def __init__(self, cfg, phase, size, num_classes):
super(SSD, self).__init__()
self.phase = phase
self.num_classes = num_classes
self.size = size
self.prior_layer = PriorLayer(cfg)
self.priorbox = PriorBox(cfg)
self.priors = self.priorbox.forward()
self.extractor = SSDarknet53(size)
self.in_channels = cfg['in_channels_darknet']
self.num_anchors = cfg['num_anchors']
self.input_fixed = cfg['input_fixed']
self.loc = nn.ModuleList()
self.conf = nn.ModuleList()
for i in range(len(self.in_channels)):
self.loc += [
nn.Conv2d(self.in_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.conf += [
nn.Conv2d(self.in_channels[i],
self.num_anchors[i] * self.num_classes,
kernel_size=3,
padding=1)
]
def __init__(self, cfg, phase, num_classes, size, channel_size='48'):
super(SSD, self).__init__()
self.phase = phase
self.num_classes = num_classes
self.arm_num_classes = 2
self.size = size
self.prior_layer = PriorLayer(cfg)
self.priorbox = PriorBox(cfg)
self.priors = self.priorbox.forward()
self.extractor = VGG16Extractor(size, channel_size)
self.arm_channels = cfg["arm_channels"]
self.odm_channels = cfg["odm_channels"]
self.num_anchors = cfg["num_anchors"]
self.input_fixed = cfg['input_fixed']
self.loc = nn.ModuleList()
self.conf = nn.ModuleList()
self.arm_loc = nn.ModuleList()
self.arm_conf = nn.ModuleList()
for i in range(len(self.arm_channels)):
self.loc += [
nn.Conv2d(self.odm_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.arm_loc += [
nn.Conv2d(self.arm_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.conf += [
nn.Conv2d(self.odm_channels[i],
self.num_anchors[i] * self.num_classes,
kernel_size=3,
padding=1)
]
self.arm_conf += [
nn.Conv2d(self.arm_channels[i],
self.num_anchors[i] * self.arm_num_classes,
kernel_size=3,
padding=1)
]
def __init__(self, cfg):
super(SSD, self).__init__()
self.cfg = cfg
self.size = cfg.MODEL.SIZE
if self.size == '300':
size_cfg = cfg.SMALL
else:
size_cfg = cfg.BIG
self.num_classes = cfg.MODEL.NUM_CLASSES
self.prior_layer = PriorLayer(cfg)
self.priorbox = PriorBox(cfg)
self.priors = self.priorbox.forward()
self.extractor = get_func(cfg.MODEL.CONV_BODY)(self.size,
cfg.TRAIN.CHANNEL_SIZE)
if cfg.MODEL.REFINE:
self.odm_channels = size_cfg.ODM_CHANNELS
self.arm_num_classes = 2
self.odm_loc = nn.ModuleList()
self.odm_conf = nn.ModuleList()
self.loc_offset_conv = nn.ModuleList()
self.dcn_convs = nn.ModuleList(add_dcn_dilas())
self.bn_layers = nn.ModuleList(BN_layers())
self.arm_loc = nn.ModuleList()
self.arm_conf = nn.ModuleList()
self.arm_channels = size_cfg.ARM_CHANNELS
self.num_anchors = size_cfg.NUM_ANCHORS
self.input_fixed = size_cfg.INPUT_FIXED
self.arm_loc = nn.ModuleList()
self.arm_conf = nn.ModuleList()
for i in range(len(self.arm_channels)):
if cfg.MODEL.REFINE:
self.arm_loc += [
nn.Conv2d(
self.arm_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.arm_conf += [
nn.Conv2d(
self.arm_channels[i],
self.num_anchors[i] * self.arm_num_classes,
kernel_size=3,
padding=1)
]
self.loc_offset_conv +=[BasicConv(self.num_anchors[i] * 2, 18, kernel_size=1)]
self.odm_loc += [
nn.Conv2d(
self.odm_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.odm_conf += [
nn.Conv2d(
self.odm_channels[i],
self.num_anchors[i] * self.num_classes,
kernel_size=3,
padding=1)
]
else:
self.arm_loc += [
nn.Conv2d(
self.arm_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.arm_conf += [
nn.Conv2d(
self.arm_channels[i],
self.num_anchors[i] * self.num_classes,
kernel_size=3,
padding=1)
]
if cfg.TRAIN.TRAIN_ON:
self._init_modules()
def __init__(self, cfg):
super(SSD, self).__init__()
self.cfg = cfg
self.size = cfg.MODEL.SIZE
if self.size == '300':
size_cfg = cfg.SMALL
else:
size_cfg = cfg.BIG
self.num_classes = cfg.MODEL.NUM_CLASSES
self.prior_layer = PriorLayer(cfg)
self.priorbox = PriorBox(cfg)
self.priors = self.priorbox.forward()
self.extractor = get_func(cfg.MODEL.CONV_BODY)(self.size)
if cfg.MODEL.CASCADE:
self.odm_channels = size_cfg.ODM_CHANNELS
self.arm_num_classes = 2
self.arm_loc = nn.ModuleList()
self.arm_conf = nn.ModuleList()
self.arm_channels = size_cfg.ARM_CHANNELS
self.num_anchors = size_cfg.NUM_ANCHORS
self.input_fixed = size_cfg.INPUT_FIXED
self.arm_loc = nn.ModuleList()
self.arm_conf = nn.ModuleList()
self.deform_conv_c = nn.ModuleList()
self.deform_conv_l1 = nn.ModuleList()
self.deform_conv_c1 = nn.ModuleList()
self.offset = nn.ModuleList()
self.offset1 = nn.ModuleList()
self.offlat1 = nn.ModuleList()
self.offlat = nn.ModuleList()
for i in range(len(self.arm_channels)):
if cfg.MODEL.CASCADE:
self.arm_loc += [
nn.Conv2d(self.arm_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.arm_conf += [
nn.Conv2d(self.arm_channels[i],
self.num_anchors[i] * self.arm_num_classes,
kernel_size=3,
padding=1)
]
self.deform_conv_c += [
ConvOffset2d(256 * 3,
self.num_anchors[i] * self.arm_num_classes,
3,
stride=1,
padding=1,
num_deformable_groups=3)
]
self.deform_conv_c1 += [
ConvOffset2d(256 * 3,
self.num_anchors[i] * self.num_classes,
3,
stride=1,
padding=1,
num_deformable_groups=3)
]
self.deform_conv_l1 += [
ConvOffset2d(256 * 3,
self.num_anchors[i] * 4,
3,
stride=1,
padding=1,
num_deformable_groups=3)
]
self.offset += [
nn.Conv2d(36,
2 * 3 * 3 * 3,
kernel_size=1,
stride=1,
padding=0,
bias=False,
groups=3)
]
self.offset1 += [
nn.Conv2d(36,
2 * 3 * 3 * 3,
kernel_size=1,
stride=1,
padding=0,
bias=False,
groups=3)
]
self.offlat += [
nn.Sequential(
nn.Conv2d(12,
36,
kernel_size=1,
stride=1,
padding=0,
groups=3), nn.ReLU())
]
self.offlat1 += [
nn.Sequential(
nn.Conv2d(12,
36,
kernel_size=1,
stride=1,
padding=0,
groups=3), nn.ReLU())
]
else:
self.arm_loc += [
nn.Conv2d(self.arm_channels[i],
self.num_anchors[i] * 4,
kernel_size=3,
padding=1)
]
self.arm_conf += [
nn.Conv2d(self.arm_channels[i],
self.num_anchors[i] * self.num_classes,
kernel_size=3,
padding=1)
]
self.deform_conv_c += [
ConvOffset2d(256 * 3,
self.num_anchors[i] * self.num_classes,
3,
stride=1,
padding=1,
num_deformable_groups=3)
]
self.offset += [
nn.Conv2d(36,
2 * 3 * 3 * 3,
kernel_size=1,
stride=1,
padding=0,
bias=False,
groups=3)
]
self.offlat += [
nn.Sequential(
nn.Conv2d(12,
36,
kernel_size=1,
stride=1,
padding=0,
groups=3), nn.ReLU())
]
if cfg.TRAIN.TRAIN_ON:
self._init_modules()