '.act', nn.ReLU()) head_conv.add_sublayer(name + '.act', nn.ReLU())
self.feat = self.add_sublayer('hm_feat', head_conv) self.feat = head_conv
bias_init = float(-np.log((1 - 0.01) / 0.01)) bias_init = float(-np.log((1 - 0.01) / 0.01))
self.head = self.add_sublayer( self.head = nn.Conv2D(
'hm_head', in_channels=ch_out,
nn.Conv2D( out_channels=num_classes,
in_channels=ch_out, kernel_size=1,
out_channels=num_classes, weight_attr=ParamAttr(initializer=Normal(0, 0.01)),
kernel_size=1, bias_attr=ParamAttr(
weight_attr=ParamAttr(initializer=Normal(0, 0.01)), learning_rate=2.,
bias_attr=ParamAttr( regularizer=L2Decay(0.),
learning_rate=2., initializer=Constant(bias_init)))
regularizer=L2Decay(0.),
initializer=Constant(bias_init))))
def forward(self, feat): def forward(self, feat):
out = self.feat(feat) out = self.feat(feat)
...@@ -139,8 +136,7 @@ class WHHead(nn.Layer): ...@@ -139,8 +136,7 @@ class WHHead(nn.Layer):
LiteConv( LiteConv(
in_channels=ch_in if i == 0 else ch_out, in_channels=ch_in if i == 0 else ch_out,
out_channels=ch_out, out_channels=ch_out,
norm_type=norm_type, norm_type=norm_type))
name=lite_name))
head_conv.add_sublayer(lite_name + '.act', nn.ReLU6()) head_conv.add_sublayer(lite_name + '.act', nn.ReLU6())
else: else:
if dcn_head: if dcn_head:
...@@ -164,16 +160,14 @@ class WHHead(nn.Layer): ...@@ -164,16 +160,14 @@ class WHHead(nn.Layer):
learning_rate=2., regularizer=L2Decay(0.)))) learning_rate=2., regularizer=L2Decay(0.))))
head_conv.add_sublayer(name + '.act', nn.ReLU()) head_conv.add_sublayer(name + '.act', nn.ReLU())
self.feat = self.add_sublayer('wh_feat', head_conv) self.feat = head_conv
self.head = self.add_sublayer( self.head = nn.Conv2D(
'wh_head', in_channels=ch_out,
nn.Conv2D( out_channels=4,
in_channels=ch_out, kernel_size=1,
out_channels=4, weight_attr=ParamAttr(initializer=Normal(0, 0.001)),
kernel_size=1, bias_attr=ParamAttr(
weight_attr=ParamAttr(initializer=Normal(0, 0.001)), learning_rate=2., regularizer=L2Decay(0.)))
bias_attr=ParamAttr(
learning_rate=2., regularizer=L2Decay(0.))))
def forward(self, feat): def forward(self, feat):
out = self.feat(feat) out = self.feat(feat)
......
ppdet/modeling/heads/yolo_head.py
浏览文件 @ 5b20aa32
...@@ -4,7 +4,6 @@ import paddle.nn.functional as F ...@@ -4,7 +4,6 @@ import paddle.nn.functional as F
from paddle import ParamAttr from paddle import ParamAttr
from paddle.regularizer import L2Decay from paddle.regularizer import L2Decay
from ppdet.core.workspace import register from ppdet.core.workspace import register
from ..backbones.darknet import ConvBNLayer
def _de_sigmoid(x, eps=1e-7): def _de_sigmoid(x, eps=1e-7):
......
ppdet/modeling/layers.py
浏览文件 @ 5b20aa32
...@@ -117,22 +117,18 @@ class ConvNormLayer(nn.Layer): ...@@ -117,22 +117,18 @@ class ConvNormLayer(nn.Layer):
norm_decay=0., norm_decay=0.,
norm_groups=32, norm_groups=32,
use_dcn=False, use_dcn=False,
norm_name=None,
bias_on=False, bias_on=False,
lr_scale=1., lr_scale=1.,
freeze_norm=False, freeze_norm=False,
initializer=Normal( initializer=Normal(
mean=0., std=0.01), mean=0., std=0.01),
skip_quant=False, skip_quant=False):
name=None):
super(ConvNormLayer, self).__init__() super(ConvNormLayer, self).__init__()
assert norm_type in ['bn', 'sync_bn', 'gn'] assert norm_type in ['bn', 'sync_bn', 'gn']
if bias_on: if bias_on:
bias_attr = ParamAttr( bias_attr = ParamAttr(
name=name + "_bias", initializer=Constant(value=0.), learning_rate=lr_scale)
initializer=Constant(value=0.),
learning_rate=lr_scale)
else: else:
bias_attr = False bias_attr = False
...@@ -145,9 +141,7 @@ class ConvNormLayer(nn.Layer): ...@@ -145,9 +141,7 @@ class ConvNormLayer(nn.Layer):
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
weight_attr=ParamAttr( weight_attr=ParamAttr(
name=name + "_weight", initializer=initializer, learning_rate=1.),
initializer=initializer,
learning_rate=1.),
bias_attr=bias_attr) bias_attr=bias_attr)
if skip_quant: if skip_quant:
self.conv.skip_quant = True self.conv.skip_quant = True
...@@ -161,24 +155,17 @@ class ConvNormLayer(nn.Layer): ...@@ -161,24 +155,17 @@ class ConvNormLayer(nn.Layer):
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=groups, groups=groups,
weight_attr=ParamAttr( weight_attr=ParamAttr(
name=name + "_weight", initializer=initializer, learning_rate=1.),
initializer=initializer,
learning_rate=1.),
bias_attr=True, bias_attr=True,
lr_scale=2., lr_scale=2.,
regularizer=L2Decay(norm_decay), regularizer=L2Decay(norm_decay),
skip_quant=skip_quant, skip_quant=skip_quant)
name=name)
norm_lr = 0. if freeze_norm else 1. norm_lr = 0. if freeze_norm else 1.
param_attr = ParamAttr( param_attr = ParamAttr(
name=norm_name + "_scale", learning_rate=norm_lr, regularizer=L2Decay(norm_decay))
learning_rate=norm_lr,
regularizer=L2Decay(norm_decay))
bias_attr = ParamAttr( bias_attr = ParamAttr(
name=norm_name + "_offset", learning_rate=norm_lr, regularizer=L2Decay(norm_decay))
learning_rate=norm_lr,
regularizer=L2Decay(norm_decay))
if norm_type == 'bn': if norm_type == 'bn':
self.norm = nn.BatchNorm2D( self.norm = nn.BatchNorm2D(
ch_out, weight_attr=param_attr, bias_attr=bias_attr) ch_out, weight_attr=param_attr, bias_attr=bias_attr)
...@@ -215,27 +202,21 @@ class LiteConv(nn.Layer): ...@@ -215,27 +202,21 @@ class LiteConv(nn.Layer):
stride=stride, stride=stride,
groups=in_channels, groups=in_channels,
norm_type=norm_type, norm_type=norm_type,
initializer=XavierUniform(), initializer=XavierUniform())
norm_name=name + '.conv1.norm',
name=name + '.conv1')
conv2 = ConvNormLayer( conv2 = ConvNormLayer(
in_channels, in_channels,
out_channels, out_channels,
filter_size=1, filter_size=1,
stride=stride, stride=stride,
norm_type=norm_type, norm_type=norm_type,
initializer=XavierUniform(), initializer=XavierUniform())
norm_name=name + '.conv2.norm',
name=name + '.conv2')
conv3 = ConvNormLayer( conv3 = ConvNormLayer(
out_channels, out_channels,
out_channels, out_channels,
filter_size=1, filter_size=1,
stride=stride, stride=stride,
norm_type=norm_type, norm_type=norm_type,
initializer=XavierUniform(), initializer=XavierUniform())
norm_name=name + '.conv3.norm',
name=name + '.conv3')
conv4 = ConvNormLayer( conv4 = ConvNormLayer(
out_channels, out_channels,
out_channels, out_channels,
...@@ -243,9 +224,7 @@ class LiteConv(nn.Layer): ...@@ -243,9 +224,7 @@ class LiteConv(nn.Layer):
stride=stride, stride=stride,
groups=out_channels, groups=out_channels,
norm_type=norm_type, norm_type=norm_type,
initializer=XavierUniform(), initializer=XavierUniform())
norm_name=name + '.conv4.norm',
name=name + '.conv4')
conv_list = [conv1, conv2, conv3, conv4] conv_list = [conv1, conv2, conv3, conv4]
self.lite_conv.add_sublayer('conv1', conv1) self.lite_conv.add_sublayer('conv1', conv1)
self.lite_conv.add_sublayer('relu6_1', nn.ReLU6()) self.lite_conv.add_sublayer('relu6_1', nn.ReLU6())
......
ppdet/modeling/necks/fpn.py
浏览文件 @ 5b20aa32
...@@ -105,10 +105,8 @@ class FPN(nn.Layer): ...@@ -105,10 +105,8 @@ class FPN(nn.Layer):
stride=1, stride=1,
norm_type=self.norm_type, norm_type=self.norm_type,
norm_decay=self.norm_decay, norm_decay=self.norm_decay,
norm_name=lateral_name + '_norm',
freeze_norm=self.freeze_norm, freeze_norm=self.freeze_norm,
initializer=XavierUniform(fan_out=in_c), initializer=XavierUniform(fan_out=in_c)))
name=lateral_name))
else: else:
lateral = self.add_sublayer( lateral = self.add_sublayer(
lateral_name, lateral_name,
...@@ -131,10 +129,8 @@ class FPN(nn.Layer): ...@@ -131,10 +129,8 @@ class FPN(nn.Layer):
stride=1, stride=1,
norm_type=self.norm_type, norm_type=self.norm_type,
norm_decay=self.norm_decay, norm_decay=self.norm_decay,
norm_name=fpn_name + '_norm',
freeze_norm=self.freeze_norm, freeze_norm=self.freeze_norm,
initializer=XavierUniform(fan_out=fan), initializer=XavierUniform(fan_out=fan)))
name=fpn_name))
else: else:
fpn_conv = self.add_sublayer( fpn_conv = self.add_sublayer(
fpn_name, fpn_name,
...@@ -166,10 +162,8 @@ class FPN(nn.Layer): ...@@ -166,10 +162,8 @@ class FPN(nn.Layer):
stride=2, stride=2,
norm_type=self.norm_type, norm_type=self.norm_type,
norm_decay=self.norm_decay, norm_decay=self.norm_decay,
norm_name=extra_fpn_name + '_norm',
freeze_norm=self.freeze_norm, freeze_norm=self.freeze_norm,
initializer=XavierUniform(fan_out=fan), initializer=XavierUniform(fan_out=fan)))
name=extra_fpn_name))
else: else:
extra_fpn_conv = self.add_sublayer( extra_fpn_conv = self.add_sublayer(
extra_fpn_name, extra_fpn_name,
......
ppdet/modeling/necks/ttf_fpn.py
浏览文件 @ 5b20aa32
...@@ -29,7 +29,7 @@ __all__ = ['TTFFPN'] ...@@ -29,7 +29,7 @@ __all__ = ['TTFFPN']
class Upsample(nn.Layer): class Upsample(nn.Layer):
def __init__(self, ch_in, ch_out, norm_type='bn', name=None): def __init__(self, ch_in, ch_out, norm_type='bn'):
super(Upsample, self).__init__() super(Upsample, self).__init__()
fan_in = ch_in * 3 * 3 fan_in = ch_in * 3 * 3
stdv = 1. / math.sqrt(fan_in) stdv = 1. / math.sqrt(fan_in)
...@@ -46,7 +46,7 @@ class Upsample(nn.Layer): ...@@ -46,7 +46,7 @@ class Upsample(nn.Layer):
regularizer=L2Decay(0.)) regularizer=L2Decay(0.))
self.bn = batch_norm( self.bn = batch_norm(
ch_out, norm_type=norm_type, initializer=Constant(1.), name=name) ch_out, norm_type=norm_type, initializer=Constant(1.))
def forward(self, feat): def forward(self, feat):
dcn = self.dcn(feat) dcn = self.dcn(feat)
...@@ -57,7 +57,7 @@ class Upsample(nn.Layer): ...@@ -57,7 +57,7 @@ class Upsample(nn.Layer):
class DeConv(nn.Layer): class DeConv(nn.Layer):
def __init__(self, ch_in, ch_out, norm_type='bn', name=None): def __init__(self, ch_in, ch_out, norm_type='bn'):
super(DeConv, self).__init__() super(DeConv, self).__init__()
self.deconv = nn.Sequential() self.deconv = nn.Sequential()
conv1 = ConvNormLayer( conv1 = ConvNormLayer(
...@@ -66,9 +66,7 @@ class DeConv(nn.Layer): ...@@ -66,9 +66,7 @@ class DeConv(nn.Layer):
stride=1, stride=1,
filter_size=1, filter_size=1,
norm_type=norm_type, norm_type=norm_type,
initializer=XavierUniform(), initializer=XavierUniform())
norm_name=name + '.conv1.norm',
name=name + '.conv1')
conv2 = nn.Conv2DTranspose( conv2 = nn.Conv2DTranspose(
in_channels=ch_out, in_channels=ch_out,
out_channels=ch_out, out_channels=ch_out,
...@@ -78,17 +76,14 @@ class DeConv(nn.Layer): ...@@ -78,17 +76,14 @@ class DeConv(nn.Layer):
groups=ch_out, groups=ch_out,
weight_attr=ParamAttr(initializer=XavierUniform()), weight_attr=ParamAttr(initializer=XavierUniform()),
bias_attr=False) bias_attr=False)
bn = batch_norm( bn = batch_norm(ch_out, norm_type=norm_type, norm_decay=0.)
ch_out, norm_type=norm_type, norm_decay=0., name=name + '.bn')
conv3 = ConvNormLayer( conv3 = ConvNormLayer(
ch_in=ch_out, ch_in=ch_out,
ch_out=ch_out, ch_out=ch_out,
stride=1, stride=1,
filter_size=1, filter_size=1,
norm_type=norm_type, norm_type=norm_type,
initializer=XavierUniform(), initializer=XavierUniform())
norm_name=name + '.conv3.norm',
name=name + '.conv3')
self.deconv.add_sublayer('conv1', conv1) self.deconv.add_sublayer('conv1', conv1)
self.deconv.add_sublayer('relu6_1', nn.ReLU6()) self.deconv.add_sublayer('relu6_1', nn.ReLU6())
...@@ -103,12 +98,10 @@ class DeConv(nn.Layer): ...@@ -103,12 +98,10 @@ class DeConv(nn.Layer):
class LiteUpsample(nn.Layer): class LiteUpsample(nn.Layer):
def __init__(self, ch_in, ch_out, norm_type='bn', name=None): def __init__(self, ch_in, ch_out, norm_type='bn'):
super(LiteUpsample, self).__init__() super(LiteUpsample, self).__init__()
self.deconv = DeConv( self.deconv = DeConv(ch_in, ch_out, norm_type=norm_type)
ch_in, ch_out, norm_type=norm_type, name=name + '.deconv') self.conv = LiteConv(ch_in, ch_out, norm_type=norm_type)
self.conv = LiteConv(
ch_in, ch_out, norm_type=norm_type, name=name + '.liteconv')
def forward(self, inputs): def forward(self, inputs):
deconv_up = self.deconv(inputs) deconv_up = self.deconv(inputs)
...@@ -139,8 +132,7 @@ class ShortCut(nn.Layer): ...@@ -139,8 +132,7 @@ class ShortCut(nn.Layer):
in_channels=in_channels, in_channels=in_channels,
out_channels=ch_out, out_channels=ch_out,
with_act=i < layer_num - 1, with_act=i < layer_num - 1,
norm_type=norm_type, norm_type=norm_type))
name=shortcut_name))
else: else:
shortcut_conv.add_sublayer( shortcut_conv.add_sublayer(
shortcut_name, shortcut_name,
...@@ -155,7 +147,7 @@ class ShortCut(nn.Layer): ...@@ -155,7 +147,7 @@ class ShortCut(nn.Layer):
if i < layer_num - 1: if i < layer_num - 1:
shortcut_conv.add_sublayer(shortcut_name + '.act', shortcut_conv.add_sublayer(shortcut_name + '.act',
nn.ReLU()) nn.ReLU())
self.shortcut = self.add_sublayer('short', shortcut_conv) self.shortcut = self.add_sublayer('shortcut', shortcut_conv)
def forward(self, feat): def forward(self, feat):
out = self.shortcut(feat) out = self.shortcut(feat)
...@@ -208,10 +200,7 @@ class TTFFPN(nn.Layer): ...@@ -208,10 +200,7 @@ class TTFFPN(nn.Layer):
upsample = self.add_sublayer( upsample = self.add_sublayer(
'upsample.' + str(i), 'upsample.' + str(i),
upsample_module( upsample_module(
in_c, in_c, out_c, norm_type=norm_type))
out_c,
norm_type=norm_type,
name='deconv_layers.' + str(i)))
self.upsample_list.append(upsample) self.upsample_list.append(upsample)
if i < self.shortcut_len: if i < self.shortcut_len:
shortcut = self.add_sublayer( shortcut = self.add_sublayer(
......
ppdet/modeling/necks/yolo_fpn.py
浏览文件 @ 5b20aa32
...@@ -25,9 +25,9 @@ from ..shape_spec import ShapeSpec ...@@ -25,9 +25,9 @@ from ..shape_spec import ShapeSpec
__all__ = ['YOLOv3FPN', 'PPYOLOFPN'] __all__ = ['YOLOv3FPN', 'PPYOLOFPN']
def add_coord(x): def add_coord(x, data_format):
b = x.shape[0] b = x.shape[0]
if self.data_format == 'NCHW': if data_format == 'NCHW':
h = x.shape[2] h = x.shape[2]
w = x.shape[3] w = x.shape[3]
else: else:
...@@ -35,14 +35,14 @@ def add_coord(x): ...@@ -35,14 +35,14 @@ def add_coord(x):
w = x.shape[2] w = x.shape[2]
gx = paddle.arange(w, dtype='float32') / (w - 1.) * 2.0 - 1. gx = paddle.arange(w, dtype='float32') / (w - 1.) * 2.0 - 1.
if self.data_format == 'NCHW': if data_format == 'NCHW':
gx = gx.reshape([1, 1, 1, w]).expand([b, 1, h, w]) gx = gx.reshape([1, 1, 1, w]).expand([b, 1, h, w])
else: else:
gx = gx.reshape([1, 1, w, 1]).expand([b, h, w, 1]) gx = gx.reshape([1, 1, w, 1]).expand([b, h, w, 1])
gx.stop_gradient = True gx.stop_gradient = True
gy = paddle.arange(h, dtype='float32') / (h - 1.) * 2.0 - 1. gy = paddle.arange(h, dtype='float32') / (h - 1.) * 2.0 - 1.
if self.data_format == 'NCHW': if data_format == 'NCHW':
gy = gy.reshape([1, 1, h, 1]).expand([b, 1, h, w]) gy = gy.reshape([1, 1, h, 1]).expand([b, 1, h, w])
else: else:
gy = gy.reshape([1, h, 1, 1]).expand([b, h, w, 1]) gy = gy.reshape([1, h, 1, 1]).expand([b, h, w, 1])
...@@ -237,7 +237,7 @@ class CoordConv(nn.Layer): ...@@ -237,7 +237,7 @@ class CoordConv(nn.Layer):
self.data_format = data_format self.data_format = data_format
def forward(self, x): def forward(self, x):
gx, gy = add_coord(x) gx, gy = add_coord(x, self.data_format)
if self.data_format == 'NCHW': if self.data_format == 'NCHW':
y = paddle.concat([x, gx, gy], axis=1) y = paddle.concat([x, gx, gy], axis=1)
else: else:
...@@ -509,7 +509,7 @@ class PPYOLOFPN(nn.Layer): ...@@ -509,7 +509,7 @@ class PPYOLOFPN(nn.Layer):
norm_type='bn', norm_type='bn',
data_format='NCHW', data_format='NCHW',
coord_conv=False, coord_conv=False,
conv_block_num=3, conv_block_num=2,
drop_block=False, drop_block=False,
block_size=3, block_size=3,
keep_prob=0.9, keep_prob=0.9,
......
ppdet/modeling/ops.py
浏览文件 @ 5b20aa32
...@@ -53,9 +53,7 @@ def batch_norm(ch, ...@@ -53,9 +53,7 @@ def batch_norm(ch,
norm_type='bn', norm_type='bn',
norm_decay=0., norm_decay=0.,
initializer=None, initializer=None,
name=None,
data_format='NCHW'): data_format='NCHW'):
bn_name = name + '.bn'
if norm_type == 'sync_bn': if norm_type == 'sync_bn':
batch_norm = nn.SyncBatchNorm batch_norm = nn.SyncBatchNorm
else: else:
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或