def __init__(self, in_channels: int, out_channels: int, index: int):
super(SSH, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.half_out_channels = int(out_channels / 2)
self.quater_out_channels = int(self.half_out_channels / 2)
self.index = index
self.ssh_3x3 = nn.Sequential(
nn.Conv2d(in_channels=self.in_channels, out_channels=self.half_out_channels, kernel_size=3, stride=1, padding=1)
)
self.ssh_dimred = nn.Sequential(
nn.Conv2d(in_channels=self.in_channels, out_channels=self.quater_out_channels, kernel_size=3, stride=1, padding=1),
nn.ReLU()
)
self.ssh_5x5 = nn.Sequential(
nn.Conv2d(in_channels=self.quater_out_channels, out_channels=self.quater_out_channels, kernel_size=3, stride=1, padding=1)
)
self.ssh_7x7 = nn.Sequential(
nn.Conv2d(in_channels=self.quater_out_channels, out_channels=self.quater_out_channels, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.Conv2d(in_channels=self.quater_out_channels, out_channels=self.quater_out_channels, kernel_size=3, stride=1, padding=1)
)
self.out_relu = nn.ReLU()
def build_conv_block(self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
stride: int = 1,
padding: int = 1,
n_conv: int = 2,
with_pool: bool = False):
layers = []
if with_pool:
layers.append(nn.MaxPool2d(kernel_size=2, stride=2))
# convx_1
layers += [
nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding),
nn.ReLU()
]
# convx_2 -> convx_(n_conv)
for i in range(1, n_conv):
add_layers = [
nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding),
nn.ReLU()
]
layers += add_layers
# return as sequential
return nn.Sequential(*layers)
def __init__(self):
super(LFFDv1, self).__init__()
self.backbone = nn.Sequential(
nn.Conv2d(in_channels=3,
out_channels=64,
kernel_size=3,
stride=2,
padding=0), # downsample by 2
nn.ReLU(),
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=2,
padding=0), # downsample by 2
ResBlock(64),
ResBlock(64),
ResBlock(64))
self.rb1 = ResBlock(64, det_out=True)
self.det1 = DetBlock(64)
self.relu_conv10 = nn.ReLU()
self.conv11 = nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=2,
padding=0)
self.rb2 = ResBlock(64)
self.det2 = DetBlock(64)
self.rb3 = ResBlock(64, det_out=True)
self.det3 = DetBlock(64)
self.relu_conv15 = nn.ReLU()
self.conv16 = nn.Conv2d(in_channels=64,
out_channels=128,
kernel_size=3,
stride=2,
padding=0)
self.rb4 = ResBlock(128)
self.det4 = DetBlock(64)
self.relu_conv18 = nn.ReLU()
self.conv19 = nn.Conv2d(in_channels=128,
out_channels=128,
kernel_size=3,
stride=2,
padding=0)
self.rb5 = ResBlock(128)
self.det5 = DetBlock(128)
self.rb6 = ResBlock(128, det_out=True)
self.det6 = DetBlock(128)
self.rb7 = ResBlock(128, det_out=True)
self.det7 = DetBlock(128)
self.relu_conv25 = nn.ReLU()
self.det8 = DetBlock(128)
def __init__(self):
super(SSH, self).__init__()
# backbone
self.vgg16 = nn.ModuleList(make_layers(vgg_cfgs['D']))
# SSH - M3
self.M3 = M_Module(512, 256, 128)
self.M3_bbox_pred = nn.Conv2d(512, 8, 1, 1, 0)
self.M3_cls_score = nn.Conv2d(512, 4, 1, 1, 0)
self.M3_cls_score_softmax = nn.Softmax(dim=1)
# SSH - M2
self.M2 = M_Module(512, 256, 128)
self.M2_bbox_pred = nn.Conv2d(512, 8, 1, 1, 0)
self.M2_cls_score = nn.Conv2d(512, 4, 1, 1, 0)
self.M2_cls_score_softmax = nn.Softmax(dim=1)
# SSH - M1
self.conv4_128 = nn.Conv2d(512, 128, 1, 1, 0)
self.conv4_128_relu = nn.ReLU(inplace=True)
self.conv5_128 = nn.Conv2d(512, 128, 1, 1, 0)
self.conv5_128_relu = nn.ReLU(inplace=True)
self.conv5_128_up = nn.ConvTranspose2d(128, 128, 4, 2, 1, groups=128, bias=False)
self.eltadd = nn.EltAdd()
self.conv4_fuse_final = nn.Conv2d(128, 128, 3, 1, 1)
self.conv4_fuse_final_relu = nn.ReLU(inplace=True)
self.M1 = M_Module(128, 128, 64)
self.M1_bbox_pred = nn.Conv2d(256, 8, 1, 1, 0)
self.M1_cls_score = nn.Conv2d(256, 4, 1, 1, 0)
self.M1_cls_score_softmax = nn.Softmax(dim=1)
def _conv_dw(in_channels, out_channels, stride):
return nn.Sequential(
nn.Conv2d(in_channels=in_channels, out_channels=in_channels, kernel_size=3, stride=stride, padding=1, groups=in_channels, bias=False),
nn.BatchNorm2d(in_channels),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0, bias=False),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True)
)
def __init__(self, in_planes, out_planes, kernel_size, stride=1, padding=0, dilation=1, groups=1, relu=True, bn=True):
super(BasicConv, self).__init__()
self.out_channels = out_planes
if bn:
self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=False)
self.bn = nn.BatchNorm2d(out_planes, eps=1e-5, momentum=0.01, affine=True)
self.relu = nn.ReLU(inplace=True) if relu else None
else:
self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=True)
self.bn = None
self.relu = nn.ReLU(inplace=True) if relu else None
def make_layers(cfg, batch_norm=False):
layers = []
in_channels = 3
for v in cfg:
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1)
if batch_norm:
layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)]
else:
layers += [conv2d, nn.ReLU(inplace=True)]
in_channels = v
return layers
def __init__(self, input_channels, output_channels):
super(DeepHeadModule, self).__init__()
self._input_channels = input_channels
self._output_channels = output_channels
self._mid_channels = min(self._input_channels, 256)
#print(self._mid_channels)
self.conv1 = nn.Conv2d(self._input_channels,
self._mid_channels,
kernel_size=3,
dilation=1,
stride=1,
padding=1)
self.conv2 = nn.Conv2d(self._mid_channels,
self._mid_channels,
kernel_size=3,
dilation=1,
stride=1,
padding=1)
self.conv3 = nn.Conv2d(self._mid_channels,
self._mid_channels,
kernel_size=3,
dilation=1,
stride=1,
padding=1)
self.conv4 = nn.Conv2d(self._mid_channels,
self._output_channels,
kernel_size=1,
dilation=1,
stride=1,
padding=0)
self.relu = nn.ReLU(inplace=True)
def __init__(self,
in_channels,
out_channels,
kernel_sizes,
strides=None,
paddings=None,
with_pool=True):
super(Conv_Block, self).__init__()
assert len(in_channels) == len(out_channels)
assert len(out_channels) == len(kernel_sizes)
if strides is not None:
assert len(kernel_sizes) == len(strides)
self.pool = None
if with_pool:
self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
groups = len(in_channels)
convs = []
for i in range(groups):
convs.append(
nn.Conv2d(in_channels=in_channels[i],
out_channels=out_channels[i],
kernel_size=kernel_sizes[i],
stride=strides[i],
padding=paddings[i]))
convs.append(nn.ReLU(inplace=True))
self.feature = nn.Sequential(*convs)
def __init__(self, in_channels, out_channels, kernel_size, stride, padding, **kwargs):
super(ConvBNReLU, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, bias=True, **kwargs)
self.bn = nn.BatchNorm2d(out_channels)
self.relu = nn.ReLU(inplace=True)
def __init__(self, in_planes, out_planes, stride=1, scale=0.1, map_reduce=8, vision=1, groups=1):
super(BasicRFB, self).__init__()
self.scale = scale
self.out_channels = out_planes
inter_planes = in_planes // map_reduce
self.branch0 = nn.Sequential(
BasicConv(in_planes, inter_planes, kernel_size=1, stride=1, groups=groups, relu=False),
BasicConv(inter_planes, 2 * inter_planes, kernel_size=(3, 3), stride=stride, padding=(1, 1), groups=groups),
BasicConv(2 * inter_planes, 2 * inter_planes, kernel_size=3, stride=1, padding=vision + 1, dilation=vision + 1, relu=False, groups=groups)
)
self.branch1 = nn.Sequential(
BasicConv(in_planes, inter_planes, kernel_size=1, stride=1, groups=groups, relu=False),
BasicConv(inter_planes, 2 * inter_planes, kernel_size=(3, 3), stride=stride, padding=(1, 1), groups=groups),
BasicConv(2 * inter_planes, 2 * inter_planes, kernel_size=3, stride=1, padding=vision + 2, dilation=vision + 2, relu=False, groups=groups)
)
self.branch2 = nn.Sequential(
BasicConv(in_planes, inter_planes, kernel_size=1, stride=1, groups=groups, relu=False),
BasicConv(inter_planes, (inter_planes // 2) * 3, kernel_size=3, stride=1, padding=1, groups=groups),
BasicConv((inter_planes // 2) * 3, 2 * inter_planes, kernel_size=3, stride=stride, padding=1, groups=groups),
BasicConv(2 * inter_planes, 2 * inter_planes, kernel_size=3, stride=1, padding=vision + 4, dilation=vision + 4, relu=False, groups=groups)
)
self.ConvLinear = BasicConv(6 * inter_planes, out_planes, kernel_size=1, stride=1, relu=False)
self.shortcut = BasicConv(in_planes, out_planes, kernel_size=1, stride=stride, relu=False)
self.relu = nn.ReLU(inplace=False)
self.eltadd = nn.EltAdd()
def __init__(self, up_from_channels, up_to_channels):
super(LFPN, self).__init__()
self.conv1 = nn.Conv2d(up_from_channels, up_to_channels, kernel_size=1)
self.conv1_relu = nn.ReLU(inplace=True)
self.upsampling = nn.ConvTranspose2d(up_to_channels,
up_to_channels,
kernel_size=4,
stride=2,
padding=1,
groups=up_to_channels,
bias=False)
self.conv2 = nn.Conv2d(up_to_channels, up_to_channels, kernel_size=1)
self.conv2_relu = nn.ReLU(inplace=True)
self.eltmul = nn.EltMul()
def __init__(self, in_channels, out_channels_left, out_channels_right):
super(M_Module, self).__init__()
inc = in_channels
ocl, ocr = out_channels_left, out_channels_right
# left branch
self.ssh_3x3 = nn.Conv2d(inc, ocl, 3, 1, 1)
# right branch
self.ssh_dimred = nn.Conv2d(inc, ocr, 3, 1, 1)
self.ssh_dimred_relu = nn.ReLU(inplace=True)
self.ssh_5x5 = nn.Conv2d(ocr, ocr, 3, 1, 1)
self.ssh_7x7_1 = nn.Conv2d(ocr, ocr, 3, 1, 1)
self.ssh_7x7_1_relu = nn.ReLU(inplace=True)
self.ssh_7x7 = nn.Conv2d(ocr, ocr, 3, 1, 1)
self.ssh_output_relu = nn.ReLU(inplace=True)
def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1):
super(IdentityBlock, self).__init__()
out_channels_1, out_channels_2, out_channels_3 = out_channels//4, out_channels//4, out_channels
self.conv1 = nn.Conv2d(in_channels, out_channels_1, kernel_size=(1, 1))
self.bn1 = nn.BatchNorm2d(out_channels_1)
self.relu1 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(out_channels_1, out_channels_2, kernel_size=(kernel_size, kernel_size), padding=(padding, padding), dilation=(dilation, dilation))
self.bn2 = nn.BatchNorm2d(out_channels_2)
self.relu2 = nn.ReLU(inplace=True)
self.conv3 = nn.Conv2d(out_channels_2, out_channels_3, kernel_size=(1, 1))
self.bn3 = nn.BatchNorm2d(out_channels_3)
self.eltadd = nn.EltAdd()
self.relu_f = nn.ReLU(inplace=True)
def __init__(self, mode='slim'):
super(ULFG, self).__init__()
self.mode = mode
self.base_channel = 8 * 2
self.backbone = nn.Sequential(
_conv_bn(3, self.base_channel, 2), # 160*120
_conv_dw(self.base_channel, self.base_channel * 2, 1),
_conv_dw(self.base_channel * 2, self.base_channel * 2, 2), # 80*60
_conv_dw(self.base_channel * 2, self.base_channel * 2, 1),
_conv_dw(self.base_channel * 2, self.base_channel * 4, 2), # 40*30
_conv_dw(self.base_channel * 4, self.base_channel * 4, 1),
_conv_dw(self.base_channel * 4, self.base_channel * 4, 1),
_conv_dw(self.base_channel * 4, self.base_channel * 4, 1),
_conv_dw(self.base_channel * 4, self.base_channel * 8, 2), # 20*15
_conv_dw(self.base_channel * 8, self.base_channel * 8, 1),
_conv_dw(self.base_channel * 8, self.base_channel * 8, 1),
_conv_dw(self.base_channel * 8, self.base_channel * 16, 2), # 10*8
_conv_dw(self.base_channel * 16, self.base_channel * 16, 1)
)
if self.mode == 'rfb':
self.backbone[7] = BasicRFB(self.base_channel * 4, self.base_channel * 4, stride=1, scale=1.0)
self.source_layer_indexes = [8, 11, 13]
self.extras = nn.Sequential(
nn.Conv2d(in_channels=self.base_channel * 16, out_channels=self.base_channel * 4, kernel_size=1),
nn.ReLU(),
_seperable_conv2d(in_channels=self.base_channel * 4, out_channels=self.base_channel * 16, kernel_size=3, stride=2, padding=1),
nn.ReLU()
)
self.regression_headers = nn.ModuleList([
_seperable_conv2d(in_channels=self.base_channel * 4, out_channels=3 * 4, kernel_size=3, padding=1),
_seperable_conv2d(in_channels=self.base_channel * 8, out_channels=2 * 4, kernel_size=3, padding=1),
_seperable_conv2d(in_channels=self.base_channel * 16, out_channels=2 * 4, kernel_size=3, padding=1),
nn.Conv2d(in_channels=self.base_channel * 16, out_channels=3 * 4, kernel_size=3, padding=1)
])
self.classification_headers = nn.ModuleList([
_seperable_conv2d(in_channels=self.base_channel * 4, out_channels=3 * 2, kernel_size=3, padding=1),
_seperable_conv2d(in_channels=self.base_channel * 8, out_channels=2 * 2, kernel_size=3, padding=1),
_seperable_conv2d(in_channels=self.base_channel * 16, out_channels=2 * 2, kernel_size=3, padding=1),
nn.Conv2d(in_channels=self.base_channel * 16, out_channels=3 * 2, kernel_size=3, padding=1)
])
self.softmax = nn.Softmax(dim=2)
def __init__(self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=False, add_relu=True, add_bn=True, eps=1e-5):
super(ConvBlock, self).__init__()
self.conv = nn.Conv2d(in_channel, out_channel, kernel_size, stride, padding, bias=bias)
self.relu = None
self.bn = None
if add_relu:
self.relu = nn.ReLU()
if add_bn:
self.bn = nn.BatchNorm2d(out_channel, eps=eps)
def __init__(self, channels, det_out=False):
super(ResBlock, self).__init__()
self.channels = channels
self.det_out = det_out
self.relu = nn.ReLU()
self.block = nn.Sequential(
nn.Conv2d(in_channels=self.channels,
out_channels=self.channels,
kernel_size=3,
stride=1,
padding=1), nn.ReLU(),
nn.Conv2d(in_channels=self.channels,
out_channels=self.channels,
kernel_size=3,
stride=1,
padding=1))
self.eltadd = nn.EltAdd()
def __init__(self, in_channels):
super(CPM, self).__init__()
# residual
self.branch1 = Conv_BN(in_channels, 1024, 1, 1, 0, act=None)
self.branch2a = Conv_BN(in_channels, 256, 1, 1, 0, act='relu')
self.branch2b = Conv_BN(256, 256, 3, 1, 1, act='relu')
self.branch2c = Conv_BN(256, 1024, 1, 1, 0, act=None)
self.eltadd = nn.EltAdd()
self.rescomb_relu = nn.ReLU(inplace=True)
# ssh
self.ssh_1_conv = nn.Conv2d(1024, 256, 3, 1, 1)
self.ssh_dimred_conv = nn.Conv2d(1024, 128, 3, 1, 1)
self.ssh_dimred_relu = nn.ReLU(inplace=True)
self.ssh_2_conv = nn.Conv2d(128, 128, 3, 1, 1)
self.ssh_3a_conv = nn.Conv2d(128, 128, 3, 1, 1)
self.ssh_3a_relu = nn.ReLU(inplace=True)
self.ssh_3b_conv = nn.Conv2d(128, 128, 3, 1, 1)
self.concat_relu = nn.ReLU(inplace=True)
def __init__(self, channel_size):
super(FEM, self).__init__()
self.cs = channel_size
self.cpm1 = nn.Conv2d(self.cs,
256,
kernel_size=3,
dilation=1,
stride=1,
padding=1)
self.cpm2 = nn.Conv2d(self.cs,
256,
kernel_size=3,
dilation=2,
stride=1,
padding=2)
self.cpm3 = nn.Conv2d(256,
128,
kernel_size=3,
dilation=1,
stride=1,
padding=1)
self.cpm4 = nn.Conv2d(256,
128,
kernel_size=3,
dilation=2,
stride=1,
padding=2)
self.cpm5 = nn.Conv2d(128,
128,
kernel_size=3,
dilation=1,
stride=1,
padding=1)
self.relu1 = nn.ReLU(inplace=True)
self.relu2 = nn.ReLU(inplace=True)
self.relu3 = nn.ReLU(inplace=True)
self.relu4 = nn.ReLU(inplace=True)
self.relu5 = nn.ReLU(inplace=True)
def build_conv_block(self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
stride: int = 1,
padding: int = 1,
dilation: int = 1,
n_conv: int = 2,
with_pool: bool = False):
layers = [
nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding,
dilation),
nn.ReLU()
]
for i in range(1, n_conv):
layers += [
nn.Conv2d(out_channels, out_channels, kernel_size, stride,
padding),
nn.ReLU()
]
if with_pool:
layers += [nn.MaxPool2d(2, 2)]
return nn.Sequential(*layers)
def __init__(self):
super(LFFDv2, self).__init__()
self.backbone = nn.Sequential(
nn.Conv2d(in_channels=3,
out_channels=64,
kernel_size=3,
stride=2,
padding=0), # downsample by 2
nn.ReLU(),
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=2,
padding=0), # downsample by 2
ResBlock(64),
ResBlock(64),
ResBlock(64))
self.relu_conv8 = nn.ReLU()
self.conv9 = nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=2,
padding=0) # downsample by 2
self.rb1 = ResBlock(64)
self.det1 = DetBlock(64)
self.relu_conv11 = nn.ReLU()
self.conv12 = nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=2,
padding=0) # downsample by 2
self.rb2 = ResBlock(64)
self.det2 = DetBlock(64)
self.relu_conv14 = nn.ReLU()
self.conv15 = nn.Conv2d(in_channels=64,
out_channels=128,
kernel_size=3,
stride=2,
padding=0) # downsample by 2
self.rb3 = ResBlock(128)
self.det3 = DetBlock(64)
self.relu_conv17 = nn.ReLU()
self.conv18 = nn.Conv2d(in_channels=128,
out_channels=128,
kernel_size=3,
stride=2,
padding=0) # downsample by 2
self.rb4 = ResBlock(128)
self.det4 = DetBlock(128)
self.relu_conv20 = nn.ReLU()
self.det5 = DetBlock(128)
def __init__(self, in_channels):
super(DetBlock, self).__init__()
self.in_channels = in_channels
self.det_channels = 128
self.det_conv = nn.Conv2d(in_channels=self.in_channels,
out_channels=self.det_channels,
kernel_size=1,
stride=1,
padding=0)
self.det_relu = nn.ReLU()
self.bbox_conv = nn.Conv2d(in_channels=self.det_channels,
out_channels=self.det_channels,
kernel_size=1,
stride=1,
padding=0)
self.bbox_relu = nn.ReLU()
self.bbox_out_conv = nn.Conv2d(in_channels=self.det_channels,
out_channels=4,
kernel_size=1,
stride=1,
padding=0)
self.score_conv = nn.Conv2d(in_channels=self.det_channels,
out_channels=self.det_channels,
kernel_size=1,
stride=1,
padding=0)
self.score_relu = nn.ReLU()
self.score_out_conv = nn.Conv2d(in_channels=self.det_channels,
out_channels=2,
kernel_size=1,
stride=1,
padding=0)
self.softmax = nn.Softmax(dim=1)
def upsample(in_channels, out_channels): # should use F.inpterpolate
return nn.Sequential(
nn.Conv2d(in_channels=in_channels,
out_channels=in_channels,
kernel_size=(3, 3),
stride=1,
padding=1,
groups=in_channels,
bias=False),
nn.Conv2d(in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
stride=1,
padding=0,
bias=False), nn.BatchNorm2d(out_channels), nn.ReLU())
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(Bottleneck, self).__init__()
self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes,
planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * 4)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
self.eltadd = nn.EltAdd()
def __init__(self,
in_channels,
out_channels,
kernel_sizes,
strides=1,
paddings=0,
act='relu',
bias=False):
super(Conv_BN, self).__init__()
self.conv = nn.Conv2d(in_channels,
out_channels,
kernel_sizes,
strides,
paddings,
bias=bias)
self.bn = nn.BatchNorm2d(out_channels, eps=0.001, momentum=0.999)
self.act = None
if act == 'relu':
self.act = nn.ReLU(inplace=True)
def __init__(self, in_channel, out_channel):
super(SSH, self).__init__()
assert out_channel % 4 == 0
leaky = 0
if (out_channel <= 64):
leaky = 0.1
self.conv3X3 = conv_bn_no_relu(in_channel, out_channel // 2, stride=1)
self.conv5X5_1 = conv_bn(in_channel,
out_channel // 4,
stride=1,
leaky=leaky)
self.conv5X5_2 = conv_bn_no_relu(out_channel // 4,
out_channel // 4,
stride=1)
self.conv7X7_2 = conv_bn(out_channel // 4,
out_channel // 4,
stride=1,
leaky=leaky)
self.conv7x7_3 = conv_bn_no_relu(out_channel // 4,
out_channel // 4,
stride=1)
self.relu = nn.ReLU()
def __init__(self):
super(S3FD, self).__init__()
# backbone
self.vgg16 = nn.ModuleList(make_layers(vgg_cfgs['D']))
# s3fd specific
self.conv_fc6 = nn.Conv2d(512, 1024, 3, 1, 1)
self.relu_fc6 = nn.ReLU()
self.conv_fc7 = nn.Conv2d(1024, 1024, 1, 1, 0)
self.relu_fc7 = nn.ReLU()
self.conv6_1 = nn.Conv2d(1024, 256, 1, 1, 0)
self.relu_conv6_1 = nn.ReLU()
self.conv6_2 = nn.Conv2d(256, 512, 3, 2, 1)
self.relu_conv6_2 = nn.ReLU()
self.conv7_1 = nn.Conv2d(512, 128, 1, 1, 0)
self.relu_conv7_1 = nn.ReLU()
self.conv7_2 = nn.Conv2d(128, 256, 3, 2, 1)
self.relu_conv7_2 = nn.ReLU()
self.l2norm_conv3_3 = nn.L2Norm2d(256, 10)
self.l2norm_conv4_3 = nn.L2Norm2d(512, 8)
self.l2norm_conv5_3 = nn.L2Norm2d(512, 5)
# Detection Head - mbox_loc
self.mbox_loc_conv3_3_norm = nn.Conv2d(256, 4, 3, 1, 1)
self.mbox_loc_conv4_3_norm = nn.Conv2d(512, 4, 3, 1, 1)
self.mbox_loc_conv5_3_norm = nn.Conv2d(512, 4, 3, 1, 1)
self.mbox_loc_conv_fc7 = nn.Conv2d(1024, 4, 3, 1, 1)
self.mbox_loc_conv6_2 = nn.Conv2d(512, 4, 3, 1, 1)
self.mbox_loc_conv7_2 = nn.Conv2d(256, 4, 3, 1, 1)
# Detection Head - mbox_conf
self.mbox_conf_conv3_3_norm = nn.Conv2d(
256, 4, 3, 1, 1) # 4->2 through maxout at channels 0~2
self.mbox_conf_conv4_3_norm = nn.Conv2d(512, 2, 3, 1, 1)
self.mbox_conf_conv5_3_norm = nn.Conv2d(512, 2, 3, 1, 1)
self.mbox_conf_conv_fc7 = nn.Conv2d(1024, 2, 3, 1, 1)
self.mbox_conf_conv6_2 = nn.Conv2d(512, 2, 3, 1, 1)
self.mbox_conf_conv7_2 = nn.Conv2d(256, 2, 3, 1, 1)
# Detection Head - mbox_conf - softmax
self.softmax = nn.Softmax(dim=-1)
def __init__(self, in_channels, out_channels, **kwargs):
super(BasicConv2d, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs)
self.bn = nn.BatchNorm2d(out_channels, eps=1e-5)
self.relu = nn.ReLU()
def __init__(self):
super(DSFD, self).__init__()
self.size = 640
self.num_classes = 2
######
# build backbone
######
resnet152 = vision.models.resnet152()
self.layer1 = nn.Sequential(resnet152.conv1, resnet152.bn1,
resnet152.relu, resnet152.maxpool,
resnet152.layer1)
self.layer2 = nn.Sequential(resnet152.layer2)
self.layer3 = nn.Sequential(resnet152.layer3)
self.layer4 = nn.Sequential(resnet152.layer4)
self.layer5 = nn.Sequential(*[
nn.Conv2d(2048, 512, kernel_size=1),
nn.BatchNorm2d(512),
nn.ReLU(inplace=True),
nn.Conv2d(512, 512, kernel_size=3, padding=1, stride=2),
nn.BatchNorm2d(512),
nn.ReLU(inplace=True)
])
self.layer6 = nn.Sequential(*[
nn.Conv2d(
512,
128,
kernel_size=1,
),
nn.BatchNorm2d(128),
nn.ReLU(inplace=True),
nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2),
nn.BatchNorm2d(256),
nn.ReLU(inplace=True)
])
######
# dsfd specific layers
######
output_channels = [256, 512, 1024, 2048, 512, 256]
# fpn
fpn_in = output_channels
self.latlayer3 = nn.Conv2d(fpn_in[3],
fpn_in[2],
kernel_size=1,
stride=1,
padding=0)
self.latlayer2 = nn.Conv2d(fpn_in[2],
fpn_in[1],
kernel_size=1,
stride=1,
padding=0)
self.latlayer1 = nn.Conv2d(fpn_in[1],
fpn_in[0],
kernel_size=1,
stride=1,
padding=0)
self.smooth3 = nn.Conv2d(fpn_in[2],
fpn_in[2],
kernel_size=1,
stride=1,
padding=0)
self.smooth2 = nn.Conv2d(fpn_in[1],
fpn_in[1],
kernel_size=1,
stride=1,
padding=0)
self.smooth1 = nn.Conv2d(fpn_in[0],
fpn_in[0],
kernel_size=1,
stride=1,
padding=0)
self.upsample = nn.Upsample(scale_factor=2,
mode='bilinear',
align_corners=False)
self.eltmul = nn.EltMul()
# fem
cpm_in = output_channels
self.cpm3_3 = FEM(cpm_in[0])
self.cpm4_3 = FEM(cpm_in[1])
self.cpm5_3 = FEM(cpm_in[2])
self.cpm7 = FEM(cpm_in[3])
self.cpm6_2 = FEM(cpm_in[4])
self.cpm7_2 = FEM(cpm_in[5])
# pa
cfg_mbox = [1, 1, 1, 1, 1, 1]
head = pa_multibox(output_channels, cfg_mbox, self.num_classes)
# detection head
self.loc = nn.ModuleList(head[0])
self.conf = nn.ModuleList(head[1])
self.softmax = nn.Softmax(dim=-1)
from _utils import test_on
import sys
sys.path.append('.')
from flops_counter import nn
from flops_counter.tensorsize import TensorSize
######
# test on ReLU
######
relu = {
'layers': [
nn.ReLU() # same shape
],
'ins': [TensorSize([1, 64, 112, 112])],
'out_shape': [TensorSize([1, 64, 112, 112])],
'out_flops': [1605632]
}
test_on(relu)
######
# test on Sigmoid
######
sigmoid = {
'layers': [
nn.Sigmoid() # same shape
],
'ins': [TensorSize([1, 1, 56, 56])],
'out_shape': [TensorSize([1, 1, 56, 56])],
'out_flops': [9408]
