def __init__(self, dim):
super(CrossPooling, self).__init__()
self.p1_conv1 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.p2_conv1 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.p_conv1 = nn.Conv2D(dim, 3, 1, 1, use_bias=False, in_channels=128)
self.p_bn1 = nn.BatchNorm(in_channels=dim)
self.conv1 = nn.Conv2D(dim, 1, 1, 0, use_bias=False, in_channels=dim)
self.bn1 = nn.BatchNorm(in_channels=dim)
self.relu1 = nn.Activation(activation='relu')
self.conv2 = BasicConvBlock(dim, 3, 1, 1, use_bias=False, in_c=dim)
def __init__(self, dim, pool1, pool2, pool3, pool4):
super(pool_cross, self).__init__()
self.p1_conv1 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.p2_conv1 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.p_conv1 = nn.Conv2D(dim, 3, 1, 1, use_bias=False, in_channels=128)
self.p_bn1 = nn.BatchNorm(in_channels=dim)
self.conv1 = nn.Conv2D(dim, 1, 1, 0, use_bias=False, in_channels=dim)
self.bn1 = nn.BatchNorm(in_channels=dim)
self.relu1 = nn.Activation(activation='relu')
self.conv2 = BasicConvBlock(dim, 3, 1, 1, use_bias=False, in_c=dim)
self.pool1 = pool1
self.pool2 = pool2
self.pool3 = pool3
self.pool4 = pool4
def __init__(self, dim):
super(BottomRightPooling, self).__init__()
self.p1_conv1 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.p2_conv1 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.p_conv1 = nn.Conv2D(dim, 3, 1, 1, use_bias=False, in_channels=128)
self.p_bn1 = nn.BatchNorm(in_channels=dim)
self.conv1 = nn.Conv2D(dim, 1, 1, 0, use_bias=False, in_channels=dim)
self.bn1 = nn.BatchNorm(in_channels=dim)
self.relu1 = nn.Activation(activation='relu')
self.conv2 = BasicConvBlock(dim, 3, 1, 1, use_bias=False, in_c=dim)
# self.pool1 = pool1 # Bottom Pooling
# self.pool2 = pool2 # Right Pooling
self.look_conv1 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.look_conv2 = BasicConvBlock(128, 3, 1, 1, use_bias=False, in_c=dim)
self.P1_look_conv = nn.Conv2D(128, 3, 1, 1, use_bias=False, in_channels=128)
self.P2_look_conv = nn.Conv2D(128, 3, 1, 1, use_bias=False, in_channels=128)
def __init__(self, backbone, n_cls, **kwargs):
''' CenterNet
Backbone: Ghost
Neck : Stair Upsampling
Head : CenterNet (Cascade Corner Pooling, Center Pooling, Heatmap, Embeding map etc.)
@params backbone: str or gluon Modules
@params n_cls: class nums, e.g. 80 for coco dataset, 20 for voc
'''
super(CenterNetTri, self).__init__(**kwargs)
## backbone
with self.name_scope():
## Stem Block, downscale by 4 total
# TODO: add stem block here: BasicConvBlock(128, 7, 2, 3, use_bias=False, in_c=3) & ResConvBlock(256, 3, 2, 1, False, 128)
## Backbone
if isinstance(backbone, str):
self.backbone = model_zoo.get_model(backbone, **kwargs)
else:
self.backbone = backbone
## Neck
self.neck = StairUpsampling(cout=256, n_stages=3)
## Head
self.cnvs = BasicConvBlock(256, 3, 1, 1, False, 256, True, True)
self.tl_cnvs = TopLeftPooling(256) # cascade top left pooling
self.br_cnvs = BottomRightPooling(
256) # cascade bottom right pooling
self.ct_cnvs = CrossPooling(256) # center pooling
## Keypoint heatmaps
self.tl_heats = HeatmapConvBlock(n_cls, 256, 3, 1, 1, True, 256)
self.br_heats = HeatmapConvBlock(n_cls, 256, 3, 1, 1, True, 256)
self.ct_heats = HeatmapConvBlock(n_cls, 256, 3, 1, 1, True, 256)
## Tags, same structure as heatmaps except the output channel nums
self.tl_tags = TagConvBlock(1, 256, 3, 1, 1, True, 256)
self.br_tags = TagConvBlock(1, 256, 3, 1, 1, True, 256)
## Box Regress, same as tag & heatmap output layers, only different at the output channels
self.tl_regs = RegConvBlock(2, 256, 3, 1, 1, True, 256)
self.br_regs = RegConvBlock(2, 256, 3, 1, 1, True, 256)
self.ct_regs = RegConvBlock(2, 256, 3, 1, 1, True, 256)
## for training
self.transpose_gather_feats = TransposeGatherFeats()
def __init__(self, c_out=2, c_mid=256, kernel_size=3, strides=1, padding=1, use_bias=True, in_c=0, **kwargs):
super(RegConvBlock, self).__init__(**kwargs)
with self.name_scope():
self.conv0 = BasicConvBlock(c_mid, kernel_size, strides, padding, use_bias, in_c, False)
self.out = nn.Conv2D(c_out, 1, 1, 0, in_channels=c_mid)