README.md 8.15 KB
Newer Older
1
2
3
4
5
# ResNet50

## 论文
`Deep Residual Learning for Image Recognition`
- https://arxiv.org/abs/1512.03385
qianyj's avatar
qianyj committed
6
7
## 模型结构
ResNet50网络中包含了49个卷积层、1个全连接层等
8
9
10
11
12
13
14
15
16

![img](./doc/ResNet50.png)
## 算法原理
ResNet50使用了多个具有残差连接的残差块来解决梯度消失或梯度爆炸问题,并使得网络可以向更深层发展。

![img](./doc/Residual_Block.png)
## 环境配置
### Docker(方法一)
```
zhanggzh's avatar
zhanggzh committed
17
docker pull image.sourcefind.cn:5000/dcu/admin/base/tensorflow:2.11.0-centos7.6-dtk23.10-py310
18
# <Your Image ID>用上面拉取docker镜像的ID替换
“qianyj”'s avatar
“qianyj” committed
19
docker run --shm-size 16g --network=host --name=resnet50_tensorFlow --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v $PWD/resnet50_tensorflow:/home/resnet50_tensorflow -it <Your Image ID> bash
“qianyj”'s avatar
“qianyj” committed
20
pip install -r requirements.txt --no-deps
qianyj's avatar
qianyj committed
21
22
wget  https://cancon.hpccube.com:65024/directlink/1/DTK-23.10/hyhal.tar.gz
tar -xzf hyhal.tar.gz -C /opt/
23
24
25
```
### Dockerfile(方法二)
```
“qianyj”'s avatar
“qianyj” committed
26
27
28
cd resnet50_tensorflow/docker
docker build --no-cache -t resnet50_tensorflow:latest .
docker run --rm --shm-size 16g --network=host --name=resnet50_tensorflow --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v $PWD/../../resnet50_tensorflow:/home/resnet50_tensorflow -it resnet50_tensorflow:latest bash
29
30
31

```
### Anaconda(方法三)
“qianyj”'s avatar
“qianyj” committed
32
1、关于本项目DCU显卡所需的特殊深度学习库可以从开发者社区下载安装:
33
34
https://developer.hpccube.com/tool/
```
qianyj's avatar
qianyj committed
35
DTK版本:dtk23.10
zhanggzh's avatar
zhanggzh committed
36
python:  3.10
zhanggezhong's avatar
zhanggezhong committed
37
tensorflow: 2.13.1
zhanggezhong's avatar
zhanggezhong committed
38
tf-models-official: 2.13.1
zhanggezhong's avatar
zhanggezhong committed
39
40
keras: 2.13.1
tensorboard: 2.13
qianyj's avatar
qianyj committed
41
hyhal
42
43
```
`Tips:以上dtk、python、tensorflow等DCU相关工具版本需要严格一一对应`
qianyj's avatar
qianyj committed
44

45
46
2、其他非特殊库参照requirements.txt安装
```
“qianyj”'s avatar
“qianyj” committed
47
pip3 install -r requirements.txt  --no-deps
48
49
```

qianyj's avatar
qianyj committed
50
## 数据集
“qianyj”'s avatar
“qianyj” committed
51
52

1、真实数据
“qianyj”'s avatar
“qianyj” committed
53

qianyj's avatar
qianyj committed
54
55
56
使用ImageNet数据集,并且需要转成TFRecord格式
ImageNet数据集可以[官网](https://image-net.org/ "ImageNet数据集官网")下载、百度搜索或者联系我们
ImageNet数据集转成TFRecord格式,可以参考以下[script](https://github.com/tensorflow/tpu/blob/master/tools/datasets/imagenet_to_gcs.py)[README](https://github.com/tensorflow/tpu/tree/master/tools/datasets#imagenet_to_gcspy)
“qianyj”'s avatar
“qianyj” committed
57
制作完成的TFRrecord数据形式如下:
“qianyj”'s avatar
“qianyj” committed
58
```
“qianyj”'s avatar
“qianyj” committed
59
60
61
tfrecord-imagenet
                | 
                train-00000-of-01024
“qianyj”'s avatar
“qianyj” committed
62
                train-00001-of-01024
“qianyj”'s avatar
“qianyj” committed
63
                ...
“qianyj”'s avatar
“qianyj” committed
64
                train-01022-of-01024
“qianyj”'s avatar
“qianyj” committed
65
66
67
68
                train-01023-of-01024
                validation-00000-of-00128
                validation-00001-of-00128
                ...
“qianyj”'s avatar
“qianyj” committed
69
                validation-00126-of-00128
“qianyj”'s avatar
“qianyj” committed
70
                validation-00127-of-00128
“qianyj”'s avatar
“qianyj” committed
71
```
“qianyj”'s avatar
“qianyj” committed
72
2、合成数据
“qianyj”'s avatar
“qianyj” committed
73

“qianyj”'s avatar
“qianyj” committed
74
基于随机合成的数据,不需要下载ImageNet数据集,执行网络训练时只需要把程序执行语句中的--use_synthetic_data设置为true即可
qianyj's avatar
qianyj committed
75
76
77
78

## 训练
### fp32训练
#### 单机单卡训练命令:
qianyj's avatar
qianyj committed
79

qianyj's avatar
qianyj committed
80
不打开xla:
qianyj's avatar
qianyj committed
81

qianyj's avatar
qianyj committed
82
    export PYTHONPATH=/home/resnet50_tensorFlow:$PYTHONPATH  
qianyj's avatar
qianyj committed
83
    python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=128 --num_gpus=1  --use_synthetic_data=false  --train_epochs=90  --dtype=fp32
qianyj's avatar
qianyj committed
84
85

打开xla:
qianyj's avatar
qianyj committed
86
87
    
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
88
    TF_XLA_FLAGS="--tf_xla_auto_jit=1" python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=128 --num_gpus=1  --use_synthetic_data=false  --train_epochs=90  --dtype=fp32
qianyj's avatar
qianyj committed
89
90
91

#### 单机四卡训练指令:
不打开xla:
qianyj's avatar
qianyj committed
92

qianyj's avatar
qianyj committed
93
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
94
    python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=512 --num_gpus=4  --use_synthetic_data=false  --train_epochs=90  --dtype=fp32
qianyj's avatar
qianyj committed
95
96

打开xla:
qianyj's avatar
qianyj committed
97

qianyj's avatar
qianyj committed
98
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
99
    TF_XLA_FLAGS="--tf_xla_auto_jit=1" python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=512 --num_gpus=4  --train_epochs=90  --use_synthetic_data=false --dtype=fp32
qianyj's avatar
qianyj committed
100
101

#### 多机多卡训练指令(以单机四卡模拟四卡四进程为例):
qianyj's avatar
qianyj committed
102

qianyj's avatar
qianyj committed
103
sed指令只需要执行一次,添加支持多卡运行的代码
qianyj's avatar
qianyj committed
104

qianyj's avatar
qianyj committed
105
    sed -i '100 r configfile' official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py
qianyj's avatar
qianyj committed
106
107

不打开xla:
qianyj's avatar
qianyj committed
108

qianyj's avatar
qianyj committed
109
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
110
    mpirun -np 4 --hostfile hostfile  -mca btl self,tcp  --allow-run-as-root  --bind-to none scripts-run/single_process.sh
qianyj's avatar
qianyj committed
111
112

打开xla:
qianyj's avatar
qianyj committed
113
114
    
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
115
    mpirun -np 4 --hostfile hostfile  -mca btl self,tcp  --allow-run-as-root  --bind-to none scripts-run/single_process_xla.sh
qianyj's avatar
qianyj committed
116
117
118
    
### fp16训练
#### 单机单卡训练指令
qianyj's avatar
qianyj committed
119

qianyj's avatar
qianyj committed
120
不打开xla:
qianyj's avatar
qianyj committed
121
   
qianyj's avatar
qianyj committed
122
    export PYTHONPATH=/home/resnet50_tensorFlow:$PYTHONPATH
qianyj's avatar
qianyj committed
123
    python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=128 --num_gpus=1  --use_synthetic_data=false --train_epochs=90  --dtype=fp16
qianyj's avatar
qianyj committed
124
125

打开xla:
qianyj's avatar
qianyj committed
126
  
qianyj's avatar
qianyj committed
127
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
128
    TF_XLA_FLAGS="--tf_xla_auto_jit=1" python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=128 --num_gpus=1  --train_epochs=90  --use_synthetic_data=false --dtype=fp16
qianyj's avatar
qianyj committed
129
130

#### 单机四卡训练指令
qianyj's avatar
qianyj committed
131

qianyj's avatar
qianyj committed
132
不打开xla:
qianyj's avatar
qianyj committed
133
  
qianyj's avatar
qianyj committed
134
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
135
    python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=512 --num_gpus=4  --train_epochs=90  --use_synthetic_data=false --dtype=fp16
qianyj's avatar
qianyj committed
136
137

打开xla:
qianyj's avatar
qianyj committed
138
139
    
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
140
    TF_XLA_FLAGS="--tf_xla_auto_jit=1" python3 official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py --data_dir=/path/to/{ImageNet-tensorflow_data_dir} --model_dir=/path/to/{model_save_dir} --batch_size=512 --num_gpus=4  --train_epochs=90  --use_synthetic_data=false --dtype=fp16
qianyj's avatar
qianyj committed
141
142

#### 多机多卡训练指令(以单机四卡模拟四卡四进程为例)
qianyj's avatar
qianyj committed
143

qianyj's avatar
qianyj committed
144
sed指令只需要执行一次,添加支持多卡运行的代码
qianyj's avatar
qianyj committed
145
    
qianyj's avatar
qianyj committed
146
    sed -i '100 r configfile' official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py
qianyj's avatar
qianyj committed
147

qianyj's avatar
qianyj committed
148
149
150
修改scripts-run/single_process.sh和scripts-run/single_process_xla.sh文件里的--dtype=fp16

不打开xla:
qianyj's avatar
qianyj committed
151

qianyj's avatar
qianyj committed
152
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
153
154
    mpirun -np 4 --hostfile hostfile  -mca btl self,tcp  --allow-run-as-root  --bind-to none scripts-run/single_process.sh

qianyj's avatar
qianyj committed
155
打开xla:
qianyj's avatar
qianyj committed
156
157
 
    export PYTHONPATH=/home/resnet50_tensorflow:$PYTHONPATH
qianyj's avatar
qianyj committed
158
    mpirun -np 4 --hostfile hostfile  -mca btl self,tcp  --allow-run-as-root  --bind-to none scripts-run/single_process_xla.sh
qianyj's avatar
qianyj committed
159

qianyj's avatar
qianyj committed
160
161
162
### result
![img](./doc/ILSVRC2012_val_00001915.PNG)
![img](./doc/ILSVRC2012_val_00003386.PNG)
qianyj's avatar
qianyj committed
163

qianyj's avatar
qianyj committed
164
## 精度
qianyj's avatar
qianyj committed
165
测试数据:[ImageNet的测试数据集](https://image-net.org/ "ImageNet数据集官网"),使用的加速卡:DCU-Z100-16G
qianyj's avatar
qianyj committed
166

167
168
| 卡数 | batch size | 类型 |  Accuracy | 是否打开xla | 进程数 |
| :------: | :------: |  :------: | :------: | :------:| -------- |
qianyj's avatar
qianyj committed
169
170
171
172
| 4 | 512 | fp32 |  0.763  | 否 | 单进程 |
| 4 | 512 | fp16 |  0.764  | 否 | 单进程 |
| 4 | 512 | fp32 |  0.764  | 否 | 四进程 |
| 4 | 512 | fp16 |  0.763  | 否 | 四进程 |
173

“qianyj”'s avatar
“qianyj” committed
174
175
176
177
178
179
## 应用场景
### 算法类别
`图像分类`
### 热点应用行业
`制造,政府,医疗,科研`

180
## 源码仓库及问题反馈
181
182
* https://developer.hpccube.com/codes/modelzoo/resnet50_tensorflow

qianyj's avatar
qianyj committed
183
184
185
## 参考
* https://github.com/tensorflow/models/tree/master
* https://www.tensorflow.org/api_docs/python/tf/distribute/MultiWorkerMirroredStrategy