Loading detection/modeling/faster_rcnn.py +10 −1 Original line number Diff line number Diff line Loading @@ -18,6 +18,7 @@ class Dis(nn.Module): focal_loss_gamma=5, pool_type='avg', loss_weight=1.0, window_strides=None, window_sizes=(3, 9, 15, 21, -1)): super().__init__() # fmt:off Loading @@ -29,6 +30,11 @@ class Dis(nn.Module): self.num_windows = len(window_sizes) self.num_anchors = num_anchors self.window_sizes = window_sizes if window_strides is None: self.window_strides = [None] * len(window_sizes) else: assert len(window_strides) == len(window_sizes), 'window_strides and window_sizes should has same len' self.window_strides = window_strides if pool_type == 'avg': channel_multiply = 1 Loading Loading @@ -120,7 +126,10 @@ class Dis(nn.Module): if k == -1: x = self.pool_func(feature, kernel_size=(H, W)) else: x = self.pool_func(feature, kernel_size=k, stride=i + 1) stride = self.window_strides[i] if stride is None: stride = i + 1 # default x = self.pool_func(feature, kernel_size=k, stride=stride) _, _, h, w = x.shape semantic_map_per_level = F.interpolate(semantic_map, size=(h, w), mode='bilinear', align_corners=True) semantic_map_per_level = self.semantic_list[i](semantic_map_per_level) Loading Loading
detection/modeling/faster_rcnn.py +10 −1 Original line number Diff line number Diff line Loading @@ -18,6 +18,7 @@ class Dis(nn.Module): focal_loss_gamma=5, pool_type='avg', loss_weight=1.0, window_strides=None, window_sizes=(3, 9, 15, 21, -1)): super().__init__() # fmt:off Loading @@ -29,6 +30,11 @@ class Dis(nn.Module): self.num_windows = len(window_sizes) self.num_anchors = num_anchors self.window_sizes = window_sizes if window_strides is None: self.window_strides = [None] * len(window_sizes) else: assert len(window_strides) == len(window_sizes), 'window_strides and window_sizes should has same len' self.window_strides = window_strides if pool_type == 'avg': channel_multiply = 1 Loading Loading @@ -120,7 +126,10 @@ class Dis(nn.Module): if k == -1: x = self.pool_func(feature, kernel_size=(H, W)) else: x = self.pool_func(feature, kernel_size=k, stride=i + 1) stride = self.window_strides[i] if stride is None: stride = i + 1 # default x = self.pool_func(feature, kernel_size=k, stride=stride) _, _, h, w = x.shape semantic_map_per_level = F.interpolate(semantic_map, size=(h, w), mode='bilinear', align_corners=True) semantic_map_per_level = self.semantic_list[i](semantic_map_per_level) Loading