def __init__(self, cfg, in_channels, out_channels):
super(CropRoi, self).__init__()
self.num_scales = len(cfg.rpn_scales)
self.scales = cfg.rpn_scales
self.crop_size = cfg.crop_size
self.convs = nn.ModuleList()
for l in range(self.num_scales):
if self.scales[l] == 1:
self.convs.append(
nn.Sequential(
nn.Conv2d(in_channels,
out_channels // self.num_scales,
kernel_size=3,
padding=1), ))
else:
self.convs.append(
nn.Sequential(
nn.Conv2d(in_channels,
out_channels // self.num_scales,
kernel_size=3,
padding=1),
nn.Upsample(scale_factor=self.scales[l],
mode='bilinear'),
))
self.crop = Crop(self.crop_size, self.crop_size, 1)
def __init__(self, cfg, crop_size):
super(CropRoi_1layer, self).__init__()
self.num_scales = len(cfg.rpn_scales)
self.crop_size = crop_size
self.sizes = cfg.rpn_base_sizes
self.scales = cfg.rpn_scales
self.crops = Crop(self.crop_size, self.crop_size, 1 / self.scales[0])
def __init__(self, cfg):
super(CropRoi, self).__init__()
self.num_heads = cfg.rpn_num_heads
self.pool_size = cfg.pool_size
self.size_thresholds = cfg.rcnn_select_size_thresholds
self.crops = nn.ModuleList()
for i in range(cfg.rpn_num_heads):
self.crops.append(
Crop(cfg.pool_size, cfg.pool_size, 1 / cfg.rpn_strides[i]))
def __init__(self, cfg, crop_size):
super(CropRoi, self).__init__()
self.num_scales = len(cfg.rpn_scales)
self.crop_size = crop_size
self.sizes = cfg.rpn_base_sizes
self.scales = cfg.rpn_scales
self.crops = nn.ModuleList()
for l in range(self.num_scales):
self.crops.append(Crop(self.crop_size, self.crop_size, 1 / self.scales[l]))
def __init__(self, cfg, crop_size):
super(CropRoi, self).__init__()
self.num_scales = len(cfg.rpn_scales)
self.crop_size = crop_size
self.sizes = cfg.rpn_base_sizes
self.sizes = torch.from_numpy(np.array(self.sizes, np.float32))
self.sizes = self.sizes.cuda() if USE_CUDA else self.sizes
self.scales = cfg.rpn_scales
self.crops = nn.ModuleList()
for l in range(self.num_scales):
self.crops.append(
Crop(self.crop_size, self.crop_size, 1 / self.scales[l])
)