From ac213d3bf1481474e4fcf5ec091f97ddef396bb4 Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Tue, 18 Oct 2022 16:32:04 +0800
Subject: [PATCH 01/11] [Doc] Update FAQ doc about binary segmentation and
 ReduceZeroLabel

---
 docs/en/faq.md    | 80 +++++++++++++++++++++++++++++++++++++++++++++++
 docs/zh_cn/faq.md | 80 +++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 160 insertions(+)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index 5151f1fd36..09c69b915f 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -66,3 +66,83 @@ In the test script, we provide `show-dir` argument to control whether output the
 ```shell
 python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --opacity 1
 ```
+
+## Why is the IoU always 0, NaN or very low in binary segmentation task
+
+Sometimes when training customized dataset, the IoU of certain class is 0, NaN or very low like below:
+
+```
++--------------+-------+-------+
+|    Class     |  IoU  |  Acc  |
++--------------+-------+-------+
+| label_a | 80.19 | 100.0 |
+| label_b  |  nan  |  nan  |
++--------------+-------+-------+
+2022-10-18 10:56:56,032 - mmseg - INFO - Summary:
+2022-10-18 10:56:56,032 - mmseg - INFO -
++-------+-------+-------+
+|  aAcc |  mIoU |  mAcc |
++-------+-------+-------+
+| 100.0 | 80.19 | 100.0 |
++-------+-------+-------+
+```
+
+or
+
+```
++------------+------+-------+
+|   Class    | IoU  |  Acc  |
++------------+------+-------+
+| label_a | 0.0  |  0.0  |
+|   label_b   | 1.77 | 100.0 |
++------------+------+-------+
+2022-10-18 00:57:12,082 - mmseg - INFO - Summary:
+2022-10-18 00:57:12,083 - mmseg - INFO -
++------+------+------+
+| aAcc | mIoU | mAcc |
++------+------+------+
+| 1.77 | 0.88 | 50.0 |
++------+------+------+
+```
+
+- Solution One: You can follow our config file of dataset [`DRIVE`](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#drive) for reference, whose [dataset class](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/drive.py) is like below:
+
+```python
+class DRIVEDataset(CustomDataset):
+    CLASSES = ('background', 'vessel')
+
+    PALETTE = [[120, 120, 120], [6, 230, 230]]
+
+    def __init__(self, **kwargs):
+        super(DRIVEDataset, self).__init__(
+            img_suffix='.png',
+            seg_map_suffix='_manual1.png',
+            reduce_zero_label=False,
+            **kwargs)
+        assert self.file_client.exists(self.img_dir)
+```
+
+And in corresponding config files of [dataset](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/datasets/drive.py) and [model](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/fcn_unet_s5-d16.py#L23-L48):
+
+```python
+xxx_head=dict(
+    num_classes=2,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0))
+```
+
+- Solution Two: In [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), we fix the binary segmentation task when `num_classes=1`. You can follow this [#2201](https://github.com/open-mmlab/mmsegmentation/issues/2201) by setting `num_classes=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
+
+## What does `reduce_zero_label` work for?
+
+When [loading annotation](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) in MMSegmentation, `reduce_zero_label (bool)` is provided to determine whether reduce all label value by 1:
+
+```python
+if self.reduce_zero_label:
+    # avoid using underflow conversion
+    gt_semantic_seg[gt_semantic_seg == 0] = 255
+    gt_semantic_seg = gt_semantic_seg - 1
+    gt_semantic_seg[gt_semantic_seg == 254] = 255
+```
+
+`reduce_zero_label` is usually used for datasets where 0 is background label, if `reduce_zero_label=True`, the pixels whose corresponding label is 0 would not be involved in loss calculation.
diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index 9a55b8e3c8..e71fdb2f38 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -66,3 +66,83 @@
 ```shell
 python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --opacity 1
 ```
+
+## 为什么在二值分割任务里 IoU 总是 0, NaN 或者非常低?
+
+有时候在训练自定义的数据集时, 如下所示, 某个类别的 IoU 总是 0, NaN 或者很低:
+
+```
++--------------+-------+-------+
+|    Class     |  IoU  |  Acc  |
++--------------+-------+-------+
+| label_a | 80.19 | 100.0 |
+| label_b  |  nan  |  nan  |
++--------------+-------+-------+
+2022-10-18 10:56:56,032 - mmseg - INFO - Summary:
+2022-10-18 10:56:56,032 - mmseg - INFO -
++-------+-------+-------+
+|  aAcc |  mIoU |  mAcc |
++-------+-------+-------+
+| 100.0 | 80.19 | 100.0 |
++-------+-------+-------+
+```
+
+或者
+
+```
++------------+------+-------+
+|   Class    | IoU  |  Acc  |
++------------+------+-------+
+| label_a | 0.0  |  0.0  |
+|   label_b   | 1.77 | 100.0 |
++------------+------+-------+
+2022-10-18 00:57:12,082 - mmseg - INFO - Summary:
+2022-10-18 00:57:12,083 - mmseg - INFO -
++------+------+------+
+| aAcc | mIoU | mAcc |
++------+------+------+
+| 1.77 | 0.88 | 50.0 |
++------+------+------+
+```
+
+- 解决方案 (一): 您可以参考数据集 [`DRIVE`](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#drive) 的配置文件, 它的 [数据集类](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/drive.py) 如下所示:
+
+```python
+class DRIVEDataset(CustomDataset):
+    CLASSES = ('background', 'vessel')
+
+    PALETTE = [[120, 120, 120], [6, 230, 230]]
+
+    def __init__(self, **kwargs):
+        super(DRIVEDataset, self).__init__(
+            img_suffix='.png',
+            seg_map_suffix='_manual1.png',
+            reduce_zero_label=False,
+            **kwargs)
+        assert self.file_client.exists(self.img_dir)
+```
+
+并且在 [数据集](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/datasets/drive.py) 和 [模型](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/fcn_unet_s5-d16.py#L23-L48) 对应的配置文件里设置:
+
+```python
+xxx_head=dict(
+    num_classes=2,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0))
+```
+
+- 解决方案 (二): 在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), 我们修复了当 `num_classes=1` 时的二值分割的问题. 您可以参考这个 issue [#2201](https://github.com/open-mmlab/mmsegmentation/issues/2201), 设置 `num_classes=1` 和 `CrossEntropyLoss` 里的 `use_sigmoid=True`.
+
+## `reduce_zero_label` 的作用
+
+在 MMSegmentation 里面, 当 [加载注释](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) 时, `reduce_zero_label (bool)` 被用来决定是否将所有 label 减去 1:
+
+```python
+if self.reduce_zero_label:
+    # avoid using underflow conversion
+    gt_semantic_seg[gt_semantic_seg == 0] = 255
+    gt_semantic_seg = gt_semantic_seg - 1
+    gt_semantic_seg[gt_semantic_seg == 254] = 255
+```
+
+`reduce_zero_label` 常常被用来处理 label 0 是背景的数据集, 如果 `reduce_zero_label=True`, label 0 对应的像素将不会参与损失函数的计算.

From 0255eb969cb4bb5f9d062c4151478c8084d5d884 Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Tue, 18 Oct 2022 17:40:36 +0800
Subject: [PATCH 02/11] update

---
 docs/en/faq.md                           | 33 ++++++++++++++++++++++++
 docs/zh_cn/faq.md                        | 33 ++++++++++++++++++++++++
 mmseg/models/decode_heads/decode_head.py |  2 +-
 3 files changed, 67 insertions(+), 1 deletion(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index 09c69b915f..b526c80d26 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -133,6 +133,39 @@ xxx_head=dict(
 
 - Solution Two: In [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), we fix the binary segmentation task when `num_classes=1`. You can follow this [#2201](https://github.com/open-mmlab/mmsegmentation/issues/2201) by setting `num_classes=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
 
+In summary, we encourage beginners to use `num_classes=2`, `out_channels=1` and `use_sigmoid=True` in `CrossEntropyLoss`, below is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
+
+```python
+decode_head=dict(
+    type='PSPHead',
+    in_channels=64,
+    in_index=4,
+    num_classes=2,
+    out_channels=1,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0)),
+auxiliary_head=dict(
+    type='FCNHead',
+    in_channels=128,
+    in_index=3,
+    num_classes=2,
+    out_channels=1,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.4)),
+```
+
+In [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) we also provide a parameter `threshold` for binary segmentation in the case of `out_channels=1`. It would be used to calculate segmentation prediction in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
+
+```python
+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)
+```
+
+By setting different value of `threshold`, users can calculate ROC(Receiver Operating Characteristic) Curve and AUC(Area Under Curve).
+
 ## What does `reduce_zero_label` work for?
 
 When [loading annotation](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) in MMSegmentation, `reduce_zero_label (bool)` is provided to determine whether reduce all label value by 1:
diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index e71fdb2f38..20a1cd5218 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -133,6 +133,39 @@ xxx_head=dict(
 
 - 解决方案 (二): 在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), 我们修复了当 `num_classes=1` 时的二值分割的问题. 您可以参考这个 issue [#2201](https://github.com/open-mmlab/mmsegmentation/issues/2201), 设置 `num_classes=1` 和 `CrossEntropyLoss` 里的 `use_sigmoid=True`.
 
+综上所述, 我们鼓励初学者设置 `num_classes=2`, `out_channels=1` 和 `CrossEntropyLoss` 里的 `use_sigmoid=True`, 下面是 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 的一个修改样例:
+
+```python
+decode_head=dict(
+    type='PSPHead',
+    in_channels=64,
+    in_index=4,
+    num_classes=2,
+    out_channels=1,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0)),
+auxiliary_head=dict(
+    type='FCNHead',
+    in_channels=128,
+    in_index=3,
+    num_classes=2,
+    out_channels=1,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.4)),
+```
+
+对于二值分割任务, 在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 里面我们还提供了一个 `threshold` 参数来处理 `out_channels=1` 的情况. 它会在 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py) 里面用来得到预测的结果:
+
+```python
+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)
+```
+
+通过设置 `threshold` 不同的值, 用户可以由此计算出 ROC 曲线和 AUC 的值.
+
 ## `reduce_zero_label` 的作用
 
 在 MMSegmentation 里面, 当 [加载注释](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) 时, `reduce_zero_label (bool)` 被用来决定是否将所有 label 减去 1:
diff --git a/mmseg/models/decode_heads/decode_head.py b/mmseg/models/decode_heads/decode_head.py
index f6b05dd3eb..c893f76e75 100644
--- a/mmseg/models/decode_heads/decode_head.py
+++ b/mmseg/models/decode_heads/decode_head.py
@@ -21,7 +21,7 @@ class BaseDecodeHead(BaseModule, metaclass=ABCMeta):
         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.
+            `out_channels==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.

From 207f3c400decc754bcf577c4c08d7b94a5eca278 Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Tue, 18 Oct 2022 23:55:39 +0800
Subject: [PATCH 03/11] modify

---
 docs/en/faq.md    | 117 +++++++++++++++++++--------------------------
 docs/zh_cn/faq.md | 118 +++++++++++++++++++---------------------------
 2 files changed, 98 insertions(+), 137 deletions(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index b526c80d26..ecabb3316d 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -67,73 +67,65 @@ In the test script, we provide `show-dir` argument to control whether output the
 python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --opacity 1
 ```
 
-## Why is the IoU always 0, NaN or very low in binary segmentation task
+## How to handle binary segmentation task
 
-Sometimes when training customized dataset, the IoU of certain class is 0, NaN or very low like below:
+MMSegmentation uses `num_classes` and `out_channels` to control output of last layer `self.conv_seg` (More details could be found [here](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).):
 
-```
-+--------------+-------+-------+
-|    Class     |  IoU  |  Acc  |
-+--------------+-------+-------+
-| label_a | 80.19 | 100.0 |
-| label_b  |  nan  |  nan  |
-+--------------+-------+-------+
-2022-10-18 10:56:56,032 - mmseg - INFO - Summary:
-2022-10-18 10:56:56,032 - mmseg - INFO -
-+-------+-------+-------+
-|  aAcc |  mIoU |  mAcc |
-+-------+-------+-------+
-| 100.0 | 80.19 | 100.0 |
-+-------+-------+-------+
-```
-
-or
-
-```
-+------------+------+-------+
-|   Class    | IoU  |  Acc  |
-+------------+------+-------+
-| label_a | 0.0  |  0.0  |
-|   label_b   | 1.77 | 100.0 |
-+------------+------+-------+
-2022-10-18 00:57:12,082 - mmseg - INFO - Summary:
-2022-10-18 00:57:12,083 - mmseg - INFO -
-+------+------+------+
-| aAcc | mIoU | mAcc |
-+------+------+------+
-| 1.77 | 0.88 | 50.0 |
-+------+------+------+
+```python
+def __init__(self,
+             ...,
+             ):
+  ...
+  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
+  ...
+  self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
 ```
 
-- Solution One: You can follow our config file of dataset [`DRIVE`](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#drive) for reference, whose [dataset class](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/drive.py) is like below:
+There are two types of calculating binary segmentation methods. First, when `out_channels=2`, using `F.softmax()` and `argmax()` to get prediction and then calculating by Cross Entropy Loss. Second, in the case of `out_channels=1`, in [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) we provide a parameter `threshold(default to 0.3)` for binary segmentation. Using `F.sigmoid()` and `threshold` to get prediction and then calculating by Binacry Cross Entropy Loss:
 
 ```python
-class DRIVEDataset(CustomDataset):
-    CLASSES = ('background', 'vessel')
-
-    PALETTE = [[120, 120, 120], [6, 230, 230]]
-
-    def __init__(self, **kwargs):
-        super(DRIVEDataset, self).__init__(
-            img_suffix='.png',
-            seg_map_suffix='_manual1.png',
-            reduce_zero_label=False,
-            **kwargs)
-        assert self.file_client.exists(self.img_dir)
-```
+...
+if self.out_channels == 1:
+    seg_logit = F.sigmoid(seg_logit)
+else:
+    seg_logit = F.softmax(seg_logit, dim=1)
 
-And in corresponding config files of [dataset](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/datasets/drive.py) and [model](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/fcn_unet_s5-d16.py#L23-L48):
+...
 
-```python
-xxx_head=dict(
-    num_classes=2,
-    loss_decode=dict(
-        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0))
+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)
 ```
 
-- Solution Two: In [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), we fix the binary segmentation task when `num_classes=1`. You can follow this [#2201](https://github.com/open-mmlab/mmsegmentation/issues/2201) by setting `num_classes=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
+More details about calculating segmentation prediction could be found in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
+
+In summary, we encourage beginners to take solution (1) `num_classes=2`, `out_channels=2` and `use_sigmoid=False` in `CrossEntropyLoss` or (2) `num_classes=2`, `out_channels=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
 
-In summary, we encourage beginners to use `num_classes=2`, `out_channels=1` and `use_sigmoid=True` in `CrossEntropyLoss`, below is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
+When taking solution (2), below is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
 
 ```python
 decode_head=dict(
@@ -154,18 +146,6 @@ auxiliary_head=dict(
         type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.4)),
 ```
 
-In [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) we also provide a parameter `threshold` for binary segmentation in the case of `out_channels=1`. It would be used to calculate segmentation prediction in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
-
-```python
-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)
-```
-
-By setting different value of `threshold`, users can calculate ROC(Receiver Operating Characteristic) Curve and AUC(Area Under Curve).
-
 ## What does `reduce_zero_label` work for?
 
 When [loading annotation](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) in MMSegmentation, `reduce_zero_label (bool)` is provided to determine whether reduce all label value by 1:
@@ -179,3 +159,4 @@ if self.reduce_zero_label:
 ```
 
 `reduce_zero_label` is usually used for datasets where 0 is background label, if `reduce_zero_label=True`, the pixels whose corresponding label is 0 would not be involved in loss calculation.
+Noted that in binary segmentation task it is unnecessary to use `reduce_zero_label=True`, take solutions we mentioned above please.
diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index 20a1cd5218..a26f9e27d8 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -67,73 +67,65 @@
 python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --opacity 1
 ```
 
-## 为什么在二值分割任务里 IoU 总是 0, NaN 或者非常低?
+## 如何处理二值分割任务?
 
-有时候在训练自定义的数据集时, 如下所示, 某个类别的 IoU 总是 0, NaN 或者很低:
+MMSegmentation 使用 `num_classes` 和 `out_channels` 来控制模型最后一层 `self.conv_seg` 的输出. (更多细节可以参考 [这里](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).):
 
-```
-+--------------+-------+-------+
-|    Class     |  IoU  |  Acc  |
-+--------------+-------+-------+
-| label_a | 80.19 | 100.0 |
-| label_b  |  nan  |  nan  |
-+--------------+-------+-------+
-2022-10-18 10:56:56,032 - mmseg - INFO - Summary:
-2022-10-18 10:56:56,032 - mmseg - INFO -
-+-------+-------+-------+
-|  aAcc |  mIoU |  mAcc |
-+-------+-------+-------+
-| 100.0 | 80.19 | 100.0 |
-+-------+-------+-------+
-```
-
-或者
-
-```
-+------------+------+-------+
-|   Class    | IoU  |  Acc  |
-+------------+------+-------+
-| label_a | 0.0  |  0.0  |
-|   label_b   | 1.77 | 100.0 |
-+------------+------+-------+
-2022-10-18 00:57:12,082 - mmseg - INFO - Summary:
-2022-10-18 00:57:12,083 - mmseg - INFO -
-+------+------+------+
-| aAcc | mIoU | mAcc |
-+------+------+------+
-| 1.77 | 0.88 | 50.0 |
-+------+------+------+
+```python
+def __init__(self,
+             ...,
+             ):
+  ...
+  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
+  ...
+  self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
 ```
 
-- 解决方案 (一): 您可以参考数据集 [`DRIVE`](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/dataset_prepare.md#drive) 的配置文件, 它的 [数据集类](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/drive.py) 如下所示:
+有两种计算二值分割任务的方法. 第一种是当 `out_channels=2` 时, 使用 `F.softmax()` 然后通过 `argmax()` 得到预测结果, 再以 Cross Entropy Loss 作为损失函数. 第二种是当 `out_channels=1` 时, 使用在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 提供的参数 `threshold (默认为 0.3)`. 通过 `F.sigmoid()` 和 `threshold` 得到预测结果, 再~~~~以 Binary Cross Entropy Loss 作为损失函数.
 
 ```python
-class DRIVEDataset(CustomDataset):
-    CLASSES = ('background', 'vessel')
-
-    PALETTE = [[120, 120, 120], [6, 230, 230]]
-
-    def __init__(self, **kwargs):
-        super(DRIVEDataset, self).__init__(
-            img_suffix='.png',
-            seg_map_suffix='_manual1.png',
-            reduce_zero_label=False,
-            **kwargs)
-        assert self.file_client.exists(self.img_dir)
-```
+...
+if self.out_channels == 1:
+    seg_logit = F.sigmoid(seg_logit)
+else:
+    seg_logit = F.softmax(seg_logit, dim=1)
 
-并且在 [数据集](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/datasets/drive.py) 和 [模型](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/fcn_unet_s5-d16.py#L23-L48) 对应的配置文件里设置:
+...
 
-```python
-xxx_head=dict(
-    num_classes=2,
-    loss_decode=dict(
-        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0))
+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)
 ```
 
-- 解决方案 (二): 在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), 我们修复了当 `num_classes=1` 时的二值分割的问题. 您可以参考这个 issue [#2201](https://github.com/open-mmlab/mmsegmentation/issues/2201), 设置 `num_classes=1` 和 `CrossEntropyLoss` 里的 `use_sigmoid=True`.
+更多关于计算语义分割预测的细节可以参考 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
+
+综上所述, 我们建议使用者采取两种解决方案 (1) `num_classes=2`, `out_channels=2` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=False` 或者 (2) `num_classes=2`, `out_channels=1` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=True`.
 
-综上所述, 我们鼓励初学者设置 `num_classes=2`, `out_channels=1` 和 `CrossEntropyLoss` 里的 `use_sigmoid=True`, 下面是 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 的一个修改样例:
+当采用解决方案 (2) 时, 下面是对样例 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 的对应的修改:
 
 ```python
 decode_head=dict(
@@ -154,18 +146,6 @@ auxiliary_head=dict(
         type='CrossEntropyLoss', use_sigmoid=True, loss_weight=0.4)),
 ```
 
-对于二值分割任务, 在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 里面我们还提供了一个 `threshold` 参数来处理 `out_channels=1` 的情况. 它会在 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py) 里面用来得到预测的结果:
-
-```python
-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)
-```
-
-通过设置 `threshold` 不同的值, 用户可以由此计算出 ROC 曲线和 AUC 的值.
-
 ## `reduce_zero_label` 的作用
 
 在 MMSegmentation 里面, 当 [加载注释](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) 时, `reduce_zero_label (bool)` 被用来决定是否将所有 label 减去 1:
@@ -178,4 +158,4 @@ if self.reduce_zero_label:
     gt_semantic_seg[gt_semantic_seg == 254] = 255
 ```
 
-`reduce_zero_label` 常常被用来处理 label 0 是背景的数据集, 如果 `reduce_zero_label=True`, label 0 对应的像素将不会参与损失函数的计算.
+`reduce_zero_label` 常常被用来处理 label 0 是背景的数据集, 如果 `reduce_zero_label=True`, label 0 对应的像素将不会参与损失函数的计算. 需要说明的是在二值分割任务中没有必要设置 `reduce_zero_label=True`, 请采用上面我们提到的解决方案.

From 8a72af73aad43999bae4a27d4776f2d18729148c Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Thu, 20 Oct 2022 13:22:51 +0800
Subject: [PATCH 04/11] fix typo and add modification

---
 docs/en/faq.md    | 14 +++++++++++---
 docs/zh_cn/faq.md | 14 +++++++++++---
 2 files changed, 22 insertions(+), 6 deletions(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index ecabb3316d..146de53ddb 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -103,7 +103,7 @@ def __init__(self,
   self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
 ```
 
-There are two types of calculating binary segmentation methods. First, when `out_channels=2`, using `F.softmax()` and `argmax()` to get prediction and then calculating by Cross Entropy Loss. Second, in the case of `out_channels=1`, in [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) we provide a parameter `threshold(default to 0.3)` for binary segmentation. Using `F.sigmoid()` and `threshold` to get prediction and then calculating by Binacry Cross Entropy Loss:
+There are two types of calculating binary segmentation methods:
 
 ```python
 ...
@@ -121,9 +121,17 @@ else:
     seg_pred = seg_logit.argmax(dim=1)
 ```
 
+- When `out_channels=2`, using Cross Entropy Loss in training, using `F.softmax()` and `argmax()` to get prediction of each pixel in inference.
+
+- When `out_channels=1`, we provide a parameter `threshold(default to 0.3)` in [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), using Binary Cross Entropy Loss in training, using `F.sigmoid()` and `threshold` to get prediction of each pixel in inference.
+
 More details about calculating segmentation prediction could be found in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
 
-In summary, we encourage beginners to take solution (1) `num_classes=2`, `out_channels=2` and `use_sigmoid=False` in `CrossEntropyLoss` or (2) `num_classes=2`, `out_channels=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
+In summary, to implement binary segmentation methods users should modify below parameters in config files:
+
+- (1) `num_classes=2`, `out_channels=2`  and `use_sigmoid=False` in `CrossEntropyLoss`.
+
+- (2) `num_classes=2`, `out_channels=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
 
 When taking solution (2), below is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
 
@@ -159,4 +167,4 @@ if self.reduce_zero_label:
 ```
 
 `reduce_zero_label` is usually used for datasets where 0 is background label, if `reduce_zero_label=True`, the pixels whose corresponding label is 0 would not be involved in loss calculation.
-Noted that in binary segmentation task it is unnecessary to use `reduce_zero_label=True`, take solutions we mentioned above please.
+Noted that in binary segmentation task it is unnecessary to use `reduce_zero_label=True`, please take solutions we mentioned above.
diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index a26f9e27d8..3363ed9dff 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -103,7 +103,11 @@ def __init__(self,
   self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
 ```
 
-有两种计算二值分割任务的方法. 第一种是当 `out_channels=2` 时, 使用 `F.softmax()` 然后通过 `argmax()` 得到预测结果, 再以 Cross Entropy Loss 作为损失函数. 第二种是当 `out_channels=1` 时, 使用在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 提供的参数 `threshold (默认为 0.3)`. 通过 `F.sigmoid()` 和 `threshold` 得到预测结果, 再~~~~以 Binary Cross Entropy Loss 作为损失函数.
+有两种计算二值分割任务的方法:
+
+- 当 `out_channels=2` 时, 在训练时以 Cross Entropy Loss 作为损失函数, 在推理时使用 `F.softmax()` 归一化 logits 值, 然后通过 `argmax()` 得到每个像素的预测结果.
+
+- 当 `out_channels=1` 时, 我们在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 里提供了阈值参数 `threshold (默认为 0.3)`, 在训练时以 Binary Cross Entropy Loss 作为损失函数, 在推理时使用 `F.sigmoid()` 和在 `threshold` 得到预测结果.
 
 ```python
 ...
@@ -123,9 +127,13 @@ else:
 
 更多关于计算语义分割预测的细节可以参考 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
 
-综上所述, 我们建议使用者采取两种解决方案 (1) `num_classes=2`, `out_channels=2` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=False` 或者 (2) `num_classes=2`, `out_channels=1` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=True`.
+对于实现上述两种计算二值分割的方法, 需要分别在配置文件里修改:
+
+- (1) `num_classes=2`, `out_channels=2` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=False`
+
+- (2) `num_classes=2`, `out_channels=1` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=True`.
 
-当采用解决方案 (2) 时, 下面是对样例 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 的对应的修改:
+如果采用解决方案 (2), 下面是对样例 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 做出的对应修改:
 
 ```python
 decode_head=dict(

From f63d67561b5be3cda3c4f948e448b77506dfb643 Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Fri, 21 Oct 2022 18:06:29 +0800
Subject: [PATCH 05/11] fix typo

---
 docs/en/faq.md    | 2 +-
 docs/zh_cn/faq.md | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index 146de53ddb..8ff2ecef04 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -127,7 +127,7 @@ else:
 
 More details about calculating segmentation prediction could be found in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
 
-In summary, to implement binary segmentation methods users should modify below parameters in config files:
+In summary, to implement binary segmentation methods users should modify below parameters in the `decode_head` and `auxiliary_head` configs:
 
 - (1) `num_classes=2`, `out_channels=2`  and `use_sigmoid=False` in `CrossEntropyLoss`.
 
diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index 3363ed9dff..24515a659a 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -107,7 +107,7 @@ def __init__(self,
 
 - 当 `out_channels=2` 时, 在训练时以 Cross Entropy Loss 作为损失函数, 在推理时使用 `F.softmax()` 归一化 logits 值, 然后通过 `argmax()` 得到每个像素的预测结果.
 
-- 当 `out_channels=1` 时, 我们在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 里提供了阈值参数 `threshold (默认为 0.3)`, 在训练时以 Binary Cross Entropy Loss 作为损失函数, 在推理时使用 `F.sigmoid()` 和在 `threshold` 得到预测结果.
+- 当 `out_channels=1` 时, 我们在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 里提供了阈值参数 `threshold (默认为 0.3)`, 在训练时以 Binary Cross Entropy Loss 作为损失函数, 在推理时使用 `F.sigmoid()` 和 `threshold` 得到预测结果.
 
 ```python
 ...
@@ -127,7 +127,7 @@ else:
 
 更多关于计算语义分割预测的细节可以参考 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
 
-对于实现上述两种计算二值分割的方法, 需要分别在配置文件里修改:
+对于实现上述两种计算二值分割的方法, 需要在 `decode_head` 和 `auxiliary_head` 的配置里修改:
 
 - (1) `num_classes=2`, `out_channels=2` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=False`
 

From 49261e207e670acdfe2c4487b4348f0d0af91d2e Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Mon, 24 Oct 2022 12:45:08 +0800
Subject: [PATCH 06/11] fix comments

---
 docs/en/faq.md    | 83 ++++++++++++++-------------------------------
 docs/zh_cn/faq.md | 86 +++++++++++++++--------------------------------
 2 files changed, 54 insertions(+), 115 deletions(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index 8ff2ecef04..142ba976e9 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -71,69 +71,39 @@ python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --op
 
 MMSegmentation uses `num_classes` and `out_channels` to control output of last layer `self.conv_seg` (More details could be found [here](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).):
 
-```python
-def __init__(self,
-             ...,
-             ):
-  ...
-  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
-  ...
-  self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
-```
-
-There are two types of calculating binary segmentation methods:
-
-```python
-...
-if self.out_channels == 1:
-    seg_logit = F.sigmoid(seg_logit)
-else:
-    seg_logit = F.softmax(seg_logit, 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)
-```
+- Set `out_channels=2`, using Cross Entropy Loss in training, using `F.softmax()` and `argmax()` to get prediction of each pixel in inference.
 
-- When `out_channels=2`, using Cross Entropy Loss in training, using `F.softmax()` and `argmax()` to get prediction of each pixel in inference.
+- Set `out_channels=1`, using Binary Cross Entropy Loss in training, using `F.sigmoid()` and `threshold` to get prediction of each pixel in inference. `threshold` is set 0.3 as default.
 
-- When `out_channels=1`, we provide a parameter `threshold(default to 0.3)` in [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016), using Binary Cross Entropy Loss in training, using `F.sigmoid()` and `threshold` to get prediction of each pixel in inference.
+More details about implementation could be found in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
 
-More details about calculating segmentation prediction could be found in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
+In summary, to implement binary segmentation methods users should modify below parameters in the `decode_head` and `auxiliary_head` configs. Here is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
 
-In summary, to implement binary segmentation methods users should modify below parameters in the `decode_head` and `auxiliary_head` configs:
+`num_classes` should be the same as number of types of labels, in binary segmentation task, dataset only has two types of labels: foreground and background, so `num_classes=2`. `out_channels` controls the output channel of last layer of model, it usually equals to `num_classes`.
+But in binary segmentation task, there are two solutions:
 
 - (1) `num_classes=2`, `out_channels=2`  and `use_sigmoid=False` in `CrossEntropyLoss`.
 
-- (2) `num_classes=2`, `out_channels=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
+```python
+decode_head=dict(
+    type='PSPHead',
+    in_channels=64,
+    in_index=4,
+    num_classes=2,
+    out_channels=2,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+auxiliary_head=dict(
+    type='FCNHead',
+    in_channels=128,
+    in_index=3,
+    num_classes=2,
+    out_channels=2,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+```
 
-When taking solution (2), below is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
+- (2) `num_classes=2`, `out_channels=1` and `use_sigmoid=True` in `CrossEntropyLoss`.
 
 ```python
 decode_head=dict(
@@ -166,5 +136,4 @@ if self.reduce_zero_label:
     gt_semantic_seg[gt_semantic_seg == 254] = 255
 ```
 
-`reduce_zero_label` is usually used for datasets where 0 is background label, if `reduce_zero_label=True`, the pixels whose corresponding label is 0 would not be involved in loss calculation.
-Noted that in binary segmentation task it is unnecessary to use `reduce_zero_label=True`, please take solutions we mentioned above.
+Noted that in please check out label numbers of dataset when using `reduce_zero_label`. If dataset only has two types of labels (i.e., label 0 and 1), it needs to close `reduce_zero_label`, i.e., set `reduce_zero_label=True`.
diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index 24515a659a..b3ec9945bf 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -69,71 +69,40 @@ python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --op
 
 ## 如何处理二值分割任务?
 
-MMSegmentation 使用 `num_classes` 和 `out_channels` 来控制模型最后一层 `self.conv_seg` 的输出. (更多细节可以参考 [这里](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).):
+MMSegmentation 使用 `num_classes` 和 `out_channels` 来控制模型最后一层 `self.conv_seg` 的输出. 更多细节可以参考 [这里](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).
 
-```python
-def __init__(self,
-             ...,
-             ):
-  ...
-  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
-  ...
-  self.conv_seg = nn.Conv2d(channels, self.out_channels, kernel_size=1)
-```
-
-有两种计算二值分割任务的方法:
+`num_classes` 应该和数据集本身类别个数一致，当是二值分割时，数据集只有前景和背景两类，所以 `num_classes` 为 2。`out_channels` 控制模型最后一层的输出的通道数，通常和 `num_classes` 相等，但当二值分割时候，可以有两种处理方法，分别是：
 
-- 当 `out_channels=2` 时, 在训练时以 Cross Entropy Loss 作为损失函数, 在推理时使用 `F.softmax()` 归一化 logits 值, 然后通过 `argmax()` 得到每个像素的预测结果.
+- 设置 `out_channels=2`, 在训练时以 Cross Entropy Loss 作为损失函数, 在推理时使用 `F.softmax()` 归一化 logits 值, 然后通过 `argmax()` 得到每个像素的预测结果.
 
-- 当 `out_channels=1` 时, 我们在 [#2016](https://github.com/open-mmlab/mmsegmentation/pull/2016) 里提供了阈值参数 `threshold (默认为 0.3)`, 在训练时以 Binary Cross Entropy Loss 作为损失函数, 在推理时使用 `F.sigmoid()` 和 `threshold` 得到预测结果.
-
-```python
-...
-if self.out_channels == 1:
-    seg_logit = F.sigmoid(seg_logit)
-else:
-    seg_logit = F.softmax(seg_logit, 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)
-```
+- 设置 `out_channels=1`, 在训练时以 Binary Cross Entropy Loss 作为损失函数, 在推理时使用 `F.sigmoid()` 和 `threshold` 得到预测结果, `threshold` 默认为 0.3.
 
-更多关于计算语义分割预测的细节可以参考 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
+更多关于实现细节可以参考 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
 
-对于实现上述两种计算二值分割的方法, 需要在 `decode_head` 和 `auxiliary_head` 的配置里修改:
+对于实现上述两种计算二值分割的方法, 需要在 `decode_head` 和 `auxiliary_head` 的配置里修改. 下面是对样例 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 做出的对应修改.
 
 - (1) `num_classes=2`, `out_channels=2` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=False`
 
-- (2) `num_classes=2`, `out_channels=1` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=True`.
+```python
+decode_head=dict(
+    type='PSPHead',
+    in_channels=64,
+    in_index=4,
+    num_classes=2,
+    out_channels=2,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
+auxiliary_head=dict(
+    type='FCNHead',
+    in_channels=128,
+    in_index=3,
+    num_classes=2,
+    out_channels=2,
+    loss_decode=dict(
+        type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
+```
 
-如果采用解决方案 (2), 下面是对样例 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 做出的对应修改:
+- (2) `num_classes=2`, `out_channels=1` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=True`.
 
 ```python
 decode_head=dict(
@@ -156,7 +125,8 @@ auxiliary_head=dict(
 
 ## `reduce_zero_label` 的作用
 
-在 MMSegmentation 里面, 当 [加载注释](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) 时, `reduce_zero_label (bool)` 被用来决定是否将所有 label 减去 1:
+数据集中 `reduce_zero_label`参数类型为布尔类型, 默认为 False, 它的功能是为了忽略数据集 label 0. 具体做法是将 label 0 改为 255, 其余 label 相应编号减1, 同时 decode head 里将 255 设为 ignore index, 即不参与 loss 计算.
+以下是 `reduce_zero_label` 具体实现逻辑:
 
 ```python
 if self.reduce_zero_label:
@@ -166,4 +136,4 @@ if self.reduce_zero_label:
     gt_semantic_seg[gt_semantic_seg == 254] = 255
 ```
 
-`reduce_zero_label` 常常被用来处理 label 0 是背景的数据集, 如果 `reduce_zero_label=True`, label 0 对应的像素将不会参与损失函数的计算. 需要说明的是在二值分割任务中没有必要设置 `reduce_zero_label=True`, 请采用上面我们提到的解决方案.
+需要注意的是, 使用 `reduce_zero_label` 请确认数据集原始类别个数, 如果只有两类, 需要关闭 `reduce_zero_label` 即设置 `reduce_zero_label=False`.

From 1854d192e23e63da3b2b5e6ee358f80f94688c55 Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Mon, 24 Oct 2022 13:33:40 +0800
Subject: [PATCH 07/11] fix order

---
 docs/en/faq.md | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index 142ba976e9..c6814a953e 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -69,7 +69,10 @@ python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --op
 
 ## How to handle binary segmentation task
 
-MMSegmentation uses `num_classes` and `out_channels` to control output of last layer `self.conv_seg` (More details could be found [here](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).):
+MMSegmentation uses `num_classes` and `out_channels` to control output of last layer `self.conv_seg`. More details could be found [here](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).
+
+`num_classes` should be the same as number of types of labels, in binary segmentation task, dataset only has two types of labels: foreground and background, so `num_classes=2`. `out_channels` controls the output channel of last layer of model, it usually equals to `num_classes`.
+But in binary segmentation task, there are two solutions:
 
 - Set `out_channels=2`, using Cross Entropy Loss in training, using `F.softmax()` and `argmax()` to get prediction of each pixel in inference.
 
@@ -79,9 +82,6 @@ More details about implementation could be found in [encoder_decoder.py](https:/
 
 In summary, to implement binary segmentation methods users should modify below parameters in the `decode_head` and `auxiliary_head` configs. Here is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
 
-`num_classes` should be the same as number of types of labels, in binary segmentation task, dataset only has two types of labels: foreground and background, so `num_classes=2`. `out_channels` controls the output channel of last layer of model, it usually equals to `num_classes`.
-But in binary segmentation task, there are two solutions:
-
 - (1) `num_classes=2`, `out_channels=2`  and `use_sigmoid=False` in `CrossEntropyLoss`.
 
 ```python

From 35f578da51ecf54927fc4d3a7bffe74818d9b178 Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Mon, 24 Oct 2022 14:01:40 +0800
Subject: [PATCH 08/11] fix

---
 docs/en/faq.md    | 2 --
 docs/zh_cn/faq.md | 6 ++----
 2 files changed, 2 insertions(+), 6 deletions(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index c6814a953e..404ba1005c 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -78,8 +78,6 @@ But in binary segmentation task, there are two solutions:
 
 - Set `out_channels=1`, using Binary Cross Entropy Loss in training, using `F.sigmoid()` and `threshold` to get prediction of each pixel in inference. `threshold` is set 0.3 as default.
 
-More details about implementation could be found in [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
-
 In summary, to implement binary segmentation methods users should modify below parameters in the `decode_head` and `auxiliary_head` configs. Here is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
 
 - (1) `num_classes=2`, `out_channels=2`  and `use_sigmoid=False` in `CrossEntropyLoss`.
diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index b3ec9945bf..c4dda95cb5 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -71,17 +71,15 @@ python tools/test.py {config} {checkpoint} --show-dir {/path/to/save/image} --op
 
 MMSegmentation 使用 `num_classes` 和 `out_channels` 来控制模型最后一层 `self.conv_seg` 的输出. 更多细节可以参考 [这里](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).
 
-`num_classes` 应该和数据集本身类别个数一致，当是二值分割时，数据集只有前景和背景两类，所以 `num_classes` 为 2。`out_channels` 控制模型最后一层的输出的通道数，通常和 `num_classes` 相等，但当二值分割时候，可以有两种处理方法，分别是：
+`num_classes` 应该和数据集本身类别个数一致，当是二值分割时，数据集只有前景和背景两类, 所以 `num_classes` 为 2. `out_channels` 控制模型最后一层的输出的通道数，通常和 `num_classes` 相等, 但当二值分割时候, 可以有两种处理方法, 分别是：
 
 - 设置 `out_channels=2`, 在训练时以 Cross Entropy Loss 作为损失函数, 在推理时使用 `F.softmax()` 归一化 logits 值, 然后通过 `argmax()` 得到每个像素的预测结果.
 
 - 设置 `out_channels=1`, 在训练时以 Binary Cross Entropy Loss 作为损失函数, 在推理时使用 `F.sigmoid()` 和 `threshold` 得到预测结果, `threshold` 默认为 0.3.
 
-更多关于实现细节可以参考 [encoder_decoder.py](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/segmentors/encoder_decoder.py):
-
 对于实现上述两种计算二值分割的方法, 需要在 `decode_head` 和 `auxiliary_head` 的配置里修改. 下面是对样例 [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py) 做出的对应修改.
 
-- (1) `num_classes=2`, `out_channels=2` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=False`
+- (1) `num_classes=2`, `out_channels=2` 并在 `CrossEntropyLoss` 里面设置 `use_sigmoid=False`.
 
 ```python
 decode_head=dict(

From 294b93b7e7ff658e442796a0d3ff6d734f021d6d Mon Sep 17 00:00:00 2001
From: MengzhangLI <mcmong@pku.edu.cn>
Date: Mon, 24 Oct 2022 14:03:42 +0800
Subject: [PATCH 09/11] fix

---
 docs/zh_cn/faq.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index c4dda95cb5..bf8e4514ed 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -123,7 +123,7 @@ auxiliary_head=dict(
 
 ## `reduce_zero_label` 的作用
 
-数据集中 `reduce_zero_label`参数类型为布尔类型, 默认为 False, 它的功能是为了忽略数据集 label 0. 具体做法是将 label 0 改为 255, 其余 label 相应编号减1, 同时 decode head 里将 255 设为 ignore index, 即不参与 loss 计算.
+数据集中 `reduce_zero_label` 参数类型为布尔类型, 默认为 False, 它的功能是为了忽略数据集 label 0. 具体做法是将 label 0 改为 255, 其余 label 相应编号减 1, 同时 decode head 里将 255 设为 ignore index, 即不参与 loss 计算.
 以下是 `reduce_zero_label` 具体实现逻辑:
 
 ```python

From 56bbfcc7163f8d41c01c56b6eb08098798f1dc71 Mon Sep 17 00:00:00 2001
From: Miao Zheng <76149310+MeowZheng@users.noreply.github.com>
Date: Fri, 28 Oct 2022 21:26:34 +0800
Subject: [PATCH 10/11] Update docs/en/faq.md

---
 docs/en/faq.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/en/faq.md b/docs/en/faq.md
index 404ba1005c..2ca2fd8293 100644
--- a/docs/en/faq.md
+++ b/docs/en/faq.md
@@ -134,4 +134,4 @@ if self.reduce_zero_label:
     gt_semantic_seg[gt_semantic_seg == 254] = 255
 ```
 
-Noted that in please check out label numbers of dataset when using `reduce_zero_label`. If dataset only has two types of labels (i.e., label 0 and 1), it needs to close `reduce_zero_label`, i.e., set `reduce_zero_label=True`.
+**Noted:** Please pay attention to label numbers of dataset when using `reduce_zero_label`. If dataset only has two types of labels (i.e., label 0 and 1), it needs to close `reduce_zero_label`, i.e., set `reduce_zero_label=False`.

From 84625fe9dd0b9b4e0d156b2bbb7a27907f664f9e Mon Sep 17 00:00:00 2001
From: Miao Zheng <76149310+MeowZheng@users.noreply.github.com>
Date: Fri, 28 Oct 2022 21:27:12 +0800
Subject: [PATCH 11/11] Update docs/zh_cn/faq.md

---
 docs/zh_cn/faq.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/zh_cn/faq.md b/docs/zh_cn/faq.md
index bf8e4514ed..3cab08a18a 100644
--- a/docs/zh_cn/faq.md
+++ b/docs/zh_cn/faq.md
@@ -134,4 +134,4 @@ if self.reduce_zero_label:
     gt_semantic_seg[gt_semantic_seg == 254] = 255
 ```
 
-需要注意的是, 使用 `reduce_zero_label` 请确认数据集原始类别个数, 如果只有两类, 需要关闭 `reduce_zero_label` 即设置 `reduce_zero_label=False`.
+**注意:** 使用 `reduce_zero_label` 请确认数据集原始类别个数, 如果只有两类, 需要关闭 `reduce_zero_label` 即设置 `reduce_zero_label=False`.