[Fix] Fix binary segmentation when num_classes==1 (open-mmlab#2016)

* fix binary

* add ut

* fix ut

* restore metric computation

* remove metric ut update

* set out_channels by num_classes

* replace num_classes in encoder_decoder

* update props setting and fix ut

* update ut

* minor change

* update warning

Author: xiexinch

mmseg/models/decode_heads/decode_head.py

@@ -1,4 +1,5 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+import warnings
 from abc import ABCMeta, abstractmethod
 
 import torch
@@ -18,6 +19,9 @@ class BaseDecodeHead(BaseModule, metaclass=ABCMeta):
         in_channels (int|Sequence[int]): Input channels.
         channels (int): Channels after modules, before conv_seg.
         num_classes (int): Number of classes.
+        out_channels (int): Output channels of conv_seg.
+        threshold (float): Threshold for binary segmentation in the case of
+            `num_classes==1`. Default: None.
         dropout_ratio (float): Ratio of dropout layer. Default: 0.1.
         conv_cfg (dict|None): Config of conv layers. Default: None.
         norm_cfg (dict|None): Config of norm layers. Default: None.
@@ -56,6 +60,8 @@ def __init__(self,
                  channels,
                  *,
                  num_classes,
+                 out_channels=None,
+                 threshold=None,
                  dropout_ratio=0.1,
                  conv_cfg=None,
                  norm_cfg=None,
@@ -74,7 +80,6 @@ def __init__(self,
         super(BaseDecodeHead, self).__init__(init_cfg)
         self._init_inputs(in_channels, in_index, input_transform)
         self.channels = channels
-        self.num_classes = num_classes
         self.dropout_ratio = dropout_ratio
         self.conv_cfg = conv_cfg
         self.norm_cfg = norm_cfg
@@ -84,6 +89,30 @@ def __init__(self,
         self.ignore_index = ignore_index
         self.align_corners = align_corners
 
+        if out_channels is None:
+            if num_classes == 2:
+                warnings.warn('For binary segmentation, we suggest using'
+                              '`out_channels = 1` to define the output'
+                              'channels of segmentor, and use `threshold`'
+                              'to convert seg_logist into a prediction'
+                              'applying a threshold')
+            out_channels = num_classes
+
+        if out_channels != num_classes and out_channels != 1:
+            raise ValueError(
+                'out_channels should be equal to num_classes,'
+                'except binary segmentation set out_channels == 1 and'
+                f'num_classes == 2, but got out_channels={out_channels}'
+                f'and num_classes={num_classes}')
+
+        if out_channels == 1 and threshold is None:
+            threshold = 0.3
+            warnings.warn('threshold is not defined for binary, and defaults'
+                          'to 0.3')
+        self.num_classes = num_classes
+        self.out_channels = out_channels
+        self.threshold = threshold
+
         if isinstance(loss_decode, dict):
             self.loss_decode = build_loss(loss_decode)
         elif isinstance(loss_decode, (list, tuple)):
@@ -99,7 +128,7 @@ def __init__(self,
         else:
             self.sampler = None
 
-        self.conv_seg = nn.Conv2d(channels, num_classes, kernel_size=1)
+        self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
         if dropout_ratio > 0:
             self.dropout = nn.Dropout2d(dropout_ratio)
         else:

mmseg/models/segmentors/cascade_encoder_decoder.py

@@ -47,6 +47,7 @@ def _init_decode_head(self, decode_head):
             self.decode_head.append(builder.build_head(decode_head[i]))
         self.align_corners = self.decode_head[-1].align_corners
         self.num_classes = self.decode_head[-1].num_classes
+        self.out_channels = self.decode_head[-1].out_channels
 
     def encode_decode(self, img, img_metas):
         """Encode images with backbone and decode into a semantic segmentation

mmseg/models/segmentors/encoder_decoder.py

@@ -49,6 +49,7 @@ def _init_decode_head(self, decode_head):
         self.decode_head = builder.build_head(decode_head)
         self.align_corners = self.decode_head.align_corners
         self.num_classes = self.decode_head.num_classes
+        self.out_channels = self.decode_head.out_channels
 
     def _init_auxiliary_head(self, auxiliary_head):
         """Initialize ``auxiliary_head``"""
@@ -162,10 +163,10 @@ def slide_inference(self, img, img_meta, rescale):
         h_stride, w_stride = self.test_cfg.stride
         h_crop, w_crop = self.test_cfg.crop_size
         batch_size, _, h_img, w_img = img.size()
-        num_classes = self.num_classes
+        out_channels = self.out_channels
         h_grids = max(h_img - h_crop + h_stride - 1, 0) // h_stride + 1
         w_grids = max(w_img - w_crop + w_stride - 1, 0) // w_stride + 1
-        preds = img.new_zeros((batch_size, num_classes, h_img, w_img))
+        preds = img.new_zeros((batch_size, out_channels, h_img, w_img))
         count_mat = img.new_zeros((batch_size, 1, h_img, w_img))
         for h_idx in range(h_grids):
             for w_idx in range(w_grids):
@@ -245,7 +246,10 @@ def inference(self, img, img_meta, rescale):
             seg_logit = self.slide_inference(img, img_meta, rescale)
         else:
             seg_logit = self.whole_inference(img, img_meta, rescale)
-        output = F.softmax(seg_logit, dim=1)
+        if self.out_channels == 1:
+            output = F.sigmoid(seg_logit)
+        else:
+            output = F.softmax(seg_logit, dim=1)
         flip = img_meta[0]['flip']
         if flip:
             flip_direction = img_meta[0]['flip_direction']
@@ -260,7 +264,11 @@ def inference(self, img, img_meta, rescale):
     def simple_test(self, img, img_meta, rescale=True):
         """Simple test with single image."""
         seg_logit = self.inference(img, img_meta, rescale)
-        seg_pred = seg_logit.argmax(dim=1)
+        if self.out_channels == 1:
+            seg_pred = (seg_logit >
+                        self.decode_head.threshold).to(seg_logit).squeeze(1)
+        else:
+            seg_pred = seg_logit.argmax(dim=1)
         if torch.onnx.is_in_onnx_export():
             # our inference backend only support 4D output
             seg_pred = seg_pred.unsqueeze(0)
@@ -283,7 +291,11 @@ def aug_test(self, imgs, img_metas, rescale=True):
             cur_seg_logit = self.inference(imgs[i], img_metas[i], rescale)
             seg_logit += cur_seg_logit
         seg_logit /= len(imgs)
-        seg_pred = seg_logit.argmax(dim=1)
+        if self.out_channels == 1:
+            seg_pred = (seg_logit >
+                        self.decode_head.threshold).to(seg_logit).squeeze(1)
+        else:
+            seg_pred = seg_logit.argmax(dim=1)
         seg_pred = seg_pred.cpu().numpy()
         # unravel batch dim
         seg_pred = list(seg_pred)

tests/test_models/test_heads/test_decode_head.py

@@ -43,6 +43,18 @@ def test_decode_head():
                        in_index=[-1],
                        input_transform='resize_concat')
 
+    with pytest.raises(ValueError):
+        # out_channels should be equal to num_classes
+        BaseDecodeHead(32, 16, num_classes=19, out_channels=18)
+
+    # test out_channels
+    head = BaseDecodeHead(32, 16, num_classes=2)
+    assert head.out_channels == 2
+
+    # test out_channels == 1 and num_classes == 2
+    head = BaseDecodeHead(32, 16, num_classes=2, out_channels=1)
+    assert head.out_channels == 1 and head.num_classes == 2
+
     # test default dropout
     head = BaseDecodeHead(32, 16, num_classes=19)
     assert hasattr(head, 'dropout') and head.dropout.p == 0.1

tests/test_models/test_segmentors/test_encoder_decoder.py

@@ -18,6 +18,12 @@ def test_encoder_decoder():
     segmentor = build_segmentor(cfg)
     _segmentor_forward_train_test(segmentor)
 
+    # test out_channels == 1
+    segmentor.out_channels = 1
+    segmentor.decode_head.out_channels = 1
+    segmentor.decode_head.threshold = 0.3
+    _segmentor_forward_train_test(segmentor)
+
     # test slide mode
     cfg.test_cfg = ConfigDict(mode='slide', crop_size=(3, 3), stride=(2, 2))
     segmentor = build_segmentor(cfg)