Merge pull request ultralytics#16 from Laughing-q/instance_seg

clean up

models/yolo.py

@@ -14,8 +14,6 @@
 from copy import deepcopy
 from pathlib import Path
 
-from torch import NoneType
-
 FILE = Path(__file__).resolve()
 ROOT = FILE.parents[1]  # YOLOv5 root directory
 if str(ROOT) not in sys.path:
@@ -30,7 +28,6 @@
 from utils.plots import feature_visualization
 from utils.torch_utils import (fuse_conv_and_bn, initialize_weights, model_info, profile, scale_img, select_device,
                                time_sync)
-import torch.nn.functional as F
 
 try:
     import thop  # for FLOPs computation
@@ -110,8 +107,7 @@ def __init__(self, nc=80, anchors=(), mask_dim=32, proto_channel=256, ch=(), inp
             # nn.SiLU(inplace=True),
             # nn.Conv2d(self.proto_c, self.proto_c, kernel_size=3, stride=1, padding=1),
             # nn.SiLU(inplace=True), 
-            # nn.Upsample(scale_factor=2, mode='nearest'),
-            Upsample(scale_factor=2, mode='bilinear', align_corners=False),
+            nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
             nn.Conv2d(self.proto_c, self.proto_c, kernel_size=3, stride=1, padding=1),
             nn.SiLU(inplace=True),
             nn.Conv2d(self.proto_c, self.mask_dim, kernel_size=1, padding=0),
@@ -279,7 +275,10 @@ def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is
         for mi, s in zip(m.m, m.stride):  # from
             b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
             b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
-            b.data[:, 5+m.mask_dim:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            if hasattr(m, "mask_dim"):
+                b.data[:, 5+m.mask_dim:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
+            else:
+                b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
             mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
 
     def _print_biases(self):
@@ -379,18 +378,6 @@ def parse_model(d, ch):  # model_dict, input_channels(3)
         ch.append(c2)
     return nn.Sequential(*layers), sorted(save)
 
-class Upsample(nn.Module):
-    '''
-    deterministic upsample layer
-    '''
-    def __init__(self, scale_factor, mode="bilinear", align_corners=False) -> None:
-        super().__init__()
-        self.scale_factor = scale_factor
-        self.mode = mode
-        self.align_corners = align_corners
-
-    def forward(self, x):
-        return F.interpolate(x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners)
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
@@ -424,4 +411,4 @@ def forward(self, x):
                 print(f'Error in {cfg}: {e}')
 
     else:  # report fused model summary
-        model.fuse()
+        model.fuse()

requirements.txt

@@ -12,6 +12,7 @@ scipy>=1.4.1
 torch>=1.7.0
 torchvision>=0.8.1
 tqdm>=4.64.0
+easydict>=1.9
 protobuf<=3.20.1  # https://github.com/ultralytics/yolov5/issues/8012
 
 # Logging -------------------------------------