Nigel
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!21095 fix network issue 

Merge pull request !21095 from JichenZhao/net_issue_fix
This commit is contained in:
i-robot 2021-08-02 08:45:40 +00:00 committed by Gitee
parent a81fee280c 434dce51b8
commit 60f66910eb
11 changed files with 90 additions and 109 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
model_zoo/official/cv/faster_rcnn/default_config_101.yaml
View File

@ -122,7 +122,7 @@ batch_size: 2
loss_scale: 256
momentum: 0.91
weight_decay: 0.00001
epoch_size: 20
epoch_size: 5
save_checkpoint: True
save_checkpoint_epochs: 1
keep_checkpoint_max: 20

2
model_zoo/official/cv/faster_rcnn/default_config_152.yaml
View File

@ -125,7 +125,7 @@ weight_decay: 0.00001
epoch_size: 20
save_checkpoint: True
save_checkpoint_epochs: 1
keep_checkpoint_max: 20
keep_checkpoint_max: 5
save_checkpoint_path: "./"
# Number of threads used to process the dataset in parallel

2
model_zoo/official/cv/yolov5/README.md
View File

@ -378,7 +378,7 @@ YOLOv5 on 118K images(The annotation and data format must be the same as coco201
| outputs | heatmaps |
| Loss | 53 |
| Speed | 1p 55 img/s 8p 440 img/s(shape=640) |
| Total time | 80h |
| Total time | 24h(8pcs) |
| Checkpoint for Fine tuning | 58M (.ckpt file) |
| Scripts | <https://gitee.com/mindspore/mindspore/tree/master/model_zoo/>|

88
model_zoo/official/cv/yolov5/README_CN.md
View File

@ -45,22 +45,22 @@ YOLOv5作为先进的检测器，它比所有可用的替代检测器更快（FP
- 目录结构如下，由用户定义目录和文件的名称：
```shell
©À©¤©¤ dataset
©À©¤©¤ YOLOv5
©À©¤©¤ annotations
©¦ ©À©¤ train.json
©¦ ©¸©¤ val.json
©À©¤ images
©À©¤ train
©¦ ©¸©¤images
©¦ ©À©¤picture1.jpg
©¦ ©À©¤ ...
©¦ ©¸©¤picturen.jpg
©¸©¤ val
©¸©¤images
©À©¤picture1.jpg
©À©¤ ...
©¸©¤picturen.jpg
├── dataset
├── YOLOv5
├── annotations
│ ├─ train.json
│ └─ val.json
├─ images
├─ train
│ └─images
│ ├─picture1.jpg
│ ├─ ...
│ └─picturen.jpg
└─ val
└─images
├─picture1.jpg
├─ ...
└─picturen.jpg
```
建议用户使用MS COCO数据集来体验模型，
@ -125,34 +125,34 @@ bash run_eval.sh dataset/xxx checkpoint/xxx.ckpt
## [脚本和示例代码](#目录)
```python
©¸©¤yolov5
©À©¤README.md
©À©¤mindspore_hub_conf.md # Mindspore Hub配置
©À©¤ascend310_infer # 用于310推理
©À©¤scripts
©À©¤run_standalone_train.sh # 在Ascend中启动单机训练（1卡）
©À©¤run_distribute_train.sh # 在Ascend中启动分布式训练（8卡）
©À©¤run_infer_310.sh # 在Ascend中启动310推理
©¸©¤run_eval.sh # 在Ascend中启动评估
©À©¤src
©À©¤__init__.py # Python初始化文件
©À©¤config.py # 参数配置
©À©¤yolov5_backbone.py # 网络骨干
©À©¤distributed_sampler.py # 数据集迭代器
©À©¤initializer.py # 参数初始化器
©À©¤logger.py # 日志函数
©À©¤loss.py # 损失函数
©À©¤lr_scheduler.py # 生成学习率
©À©¤transforms.py # 预处理数据
©À©¤util.py # 工具函数
©À©¤yolo.py # YOLOv5网络
©À©¤yolo_dataset.py # 为YOLOv5创建数据集
└─yolov5
├─README.md
├─mindspore_hub_conf.md # Mindspore Hub配置
├─ascend310_infer # 用于310推理
├─scripts
├─run_standalone_train.sh # 在Ascend中启动单机训练（1卡）
├─run_distribute_train.sh # 在Ascend中启动分布式训练（8卡）
├─run_infer_310.sh # 在Ascend中启动310推理
├─run_eval.sh # 在Ascend中启动评估
├─src
├─__init__.py # Python初始化文件
├─config.py # 参数配置
├─yolov5_backbone.py # 网络骨干
├─distributed_sampler.py # 数据集迭代器
├─initializer.py # 参数初始化器
├─logger.py # 日志函数
├─loss.py # 损失函数
├─lr_scheduler.py # 生成学习率
├─transforms.py # 预处理数据
├─util.py # 工具函数
├─yolo.py # YOLOv5网络
├─yolo_dataset.py # 为YOLOv5创建数据集
©À©¤eval.py # 评估验证结果
©À©¤export.py # 将MindSpore模型转换为AIR模型
©À©¤preprocess.py # 310推理前处理脚本
©À©¤postprocess.py # 310推理后处理脚本
©¸©¤train.py # 训练网络
├─eval.py # 评估验证结果
├─export.py # 将MindSpore模型转换为AIR模型
├─preprocess.py # 310推理前处理脚本
├─postprocess.py # 310推理后处理脚本
├─train.py # 训练网络
```
## [脚本参数](#目录)
@ -378,7 +378,7 @@ YOLOv5应用于118000张图像上（标注和数据格式必须与COCO 2017相
|输出|heatmaps |
| 损失 | 53 |
|速度| 1卡：55 img/s；8卡：440 img/s（shape=640）|
| 总时长 | 80小时 |
| 总时长 | 24小时(8卡) |
| 微调检查点 | 58M （.ckpt文件） |
|脚本| <https://gitee.com/mindspore/mindspore/tree/master/model_zoo/> |

10
model_zoo/research/cv/arcface/README_CN.md
View File

@ -55,13 +55,13 @@
```python
# 分布式训练运行示例
sh scripts/run_distribute_train.sh /path/dataset /path/rank_table
bash scripts/run_distribute_train.sh /path/dataset /path/rank_table
# 单机训练运行示例
sh scripts/run_standalone_train.sh /path/dataset
bash scripts/run_standalone_train.sh /path/dataset
# 运行评估示例
sh scripts/run_eval.sh /path/evalset /path/ckpt
bash scripts/run_eval.sh /path/evalset /path/ckpt
```
## 脚本说明
@ -108,7 +108,7 @@ train.py和val.py中主要参数如下：
### 分布式训练
```shell
sh scripts/run_distribute_train.sh /path/dataset /path/rank_table
bash scripts/run_distribute_train.sh /path/dataset /path/rank_table
```
上述shell脚本将在后台运行分布训练。可以通过`device[X]/train.log`文件查看结果。
@ -134,7 +134,7 @@ epoch time: 1104929.793 ms, per step time: 97.162 ms
在运行以下命令之前，请检查用于评估的检查点路径。请将检查点路径设置为绝对全路径，例如“username/arcface/arcface-11372-1.ckpt”。
```bash
sh scripts/run_eval.sh /path/evalset /path/ckpt
bash scripts/run_eval.sh /path/evalset /path/ckpt
```
上述python命令将在后台运行，您可以通过eval.log文件查看结果。测试数据集的准确性如下：

2
model_zoo/research/cv/arcface/scripts/run_distribute_train.sh
View File

@ -27,13 +27,13 @@ get_real_path(){
echo "$(realpath -m $PWD/$1)"
fi
}
RANK_SIZE=8
DATA_PATH=$(get_real_path $1)
RANK_TABLE=$(get_real_path $2)
EXEC_PATH=$(pwd)
echo "$EXEC_PATH"
export RANK_TABLE_FILE=$RANK_TABLE
export RANK_SIZE=8
for((i=0;i<RANK_SIZE;i++))
do

4
model_zoo/research/cv/hardnet/README_CN.md
View File

@ -89,7 +89,7 @@ HarDNet指的是Harmonic DenseNet: A low memory traffic network，其突出的
# 运行分布式训练示例
python3 train.py > train.log 2>&1 & --dataset_path /path/dataset --pre_ckpt_path /path/pretrained_path --isModelArts False
OR
bash run_distribute_train.sh /path/dataset /path/pretrain_path 8
bash run_distribute_train.sh /path/dataset /path/pretrain_path /path/rank_table
# 运行评估示例
python3 eval.py > eval.log 2>&1 & --dataset_path /path/dataset --ckpt_path /path/ckpt
@ -242,7 +242,7 @@ HarDNet指的是Harmonic DenseNet: A low memory traffic network，其突出的
```bash
python3 train.py > train.log 2>&1 & --dataset_path /path/dataset --pre_ckpt_path /path/pretrained_path --isModelArts False
OR
bash run_distribute_train.sh /path/dataset /path/pretrain_path 8
bash run_distribute_train.sh /path/dataset /path/pretrain_path /path/rank_table
```
上述shell脚本将在后台运行分布训练。您可以通过train_parallel[X]/log文件查看结果。采用以下方式达到损失值：

42
model_zoo/research/cv/hardnet/scripts/run_distribute_train.sh
View File

@ -16,40 +16,28 @@
echo "=============================================================================================================="
echo "Please run the script as: "
echo "bash run_distribute_train.sh DATA_PATH pretrain_path RANK_SIZE"
echo "For example: bash run_distribute_train.sh /path/dataset /path/pretrain_path 8"
echo "bash run_distribute_train.sh DATA_PATH pretrain_path RANK_TABLE"
echo "For example: bash run_distribute_train.sh /path/dataset /path/pretrain_path /path/rank_table"
echo "It is better to use the absolute path."
echo "=============================================================================================================="
set -e
DATA_PATH=$1
PRETRAINED_PATH=$2
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
DATA_PATH=$(get_real_path $1)
PRETRAINED_PATH=$(get_real_path $2)
RANK_TABLE=$(get_real_path $3)
export DATA_PATH=${DATA_PATH}
RANK_SIZE=$3
export RANK_SIZE=8
export RANK_TABLE_FILE=$RANK_TABLE
EXEC_PATH=$(pwd)
echo "$EXEC_PATH"
test_dist_8pcs()
{
export RANK_TABLE_FILE=${EXEC_PATH}/rank_table_8pcs.json
export RANK_SIZE=8
}
test_dist_4pcs()
{
export RANK_TABLE_FILE=${EXEC_PATH}/rank_table_4pcs.json
export RANK_SIZE=4
}
test_dist_2pcs()
{
export RANK_TABLE_FILE=${EXEC_PATH}/rank_table_2pcs.json
export RANK_SIZE=2
}
test_dist_${RANK_SIZE}pcs
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
for((i=1;i<${RANK_SIZE};i++))
@ -82,7 +70,7 @@ export DEVICE_ID=0
export RANK_ID=0
echo "start training for device 0"
env > env0.log
nohup python3 -u train.py --dataset_path ${DATA_PATH} --isModelArts False --distribute True --pre_ckpt_path ${PRETRAINED_PATH} > train0.log 2>&1
nohup python3 -u train.py --dataset_path ${DATA_PATH} --isModelArts False --distribute True --pre_ckpt_path ${PRETRAINED_PATH} > train0.log 2>&1 &
if [ $? -eq 0 ];then
echo "training success"

14
model_zoo/research/cv/simple_baselines/README.md
View File

@ -82,13 +82,13 @@ simple_baselines的总体网络架构如下：
```text
# 分布式训练
用法：sh run_distribute_train.sh --is_model_arts False --run_distribute True
用法：bash run_distribute_train.sh RANK_TABLE
# 单机训练
用法：sh run_standalone_train.sh --device_id 0 --is_model_arts False --run_distribute False
用法：bash run_standalone_train.sh DEVICE_ID
# 运行评估示例
用法：sh run_eval.sh
用法：bash run_eval.sh
```
# 脚本说明
@ -183,13 +183,13 @@ config.TEST.NMS_THRE = 1.0 # nms阈值
```text
# 分布式训练
用法：sh run_distribute_train.sh --is_model_arts False --run_distribute True
用法：bash run_distribute_train.sh RANK_TABLE
# 单机训练
用法：sh run_standalone_train.sh --device_id 0 --is_model_arts False --run_distribute False
用法：bash run_standalone_train.sh DEVICE_ID
# 运行评估示例
用法：sh run_eval.sh
用法：bash run_eval.sh
```
### 结果
@ -219,7 +219,7 @@ epoch:140 step:2340, loss is 0.0003393
```bash
# 评估
sh eval.sh
bash eval.sh
```
### 结果

31
model_zoo/research/cv/simple_baselines/scripts/run_distribute_train.sh
View File

@ -16,31 +16,24 @@
echo "========================================================================"
echo "Please run the script as: "
echo "bash run.sh RANK_SIZE"
echo "For example: bash run_distribute.sh 8"
echo "bash run.sh RANK_TABLE"
echo "For example: bash run_distribute.sh RANK_TABLE"
echo "It is better to use the absolute path."
echo "========================================================================"
set -e
RANK_SIZE=$1
export RANK_SIZE
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
RANK_TABLE=$(get_real_path $1)
EXEC_PATH=$(pwd)
echo "$EXEC_PATH"
test_dist_8pcs()
{
export RANK_TABLE_FILE=${EXEC_PATH}/rank_table_8pcs.json
export RANK_SIZE=8
}
test_dist_2pcs()
{
export RANK_TABLE_FILE=${EXEC_PATH}/rank_table_2pcs.json
export RANK_SIZE=2
}
test_dist_${RANK_SIZE}pcs
export RANK_TABLE_FILE=$RANK_TABLE
export RANK_SIZE=8
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

2
model_zoo/research/nlp/seq2seq/train.py
View File

@ -217,7 +217,7 @@ def _build_training_pipeline(config: Seq2seqConfig,
scale_update_cell=scale_manager.get_update_cell()
)
net_with_grads.set_train(True)
model = Model(net_with_grads, amp_level="O2")
model = Model(net_with_grads)
loss_monitor = LossCallBack(config)
dataset_size = dataset.get_dataset_size()
time_cb = TimeMonitor(data_size=dataset_size)