Avoid circular dependencies on import

mmdet/ops/dcn/functions/deform_conv.py

@@ -2,8 +2,6 @@
 from torch.autograd import Function
 from torch.nn.modules.utils import _pair
 
-from .. import deform_conv_cuda
-
 
 class DeformConvFunction(Function):
 
@@ -18,6 +16,8 @@ def forward(ctx,
                 groups=1,
                 deformable_groups=1,
                 im2col_step=64):
+        from .. import deform_conv_cuda
+
         if input is not None and input.dim() != 4:
             raise ValueError(
                 "Expected 4D tensor as input, got {}D tensor instead.".format(

mmdet/ops/dcn/functions/deform_pool.py

@@ -1,8 +1,6 @@
 import torch
 from torch.autograd import Function
 
-from .. import deform_pool_cuda
-
 
 class DeformRoIPoolingFunction(Function):
 
@@ -19,6 +17,8 @@ def forward(ctx,
                 part_size=None,
                 sample_per_part=4,
                 trans_std=.0):
+        from .. import deform_pool_cuda
+
         ctx.spatial_scale = spatial_scale
         ctx.out_size = out_size
         ctx.out_channels = out_channels
@@ -48,6 +48,8 @@ def forward(ctx,
 
     @staticmethod
     def backward(ctx, grad_output):
+        from .. import deform_pool_cuda
+
         if not grad_output.is_cuda:
             raise NotImplementedError
 

mmdet/ops/dcn/modules/deform_conv.py

@@ -4,8 +4,6 @@
 import torch.nn as nn
 from torch.nn.modules.utils import _pair
 
-from ..functions.deform_conv import deform_conv, modulated_deform_conv
-
 
 class DeformConv(nn.Module):
 
@@ -52,6 +50,8 @@ def reset_parameters(self):
         self.weight.data.uniform_(-stdv, stdv)
 
     def forward(self, x, offset):
+        from ..functions.deform_conv import deform_conv
+
         return deform_conv(x, offset, self.weight, self.stride, self.padding,
                            self.dilation, self.groups, self.deformable_groups)
 
@@ -76,6 +76,8 @@ def init_offset(self):
         self.conv_offset.bias.data.zero_()
 
     def forward(self, x):
+        from ..functions.deform_conv import deform_conv
+
         offset = self.conv_offset(x)
         return deform_conv(x, offset, self.weight, self.stride, self.padding,
                            self.dilation, self.groups, self.deformable_groups)
@@ -123,6 +125,8 @@ def reset_parameters(self):
             self.bias.data.zero_()
 
     def forward(self, x, offset, mask):
+        from ..functions.deform_conv import modulated_deform_conv
+
         return modulated_deform_conv(x, offset, mask, self.weight, self.bias,
                                      self.stride, self.padding, self.dilation,
                                      self.groups, self.deformable_groups)
@@ -148,6 +152,8 @@ def init_offset(self):
         self.conv_offset_mask.bias.data.zero_()
 
     def forward(self, x):
+        from ..functions.deform_conv import modulated_deform_conv
+
         out = self.conv_offset_mask(x)
         o1, o2, mask = torch.chunk(out, 3, dim=1)
         offset = torch.cat((o1, o2), dim=1)

mmdet/ops/masked_conv/functions/masked_conv.py

@@ -2,13 +2,14 @@
 import torch
 from torch.autograd import Function
 from torch.nn.modules.utils import _pair
-from .. import masked_conv2d_cuda
 
 
 class MaskedConv2dFunction(Function):
 
     @staticmethod
     def forward(ctx, features, mask, weight, bias, padding=0, stride=1):
+        from .. import masked_conv2d_cuda
+
         assert mask.dim() == 3 and mask.size(0) == 1
         assert features.dim() == 4 and features.size(0) == 1
         assert features.size()[2:] == mask.size()[1:]

mmdet/ops/nms/nms_wrapper.py

@@ -2,7 +2,6 @@
 import torch
 
 from . import nms_cuda, nms_cpu
-from .soft_nms_cpu import soft_nms_cpu
 
 
 def nms(dets, iou_thr, device_id=None):
@@ -50,6 +49,8 @@ def nms(dets, iou_thr, device_id=None):
 
 
 def soft_nms(dets, iou_thr, method='linear', sigma=0.5, min_score=1e-3):
+    from .soft_nms_cpu import soft_nms_cpu
+
     if isinstance(dets, torch.Tensor):
         is_tensor = True
         dets_np = dets.detach().cpu().numpy()

mmdet/ops/roi_align/functions/roi_align.py

@@ -1,12 +1,12 @@
 from torch.autograd import Function
 
-from .. import roi_align_cuda
-
 
 class RoIAlignFunction(Function):
 
     @staticmethod
     def forward(ctx, features, rois, out_size, spatial_scale, sample_num=0):
+        from .. import roi_align_cuda
+
         if isinstance(out_size, int):
             out_h = out_size
             out_w = out_size
@@ -37,6 +37,8 @@ def forward(ctx, features, rois, out_size, spatial_scale, sample_num=0):
 
     @staticmethod
     def backward(ctx, grad_output):
+        from .. import roi_align_cuda
+
         feature_size = ctx.feature_size
         spatial_scale = ctx.spatial_scale
         sample_num = ctx.sample_num

mmdet/ops/roi_pool/functions/roi_pool.py

@@ -1,13 +1,13 @@
 import torch
 from torch.autograd import Function
 
-from .. import roi_pool_cuda
-
 
 class RoIPoolFunction(Function):
 
     @staticmethod
     def forward(ctx, features, rois, out_size, spatial_scale):
+        from .. import roi_pool_cuda
+
         if isinstance(out_size, int):
             out_h = out_size
             out_w = out_size
@@ -36,6 +36,8 @@ def forward(ctx, features, rois, out_size, spatial_scale):
 
     @staticmethod
     def backward(ctx, grad_output):
+        from .. import roi_pool_cuda
+
         assert grad_output.is_cuda
         spatial_scale = ctx.spatial_scale
         feature_size = ctx.feature_size

mmdet/ops/sigmoid_focal_loss/functions/sigmoid_focal_loss.py

@@ -1,13 +1,13 @@
 from torch.autograd import Function
 from torch.autograd.function import once_differentiable
 
-from .. import sigmoid_focal_loss_cuda
-
 
 class SigmoidFocalLossFunction(Function):
 
     @staticmethod
     def forward(ctx, input, target, gamma=2.0, alpha=0.25):
+        from .. import sigmoid_focal_loss_cuda
+
         ctx.save_for_backward(input, target)
         num_classes = input.shape[1]
         ctx.num_classes = num_classes
@@ -21,6 +21,8 @@ def forward(ctx, input, target, gamma=2.0, alpha=0.25):
     @staticmethod
     @once_differentiable
     def backward(ctx, d_loss):
+        from .. import sigmoid_focal_loss_cuda
+
         input, target = ctx.saved_tensors
         num_classes = ctx.num_classes
         gamma = ctx.gamma