Remove ROIAlign and ROIPooling compilation Requirement. Compatible with Pytorch 1.8.

model/rcnn_discriminator.py

@@ -1,11 +1,36 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from .roi_layers import ROIAlign, ROIPool
+#from .roi_layers import ROIAlign, ROIPool
+from torchvision.ops import roi_align
+from torchvision.ops import roi_pool
 from utils.util import *
 from utils.bilinear import *
 
 
+class ROIAlign(nn.Module):
+    def __init__(self, output_size, spatial_scale, sampling_ratio):
+        super(ROIAlign, self).__init__()
+        self.output_size = output_size
+        self.spatial_scale = spatial_scale
+        self.sampling_ratio = sampling_ratio
+
+    def forward(self, input, rois):
+        return roi_align(
+            input, rois, self.output_size, self.spatial_scale, self.sampling_ratio
+        )
+
+
+class ROIPool(nn.Module):
+    def __init__(self, output_size, spatial_scale):
+        super(ROIPool, self).__init__()
+        self.output_size = output_size
+        self.spatial_scale = spatial_scale
+
+    def forward(self, input, rois):
+        return roi_pool(input, rois, self.output_size, self.spatial_scale)
+
+
 def conv2d(in_feat, out_feat, kernel_size=3, stride=1, pad=1, spectral_norm=True):
     conv = nn.Conv2d(in_feat, out_feat, kernel_size, stride, pad)
     if spectral_norm:
@@ -58,8 +83,8 @@ def forward(self, x, y=None, bbox=None):
         # obj path
         # seperate different path
         s_idx = ((bbox[:, 3] - bbox[:, 1]) < 64) * ((bbox[:, 4] - bbox[:, 2]) < 64)
-        bbox_l, bbox_s = bbox[1-s_idx], bbox[s_idx]
-        y_l, y_s = y[1-s_idx], y[s_idx]
+        bbox_l, bbox_s = bbox[~s_idx], bbox[s_idx]
+        y_l, y_s = y[~s_idx], y[s_idx]
 
         obj_feat_s = self.block_obj3(x1)
         obj_feat_s = self.block_obj4(obj_feat_s)