def __init__(self, pretrained=False):
super(Detector, self).__init__()
self.t = 100
# ---------------------------
# TODO: Param
self.a_hw = [
[32.00, 32.00],
[40.32, 40.32],
[50.80, 50.80],
[22.63, 45.25],
[28.51, 57.02],
[35.92, 71.84],
[45.25, 22.63],
[57.02, 28.51],
[71.84, 35.92],
]
self.scales = 5
# self.scales = [0,0,1,1,2]
self.first_stride = 8
# self.view_size = 1025
self.view_size = 641
self.iou_th = (0.4, 0.5)
self.classes = 20
# self.classes = 80
self.nms_th = 0.05
self.nms_iou = 0.5
self.max_detections = 3000
self.balanced_fpn = False
# 统计了BIFPN中所有的结点 没有分开算
self.w_2_in = nn.Parameter(torch.Tensor(
2, 6).fill_(1)) # 6代表两个input的结点的个数，2代表每个结点分配的权重
self.w_3_in = nn.Parameter(torch.Tensor(
3, 2).fill_(1)) # 2代表两个input的结点的个数，3代表每个结点分配的权重
self.eps = 0.0001
self.w_relu = nn.ReLU()
# ---------------------------
# fpn =======================================================
self.backbone = backbone(pretrained=pretrained)
self.relu = nn.ReLU(inplace=True)
self.conv_out6 = nn.Conv2d(2048,
256,
kernel_size=3,
padding=1,
stride=2)
self.conv_out7 = nn.Conv2d(256,
256,
kernel_size=3,
padding=1,
stride=2)
self.prj_5 = nn.Conv2d(2048, 256, kernel_size=1)
self.prj_4 = nn.Conv2d(1024, 256, kernel_size=1)
self.prj_3 = nn.Conv2d(512, 256, kernel_size=1)
self.conv_5 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
self.conv_4 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
self.conv_3 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
# bifpn ======================================================
self.bi1_prj1_5 = nn.Conv2d(2048, 256, kernel_size=1)
self.bi1_prj1_5bn = nn.BatchNorm2d(256, momentum=0.01, eps=1e-3)
self.bi1_prj1_4 = nn.Conv2d(1024, 256, kernel_size=1)
self.bi1_prj1_4bn = nn.BatchNorm2d(256, momentum=0.01, eps=1e-3)
self.bi1_prj1_3 = nn.Conv2d(512, 256, kernel_size=1)
self.bi1_prj1_3bn = nn.BatchNorm2d(256, momentum=0.01, eps=1e-3)
self.bi1_prj2_5 = ConvBlock(256)
self.bi1_prj2_4 = ConvBlock(256)
self.bi1_prj2_3 = ConvBlock(256)
self.bi1_prj3_5 = ConvBlock(256)
self.bi1_prj3_4 = ConvBlock(256)
self.bi1_prj3_3 = ConvBlock(256)
self.bi2_prj1_5 = ConvBlock(256)
self.bi2_prj1_4 = ConvBlock(256)
self.bi2_prj1_3 = ConvBlock(256)
self.bi2_prj2_5 = ConvBlock(256)
self.bi2_prj2_4 = ConvBlock(256)
self.bi2_prj2_3 = ConvBlock(256)
self.bi2_prj3_5 = ConvBlock(256)
self.bi2_prj3_4 = ConvBlock(256)
self.bi2_prj3_3 = ConvBlock(256)
# -------------------------------------------
self.bi1_prj2_5 = DepthWiseConvBlock(256)
self.bi1_prj2_4 = DepthWiseConvBlock(256)
self.bi1_prj2_3 = DepthWiseConvBlock(256)
self.bi1_prj3_5 = DepthWiseConvBlock(256)
self.bi1_prj3_4 = DepthWiseConvBlock(256)
self.bi1_prj3_3 = DepthWiseConvBlock(256)
self.bi2_prj1_5 = DepthWiseConvBlock(256)
self.bi2_prj1_4 = DepthWiseConvBlock(256)
self.bi2_prj1_3 = DepthWiseConvBlock(256)
self.bi2_prj2_5 = DepthWiseConvBlock(256)
self.bi2_prj2_4 = DepthWiseConvBlock(256)
self.bi2_prj2_3 = DepthWiseConvBlock(256)
self.bi2_prj3_5 = DepthWiseConvBlock(256)
self.bi2_prj3_4 = DepthWiseConvBlock(256)
self.bi2_prj3_3 = DepthWiseConvBlock(256)
# head =======================================================
self.conv_cls = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256,
len(self.a_hw) * self.classes,
kernel_size=3,
padding=1))
self.conv_reg = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, len(self.a_hw) * 4, kernel_size=3, padding=1))
# reinit head =======================================================
for layer in self.conv_cls.children():
if isinstance(layer, nn.Conv2d):
nn.init.constant_(layer.bias, 0)
nn.init.normal_(layer.weight, mean=0, std=0.01)
for layer in self.conv_reg.children():
if isinstance(layer, nn.Conv2d):
nn.init.constant_(layer.bias, 0)
nn.init.normal_(layer.weight, mean=0, std=0.01)
pi = 0.01
_bias = -math.log((1.0 - pi) / pi)
nn.init.constant_(self.conv_cls[-1].bias, _bias)
# generate anchors =======================================================
self._view_anchors_yxyx, self._view_anchors_yxhw = \
gen_anchors(self.a_hw, self.scales, self.view_size, self.first_stride)
self.view_hwan = self._view_anchors_yxyx.shape[0]
self.register_buffer('view_anchors_yxyx', self._view_anchors_yxyx)
self.register_buffer('view_anchors_yx', self._view_anchors_yxhw[:, :2])
self.register_buffer('view_anchors_hw', self._view_anchors_yxhw[:, 2:])
def __init__(self, pretrained=False):
super(Detector, self).__init__()
# ---------------------------
# TODO: Param
self.a_hw = [
[32.00, 32.00],
[40.32, 40.32],
[50.80, 50.80],
[22.63, 45.25],
[28.51, 57.02],
[35.92, 71.84],
[45.25, 22.63],
[57.02, 28.51],
[71.84, 35.92],
]
self.scales = 5
self.first_stride = 8
self.view_size = 1025
self.iou_th = (0.4, 0.5)
self.classes = 80
self.nms_th = 0.05
self.nms_iou = 0.5
self.max_detections = 3000
self.balanced_fpn = False
# ---------------------------
# fpn =======================================================
self.backbone = backbone(pretrained=pretrained)
self.relu = nn.ReLU(inplace=True)
self.conv_out6 = nn.Conv2d(2048,
256,
kernel_size=3,
padding=1,
stride=2)
self.conv_out7 = nn.Conv2d(256,
256,
kernel_size=3,
padding=1,
stride=2)
self.prj_5 = nn.Conv2d(2048, 256, kernel_size=1)
self.prj_4 = nn.Conv2d(1024, 256, kernel_size=1)
self.prj_3 = nn.Conv2d(512, 256, kernel_size=1)
self.conv_5 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
self.conv_4 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
self.conv_3 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
# head =======================================================
self.conv_cls = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256,
len(self.a_hw) * self.classes,
kernel_size=3,
padding=1))
self.conv_reg = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, len(self.a_hw) * 4, kernel_size=3, padding=1))
# reinit head =======================================================
for layer in self.conv_cls.children():
if isinstance(layer, nn.Conv2d):
nn.init.constant_(layer.bias, 0)
nn.init.normal_(layer.weight, mean=0, std=0.01)
for layer in self.conv_reg.children():
if isinstance(layer, nn.Conv2d):
nn.init.constant_(layer.bias, 0)
nn.init.normal_(layer.weight, mean=0, std=0.01)
pi = 0.01
_bias = -math.log((1.0 - pi) / pi)
nn.init.constant_(self.conv_cls[-1].bias, _bias)
# generate anchors =======================================================
self._view_anchors_yxyx, self._view_anchors_yxhw = \
gen_anchors(self.a_hw, self.scales, self.view_size, self.first_stride)
self.view_hwan = self._view_anchors_yxyx.shape[0]
self.register_buffer('view_anchors_yxyx', self._view_anchors_yxyx)
self.register_buffer('view_anchors_yx', self._view_anchors_yxhw[:, :2])
self.register_buffer('view_anchors_hw', self._view_anchors_yxhw[:, 2:])
def __init__(self, pretrained=False):
super(Detector, self).__init__()
# ---------------------------
# TODO: Param
self.regions = [0, 64, 128, 256, 512, 9999]
self.first_stride = 8
self.view_size = 1025
self.classes = 80
self.nms_th = 0.05
self.nms_iou = 0.6
self.max_detections = 3000
self.center_offset_ratio = 1.5
# ---------------------------
# fpn =======================================================
self.backbone = Backbone(pretrained=pretrained)
self.relu = nn.ReLU(inplace=True)
self.conv_out6 = nn.Conv2d(2048,
256,
kernel_size=3,
padding=1,
stride=2)
self.conv_out7 = nn.Conv2d(256,
256,
kernel_size=3,
padding=1,
stride=2)
self.prj_5 = nn.Conv2d(2048, 256, kernel_size=1)
self.prj_4 = nn.Conv2d(1024, 256, kernel_size=1)
self.prj_3 = nn.Conv2d(512, 256, kernel_size=1)
self.conv_5 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
self.conv_4 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
self.conv_3 = nn.Conv2d(256, 256, kernel_size=3, padding=1)
# head =======================================================
self.conv_cls = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, self.classes, kernel_size=3, padding=1))
self.conv_reg = nn.Sequential(
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.Conv2d(256, 4, kernel_size=3, padding=1))
# reinit head =======================================================
for layer in self.conv_cls.children():
if isinstance(layer, nn.Conv2d):
nn.init.constant_(layer.bias, 0)
nn.init.normal_(layer.weight, mean=0, std=0.01)
for layer in self.conv_reg.children():
if isinstance(layer, nn.Conv2d):
nn.init.constant_(layer.bias, 0)
nn.init.normal_(layer.weight, mean=0, std=0.01)
pi = 0.01
_bias = -math.log((1.0 - pi) / pi)
nn.init.constant_(self.conv_cls[-1].bias, _bias)
# learnable parameter for scales =====================================
self.scale_div = nn.Parameter(torch.ones(len(self.regions) - 1))
# generate anchors ===================================================
self._a_center_yx, self._a_tlbr_max_minmax, self._a_center_offset_max = \
gen_anchors(self.view_size, self.first_stride, self.regions, self.center_offset_ratio)
self.a_shw = self._a_center_yx.shape[0]
self.register_buffer('a_center_yx', self._a_center_yx)
self.register_buffer('a_tlbr_max_minmax', self._a_tlbr_max_minmax)
self.register_buffer('a_center_offset_max', self._a_center_offset_max)