From 697821555aed9264d6af59fbfccd1f51f61b2157 Mon Sep 17 00:00:00 2001
From: exias152 <15851197936@163.com>
Date: Sat, 20 Aug 2022 16:35:04 +0800
Subject: [PATCH 1/2] =?UTF-8?q?=20!1525=20[=E6=B8=85=E5=8D=8E=E5=A4=A7?=
 =?UTF-8?q?=E5=AD=A6=E6=B7=B1=E5=9C=B3=E5=9B=BD=E9=99=85=E7=A0=94=E7=A9=B6?=
 =?UTF-8?q?=E7=94=9F=E9=99=A2][=E9=AB=98=E6=A0=A1=E8=B4=A1=E7=8C=AE][PyTor?=
 =?UTF-8?q?ch=E8=BF=81=E7=A7=BB1.8][FCN8S]-=E5=88=9D=E6=AC=A1=E6=8F=90?=
 =?UTF-8?q?=E4=BA=A4?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../cv/semantic_segmentation/FCN8s/README.md  | 277 ++++++++++++------
 .../mmseg/models/decode_heads/decode_head.py  |  17 ++
 .../FCN8s/test/train_full_1p.sh               |   2 +-
 .../FCN8s/test/train_full_8p.sh               |   2 +-
 .../FCN8s/test/train_performance_1p.sh        |   2 +-
 .../FCN8s/test/train_performance_8p.sh        |   2 +-
 .../FCN8s/tools/train.py                      |  11 +-
 7 files changed, 216 insertions(+), 97 deletions(-)

diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md
index 08f26f52ad..99334d81c7 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md
@@ -1,151 +1,244 @@
-# FCN8s
+# XXX for PyTorch\_Owner
 
-This implements training of PSPNet on the PASCAL VOC Aug dataset, mainly modified from [mmsegmentation](https://github.com/open-mmlab/mmsegmentation). 
+-   [交付件基本信息](交付件基本信息.md)
+-   [概述](概述.md)
+-   [准备训练环境](准备训练环境.md)
+-   [开始训练](开始训练.md)
+-   [训练结果展示](训练结果展示.md)
+-   [版本说明](版本说明.md)
 
-**ps**
+# 交付件基本信息
 
-1. As of the current date, Ascend-Pytorch doesn't support SyncBN, the backbone uses BN instead. To get a similar performance to SyncBN, we set a larger batch size of 16 rather than 4 in mmsegmentation.
-2. Semantic segmentation is trained by iteration. The model trained on 1 NPU is useless, so we do not give the evaluation script for the 1p model.
+应用领域（Application Domain）：Image Segmentation
 
+模型版本（Model Version）：1.1
 
-## Environment  preparation
+修改时间（Modified）：2022.8.18
 
-The latest Ascend-Pytorch version is 1.5.0. MMSegmentation 0.10.0 and mmcv 1.2.7 are chosen as they support pytorch1.5.0.
+_大小（Size）：378 MB
 
-1. Install the latest Ascend-Pytorch.
+框架（Framework）：PyTorch\_1.8.1
 
-2. Downding the repository of Ascned Model Zoo to the folder of `$YOURMODELZOO`
+模型格式（Model Format）：pth
 
-```
-# download source code
-cd $YOURMODELZOO
-git clone https://gitee.com/KevinKe/modelzoo
-# go to fcn8s
-cd contrib/PyTorch/Research/cv/semantic_segmentation/FCN8s
-```
+精度（Precision）：Mixed
 
-Denote `$FCN` as the path of `$YOURMODELZOO/contrib/PyTorch/Research/cv/semantic_segmentation/FCN8s`.
+处理器（Processor）：Ascend 910
 
-3. Build mmcv using
+应用级别（Categories）：Research
 
-Firstly, download [mmcv1.2.7](https://github.com/open-mmlab/mmcv/tree/v1.2.7) to the path `$YOURMMVCPATH`. Then, copy the `mmcv_replace` to `$YOURMMVCPATH/mmcv`.
+描述（Description）：基于PyTorch框架的FCN8S语义分割网络训练
 
-Check the numpy version is 1.21.2.
+# 概述
 
-```
-# configure
-cd $FCN
-source env_npu.sh
-
-# copy
-rm -rf $YOURMMVCPATH/mmcv
-mkdir mmcv
-cp -r mmcv_replace/* $YOURMMVCPATH/mmcv/
-
-# compile
-cd $YOURMMVCPATH
-export MMCV_WITH_OPS=1
-export MAX_JOBS=8
-python3.7.5 setup.py build_ext
-python3.7.5 setup.py develop
-pip3.7.5 list | grep mmcv
-```
+## 简述
+FCN8S是一个经典的语义分割网络，FCN8S使用全卷积结构，可以接受任意尺寸的输入图像，采用反卷积对最后一层的特征图进行上采样，得到与输入图像相同尺寸的输出，从而对输入进行逐像素预测。本代码主要在[mmsegmentation](https://github.com/open-mmlab/mmsegmentation)进行修改， 以PSPNet为backbone在PASCAL VOC数据集上进行训练。
 
-Then go back to the $PSPNET folder
-```
-cd $FCN
-```
+- 参考实现：
 
-4. Permission configuration
-```
-chmod -R 777 ./
-```
+  ```
+  url=https://github.com/pytorch/vision.git
+  ```
 
-5. remove the `mmcv_replace` folder
-```
-rm -rf mmcv_replace
-```
+- 适配昇腾 AI 处理器的实现：
+
+  ```
+  url=https://gitee.com/ascend/ModelZoo-PyTorch.git
+  code_path=PyTorch/contrib/cv/semantic_segmentation/FCN8s
+  ```
+  
+- 通过Git获取代码方法如下：
+
+  ```
+  git clone https://gitee.com/exias152/ModelZoo-PyTorch       # 克隆仓库的代码
+  cd contrib/PyTorch/Research/cv/semantic_segmentation/FCN8s        # 切换到模型代码所在路径
+  # Denote `$FCN` as the path of `$YOURMODELZOO/contrib/PyTorch/Research/cv/semantic_segmentation/FCN8s`.
+  ```
+  
+- 通过单击“立即下载”，下载源码包。
+
+# 准备训练环境
+
+## 准备环境
+
+1. 当前模型支持的固件与驱动、 CANN 以及 PyTorch 如下表所示。
+
+  **表 1**  版本配套表
+
+  | 配套       | 版本                                                         |
+  | ---------- | ------------------------------------------------------------ |
+  | 固件与驱动 | [1.0.15](https://www.hiascend.com/hardware/firmware-drivers?tag=commercial) |
+  | CANN       | [5.1.RC1](https://www.hiascend.com/software/cann/commercial?version=5.1.RC1) |
+  | PyTorch    | [1.8](https://gitee.com/ascend/pytorch/tree/master/))|
+
+- 环境准备指导。
+
+  请参考《[Pytorch框架训练环境准备](https://www.hiascend.com/document/detail/zh/ModelZoo/pytorchframework/ptes)》。
+  
+2. 安装依赖
+
+  ```
+  pip install -r requirements.txt
+  ```
+
+3. 构建mmcv。
+  下载[mmcv1.2.7](https://github.com/open-mmlab/mmcv/tree/v1.2.7)到路径`$YOURMMVCPATH`。然后，复制`mmcv_replace` 到 `$YOURMMVCPATH/mmcv`。
+  ```
+  # configure
+  cd $FCN
+  source env_npu.sh
+
+  # copy
+  rm -rf $YOURMMVCPATH/mmcv
+  mkdir mmcv
+  cp -r mmcv_replace/* $YOURMMVCPATH/mmcv/
+
+  # compile
+  cd $YOURMMVCPATH
+  export MMCV_WITH_OPS=1
+  export MAX_JOBS=8
+  python3.7.5 setup.py build_ext
+  python3.7.5 setup.py develop
+  pip3.7.5 list | grep mmcv
+  ```
+
+  Then go back to the $PSPNET folder
+  ```
+  cd $FCN
+
+  ```
+
+4. 权限配置。
+  ```
+  chmod -R 777 ./
+  ```
+5. 删除文件夹 `mmcv_replace` 
+  ```
+  rm -rf mmcv_replace
+  ```
 
-## Dataset Preparation
+## 准备数据集
 
-1. Download the training and validation set of [PASCAL VOC 2012 dataset](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar) and [PASCAL VOC2010 dataset](https://ascend-test-dataset.obs.cn-north-4.myhuaweicloud.com/train/zip/VOCtrainval_03-May-2010.tar). 
+1. 获取数据集。
+
+   下载[PASCAL VOC 2012 dataset](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar) 和 [PASCAL VOC2010 dataset](https://ascend-test-dataset.obs.cn-north-4.myhuaweicloud.com/train/zip/VOCtrainval_03-May-2010.tar)的训练集和验证集。
+   解压后，数据集目录结构如下所示。
 
-After decompression, the structure of the dataset folder should be:
 ```none
 ├── VOCdevkit
 │   │   ├── VOC2012
 │   │   ├── VOC2010
  ```  
 
-2. Download [PASCALAug dataset](http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/semantic_contours/benchmark.tgz).
-After depressing, copy `benchmark_REALSE/dataset` to `VOCaug` in the `VOCdevkit` folder.
-The structure of the dataset folder should be:
+   > **说明：** 
+   >该数据集的训练过程脚本只作为一种参考示例。
+
+2. 数据预处理。
+下载 [PASCALAug dataset](http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/semantic_contours/benchmark.tgz).解压后，复制`benchmark_REALSE/dataset`到`VOCdevkit`文件夹下的`VOCaug`。
+数据集目录结构如下所示。
 ```none
 ├── VOCdevkit
 │   │   ├── VOC2012
 │   │   ├── VOC2010
 │   │   ├── VOCaug
 ```
-3. Convert the VOCAug dataset using
+3. 处理VOCAug数据集。
 
 ```
 cd $PSPNET
 python tools/convert_datasets/voc_aug.py data/VOCdevkit data/VOCdevkit/VOCaug --nproc 8
 ```
+**Note: ** 可能会提示`Segmentation fault (core dumped)` ， 提示原因是mmcv需要pytorch支持。 请返回源码包根目录并运行 `source env_npu.sh`。 
 
-**Note: ** `Segmentation fault (core dumped)` may rise. The reason is that mmcv needs the support of pytorch. Go back to repo folder and run `source env_npu.sh` first. 
-
-4. [Optional] Make a soft link of the dataset to the folder of mmseg100
+4. [Optional] 建立数据集到文件夹mmseg100的软链。
 ```
 cd $FCN
 mkdir data
 ln -s VOCdevkit data # data_path=./data/VOCdevkit/VOC2012
 ```
 
-## Training
 
- **Note [Optional]:** When running scripts, the error `$'\r': command not found` may rise. Use `dos2unix  script_file_name` to change it from window format to Unix format first.
+# 开始训练
 
+## 训练模型
+ 
+ **Note [Optional]:**运行脚本的时候可能会报错 `$'\r': command not found` 。 使用 `dos2unix  script_file_name` 将脚本从windows格式转为unix格式。
 
-```bash
-cd $FCN
-source npu_env.sh
+1. 进入解压后的源码包根目录。
 
-# training 1p accuracy
-bash ./test/train_full_1p.sh --data_path=xxx 
-# --data_path=data/VOCdevkit/VOC2012
+   ```
+   cd $FCN 
+   ```
 
-# training 1p performance
-bash ./test/train_performance_1p.sh --data_path=xxx 
+2. 运行训练脚本。
 
-# training 8p accuracy
-bash ./test/train_full_8p.sh --data_path=xxx 
+   该模型支持单机单卡训练和单机8卡训练。
 
-# training 8p performance
-bash ./test/train_performance_8p.sh --data_path=xxx 
+   - 单机单卡训练
 
-# evaluation 8p accuracy
-bash ./test/train_val_8p.sh --data_path=xxx 
-```
+     启动单卡训练。
 
-Log and checkpoint path:
-```
-./output/devie_id/FCN/train_${device_id}.log          # training detail log
-./output/devie_id/FCN/FCN_bs16_8p_acc.log             # 8p training performance result log
-./output/devie_id/FCN/ckpt                            # checkpoits
-./output/devie_id/FCN_prof/FCN_bs16_8p_acc.log        # 8p training accuracy result log
+     ```
+     # training 1p accuracy
+     bash ./test/train_full_1p.sh --data_path=xxx --device_id=xxx
+     # --data_path=data/VOCdevkit/VOC2012    
+     # training 1p performance
+     bash ./test/train_performance_1p.sh --data_path=xxx --device_id=xxx
+     ```
 
-```
-    
+   - 单机8卡训练
+
+     启动8卡训练。
+
+     ```
+     # training 8p accuracy
+     bash ./test/train_full_8p.sh --data_path=xxx 
+
+     # training 8p performance
+     bash ./test/train_performance_8p.sh --data_path=xxx 
+
+     # evaluation 8p accuracy
+     bash ./test/train_val_8p.sh --data_path=xxx 
+     ```
+
+   --data\_path参数填写数据集路径。
+
+   
+   训练完成后，日志和权重文件保存在如下路径。
+    ```
+    ./output/devie_id/FCN/train_${device_id}.log          # training detail log
+    ./output/devie_id/FCN/FCN_bs16_8p_acc.log             # 8p training performance result log
+    ./output/devie_id/FCN/ckpt                            # checkpoits
+    ./output/devie_id/FCN_prof/FCN_bs16_8p_acc.log        # 8p training accuracy result log
+
+    ```  
 
-## Evaluation Details
+# 训练结果展示
 
-### FCN with 8p
+**表 2**  训练结果展示表
 
-| device | fps |  aAcc |  mIoU | mAcc |
+| 名称    |  FPS   |  aAcc |  mIoU |  mAcc |
 | :------: | :------: | :------: | :------: | :------: |
-|mmsegmentaion| |-- | 67.08| -- |
-|GPU-8p| 82.296| 93.16 | 69.19 | 78.7 |
-|NPU-8p| 135.19 | 93.23 | 69.36 | 78.88 |
+| 1p-竞品 | -----  | ----- | ----- | ----- |
+| 1p-NPU  | 20.99  | 93.28 | 69.62 | 79.29 |
+| 8p-竞品 | 135.19 | 93.23 | 69.36 | 78.88 |
+| 8p-NPU  | 156.30 | 93.24 | 69.54 | 78.69 |
+
+
+# 版本说明
+
+## 变更
+
+2022.8.18：更新内容，重新发布。
+
+2020.07.08：首次发布。
+
+
+
+
+
+
+
+
+
 
-ps: 2x training data are used for training on NPU (`bs*#NPU*#iter=16*8*10000`) v.s. GPU (`bs*#GPU*#iter=4*8*20000`)
\ No newline at end of file
diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/mmseg/models/decode_heads/decode_head.py b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/mmseg/models/decode_heads/decode_head.py
index b5a58aef8f..a132ea830e 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/mmseg/models/decode_heads/decode_head.py
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/mmseg/models/decode_heads/decode_head.py
@@ -1,3 +1,19 @@
+# Copyright (c) Soumith Chintala 2016,
+# All rights reserved
+#
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the BSD 3-Clause License (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# https://spdx.org/licenses/BSD-3-Clause.html
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 from abc import ABCMeta, abstractmethod
 
 import torch
@@ -207,6 +223,7 @@ class BaseDecodeHead(nn.Module, metaclass=ABCMeta):
     def cls_seg(self, feat):
         """Classify each pixel."""
         if self.dropout is not None:
+            torch.npu.synchronize()
             feat = self.dropout(feat)
         output = self.conv_seg(feat)
         return output
diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_1p.sh b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_1p.sh
index dd1170d992..61c629a750 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_1p.sh
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_1p.sh
@@ -65,7 +65,7 @@ if [ x"${etp_flag}" != x"true" ];then
 fi
 
 #执行训练脚本，以下传参不需要修改，其他需要模型审视修改
-python3.7.5 ./tools/train.py ./configs/fcn/fcn_r50-d8_512x512_20k_voc12aug.py \
+python3 ./tools/train.py ./configs/fcn/fcn_r50-d8_512x512_20k_voc12aug.py \
     --work-dir=${cur_path}/output/${Network}/$ASCEND_DEVICE_ID/ckpt \
     --device="npu" \
     --amp \
diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_8p.sh b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_8p.sh
index fbddc37fba..f6a51d8d35 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_8p.sh
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_full_8p.sh
@@ -98,7 +98,7 @@ CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'acc'
 
 ##获取性能数据，不需要修改
 #吞吐量
-ActualFPS=`awk 'BEGIN{printf "%.2f\n", '${RANK_SIZE}'*'${FPS}'}'`
+ActualFPS=`awk 'BEGIN{printf "%.2f\n", '${FPS}'}'`
 #单迭代训练时长
 TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
 
diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_1p.sh b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_1p.sh
index 158e51c0ff..185d5fe90d 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_1p.sh
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_1p.sh
@@ -65,7 +65,7 @@ if [ x"${etp_flag}" != x"true" ];then
 fi
 
 #执行训练脚本，以下传参不需要修改，其他需要模型审视修改
-python3.7.5 ./tools/train.py ./configs/fcn/fcn_r50-d8_512x512_20k_voc12aug.py \
+python3 ./tools/train.py ./configs/fcn/fcn_r50-d8_512x512_20k_voc12aug.py \
     --work-dir=${cur_path}/output/${Network}/$ASCEND_DEVICE_ID/ckpt \
     --device="npu" \
     --amp \
diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_8p.sh b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_8p.sh
index a280efbf93..0cd81e278f 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_8p.sh
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/test/train_performance_8p.sh
@@ -98,7 +98,7 @@ CaseName=${Network}_bs${BatchSize}_${RANK_SIZE}'p'_'acc'
 
 ##获取性能数据，不需要修改
 #吞吐量
-ActualFPS=`awk 'BEGIN{printf "%.2f\n", '${RANK_SIZE}'*'${FPS}'}'`
+ActualFPS=`awk 'BEGIN{printf "%.2f\n", '${FPS}'}'`
 #单迭代训练时长
 TrainingTime=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'*1000/'${FPS}'}'`
 
diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/tools/train.py b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/tools/train.py
index 89c4205595..3575c68e94 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/tools/train.py
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/tools/train.py
@@ -38,6 +38,9 @@ import time
 
 import mmcv
 import torch
+if torch.__version__ >="1.8":
+    import torch_npu
+#print(torch.__version__)
 from mmcv.runner import get_dist_info, init_dist
 from mmcv.utils import Config, DictAction, get_git_hash
 
@@ -115,10 +118,11 @@ def parse_args():
 
 
 def main():
+
     args = parse_args()
 
     os.environ['MASTER_ADDR'] = '127.0.0.1'  # 可以使用当前真实ip或者'127.0.0.1'
-    os.environ['MASTER_PORT'] = '29688'  # 随意一个可使用的port即可
+    os.environ['MASTER_PORT'] = '29338'  # 随意一个可使用的port即可
 
     cfg = Config.fromfile(args.config)
     if args.options is not None:
@@ -182,6 +186,8 @@ def main():
         cfg.warm_up_epochs = args.warm_up_epochs
     # weik add end
 
+
+
     # create work_dir
     mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))
     # dump config
@@ -218,6 +224,7 @@ def main():
     model = build_segmentor(
         cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)
 
+
     logger.info(model)
 
     datasets = [build_dataset(cfg.data.train)]
@@ -235,6 +242,8 @@ def main():
             PALETTE=datasets[0].PALETTE)
     # add an attribute for visualization convenience
     model.CLASSES = datasets[0].CLASSES
+    
+
     train_segmentor(
         model,
         datasets,
-- 
Gitee


From baf8e17c827e822e7cb58ec505a2dd299b7a4eb5 Mon Sep 17 00:00:00 2001
From: exias152 <15851197936@163.com>
Date: Wed, 21 Sep 2022 07:33:48 +0000
Subject: [PATCH 2/2] update
 PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md.

Signed-off-by: exias152 <15851197936@163.com>
---
 .../cv/semantic_segmentation/FCN8s/README.md  | 20 +++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md
index 4b51fbc967..6b68516deb 100644
--- a/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md
+++ b/PyTorch/contrib/cv/semantic_segmentation/FCN8s/README.md
@@ -104,7 +104,17 @@ FCN8s是一个经典的语义分割网络，FCN8s使用全卷积结构，可以
     ```none
     ├── VOCdevkit
     │   │   ├── VOC2012
+    |   |   │   │   ├── Annotations
+    |   |   │   │   ├── ImageSets
+    |   |   │   │   ├── JPEGImages
+    |   |   │   │   ├── SegmentationClass
+    |   |   │   │   ├── SegmentationObject
     │   │   ├── VOC2010
+    |   |   │   │   ├── Annotations
+    |   |   │   │   ├── ImageSets
+    |   |   │   │   ├── JPEGImages
+    |   |   │   │   ├── SegmentationClass
+    |   |   │   │   ├── SegmentationObject
     ```
 
 
@@ -116,7 +126,17 @@ FCN8s是一个经典的语义分割网络，FCN8s使用全卷积结构，可以
     ```none
     ├── VOCdevkit
     │   │   ├── VOC2012
+    |   |   │   │   ├── Annotations
+    |   |   │   │   ├── ImageSets
+    |   |   │   │   ├── JPEGImages
+    |   |   │   │   ├── SegmentationClass
+    |   |   │   │   ├── SegmentationObject
     │   │   ├── VOC2010
+    |   |   │   │   ├── Annotations
+    |   |   │   │   ├── ImageSets
+    |   |   │   │   ├── JPEGImages
+    |   |   │   │   ├── SegmentationClass
+    |   |   │   │   ├── SegmentationObject
     │   │   ├── VOCaug
     ```
 2. 数据预处理。
-- 
Gitee