mindspore-Ecosystem
/
mindspore
21
0
Fork 27
Code Issues Pull Requests 6 Projects Releases Wiki Activity

!20491 add gpu scripts to tinydarknet 

Merge pull request !20491 from ZeyangGAO/code_docs_tdngpu
This commit is contained in:
i-robot 2021-07-19 12:11:49 +00:00 committed by Gitee
parent a993f5a46a d6a9848b23
commit a5b29dc4c9
17 changed files with 616 additions and 88 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

156
model_zoo/official/cv/tinydarknet/README.md
View File

@ -60,8 +60,8 @@ Dataset used can refer to [paper](<https://ieeexplore.ieee.org/abstract/document
# [Environment Requirements](#contents)
- Hardware（Ascend/CPU）
- Prepare hardware environment with Ascend/CPU processor.
- Hardware（Ascend/CPU/GPU）
- Prepare hardware environment with Ascend/CPU processor/GPU.
- Framework
- [MindSpore](https://www.mindspore.cn/install/en)
- For more information,please check the resources below：
@ -93,6 +93,35 @@ After installing MindSpore via the official website, you can start training and
<https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools.>
- running on GPU with gpu default parameters
```python
# GPU standalone training example
python train.py \
--config_path=./imagenet_config_gpu.yaml \
--dataset_name=imagenet --train_data_dir=../dataset/imagenet_original/train --device_target=GPU
OR
cd scripts
bash run_distribute_train_gpu.sh [DEVICE_ID] [TRAIN_DATA_DIR] [cifar10 | imagenet]
# GPU distribute training example
export RANK_SIZE=8
mpirun --allow-run-as-root -n $RANK_SIZE --output-filename log_output --merge-stderr-to-stdout \
python train.py \
--config_path=./config/imagenet_config_gpu.yaml \
--dataset_name=imagenet \
--train_data_dir=../dataset/imagenet_original/train \
--device_target=GPU
OR
bash scripts/run_distribute_train_gpu.sh [RANK_SIZE] [TRAIN_DATA_DIR] [cifar10 | imagenet]
# GPU evaluation example
python eval.py -device_target=GPU --val_data_dir=../dataset/imagenet_original/val --dataset_name=imagenet --config_path=./config/imagenet_config_gpu.yaml \
--checkpoint_path=$PATH2
OR
bash scripts/run_train_gpu.sh [VAL_DATA_DIR] [cifar10|imagenet] [checkpoint_path]
```
- Running on ModelArts
If you want to run in modelarts, please check the official documentation of [modelarts](https://support.huaweicloud.com/modelarts/), and you can start training as follows.
@ -155,12 +184,20 @@ For more details, please refer the specify script.
├── README.md // descriptions about Tiny-Darknet in English
├── README_CN.md // descriptions about Tiny-Darknet in Chinese
├── ascend310_infer // application for 310 inference
├── src
├── imagenet_config.yaml // imagenet parameter configuration
├── imagenet_config_gpu.yaml // imagenet parameter configuration for GPU
├── cifar10_config.yaml // cifar10 parameter configuration
├── cifar10_config_gpu.yaml // cifar10 parameter configuration for GPU
├── scripts
├── run_standalone_train.sh // shell script for single on Ascend
├── run_standalone_train_gpu.sh // shell script for single on GPU
├── run_distribute_train.sh // shell script for distributed on Ascend
├── run_distribute_train_gpu.sh // shell script for distributed on GPU
├── run_train_cpu.sh // shell script for distributed on CPU
├── run_eval.sh // shell script for evaluation on Ascend
├── run_eval_cpu.sh // shell script for evaluation on CPU
├── run_eval_gpu.sh // shell script for evaluation on GPU
├── run_infer_310.sh // shell script for inference on Ascend310
├── src
├── lr_scheduler //learning rate scheduler
@ -179,8 +216,6 @@ For more details, please refer the specify script.
├── train.py // training script
├── eval.py // evaluation script
├── export.py // export checkpoint file into air/onnx
├── imagenet_config.yaml // imagenet parameter configuration
├── cifar10_config.yaml // cifar10 parameter configuration
├── mindspore_hub_conf.py // hub config
├── postprocess.py // postprocess script
@ -252,6 +287,29 @@ For more configuration details, please refer the script `imagenet_config.yaml`.
The model checkpoint file will be saved in the current folder.
<!-- The model checkpoint will be saved in the current directory. -->
- running on GPU：
```python
cd scripts
bash run_standalone_train_gpu.sh [DEVICE_ID] [TRAIN_DATA_DIR] [cifar10|imagenet]
```
The command above will run in the background, you can view the results through the file train.log.
After training, you'll get some checkpoint files under the script folder by default. The loss value will be achieved as follows:
<!-- After training, you'll get some checkpoint files under the script folder by default. The loss value will be achieved as follows: -->
```python
# grep "loss is " train.log
epoch: 498 step: 1251, loss is 2.7798953
Epoch time: 130690.544, per step time: 104.469
epoch: 499 step: 1251, loss is 2.9261637
Epoch time: 130511.081, per step time: 104.325
epoch: 500 step: 1251, loss is 2.69412
Epoch time: 127067.548, per step time: 101.573
...
```
- running on CPU
```python
@ -279,6 +337,25 @@ For more configuration details, please refer the script `imagenet_config.yaml`.
...
```
- running on GPU：
```python
bash scripts/run_distribute_train_gpu.sh [RANK_SIZE] [TRAIN_DATA_DIR] [cifar10|imagenet]
```
The above shell script will run distribute training in the background. You can view the results through the file train_parallel[X]/log. The loss value will be achieved as follows:
```python
# grep "result: " distribute_train_gpu/nohup.out
epoch: 498 step: 1251, loss is 2.7825122
epoch time: 200066.210 ms, per step time: 159.925 ms
epoch: 499 step: 1251, loss is 2.799798
epoch time: 199098.258 ms, per step time: 159.151 ms
epoch: 500 step: 1251, loss is 2.8718748
epoch time: 197784.661 ms, per step time: 158.101 ms
...
```
## [Evaluation Process](#contents)
### [Evaluation](#contents)
@ -307,6 +384,28 @@ For more configuration details, please refer the script `imagenet_config.yaml`.
accuracy: {'top_1_accuracy': 0.5871979166666667, 'top_5_accuracy': 0.8175280448717949}
```
- evaluation on Imagenet dataset when running on GPU:
Before running the command below, please check the checkpoint path used for evaluation. Please set the checkpoint path to be the absolute full path, e.g., "/username/tinydaeknet/train_tinydarknet.ckpt".
```python
bash scripts/run_train_gpu.sh [VAL_DATA_DIR] [cifar10|imagenet] [checkpoint_path]
```
The above python command will run in the background. You can view the results through the file "eval.log". The accuracy of the test dataset will be as follows:
```python
# grep "accuracy: " eval.log
accuracy: {'top_1_accuracy': 0.5896033653846153, 'top_5_accuracy': 0.8176482371794872}
```
Note that for evaluation after distributed training, please set the checkpoint_path to be the last saved checkpoint file. The accuracy of the test dataset will be as follows:
```python
# grep "accuracy: " eval.log
accuracy: {'top_1_accuracy': 0.5896033653846153, 'top_5_accuracy': 0.8176482371794872}
```
- evaluation on cifar-10 dataset when running on CPU:
Before running the command below, please check the checkpoint path used for evaluation. Please set the checkpoint path to be the absolute full path, e.g., "/username/tinydaeknet/train_tinydarknet.ckpt".
@ -389,34 +488,33 @@ Inference result is saved in current path, you can find result like this in acc.
### [Training Performance](#contents)
| Parameters | Ascend |
| -------------------------- | ----------------------------------------------------------- |
| Model Version | V1 |
| Resource | Ascend 910; CPU 2.60GHz, 56cores; Memory 314G; OS Euler2.8 |
| Uploaded Date | 2020/12/22 |
| MindSpore Version | 1.1.0 |
| Dataset | 1200k images |
| Training Parameters | epoch=500, steps=1251, batch_size=128, lr=0.1 |
| Optimizer | Momentum |
| Loss Function | Softmax Cross Entropy |
| Speed | 8 pc: 104 ms/step |
| Total Time | 8 pc: 17.8 hours |
| Parameters(M) | 4.0M |
| Scripts | [Tiny-Darknet Scripts](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/tinydarknet) |
| Parameters | Ascend | GPU |
| -------------------------- | ------------------------------------------------------------| ----------------------------------------------------|
| Model Version | V1 | V1 |
| Resource | Ascend 910；CPU 2.60GHz，56cores；内存 314G；系统 Euler2.8 | PCIE V100-32G |
| Uploaded Date | 2020/12/22 | 2021/07/15 |
| MindSpore Version | 1.1.0 | 1.3.0 |
| Dataset | 1200k images | 1200k images |
| Training Parameters | epoch=500, steps=1251, batch_size=128, lr=0.1 | epoch=500, steps=1251, batch_size = 128, lr=0.005 |
| Optimizer | Momentum | Momentum |
| Loss Function | Softmax Cross Entropy | Softmax Cross Entropy |
| Speed | 8pc: 104 ms/step | 8pc: 255 ms/step |
| Parameters(M) | 4.0; | 4.0; |
| Scripts | [Tiny-Darknet scripts](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/tinydarknet)
### [Evaluation Performance](#contents)
| Parameters | Ascend |
| ------------------- | --------------------------- |
| Model Version | V1 |
| Resource | Ascend 910; OS Euler2.8 |
| Uploaded Date | 2020/12/22 |
| MindSpore Version | 1.1.0 |
| Dataset | 200k images |
| batch_size | 128 |
| Outputs | probability |
| Accuracy | 8 pc Top-1: 58.7%; Top-5: 81.7% |
| Model for inference | 11.6M (.ckpt file) |
| Parameters | Ascend | GPU |
| ------------------- | ----------------------------------| ----------------------------------|
| Model Version | V1 | V1 |
| Resource | Ascend 910；Euler2.8 | PCIE V100-32G |
| Uploaded Date | 2020/12/22 | 2021/7/15 |
| MindSpore Version | 1.1.0 | 1.3.0 |
| Dataset | 200k images | 200k images |
| batch_size | 128 | 128 |
| Outputs | probability | probability |
| Accuracy | 8pcs Top-1: 58.7%; Top-5: 81.7% | 8pcs Top-1: 58.9%; Top-5: 81.7% |
| Model for inference | 11.6M (.ckpt file) | 10.06M (.ckpt file) |
### [Inference Performance](#contents)

177
model_zoo/official/cv/tinydarknet/README_CN.md
View File

@ -68,8 +68,8 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的
# [环境要求](#目录)
- 硬件（Ascend/CPU）
- 请准备具有Ascend/CPU处理器的硬件环境.
- 硬件（Ascend/CPU/GPU）
- 请准备具有Ascend/CPU处理器/GPU的硬件环境.
- 框架
- [MindSpore](https://www.mindspore.cn/install)
- 更多的信息请访问以下链接：
@ -101,6 +101,35 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的
<https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools.>
- running on GPU with gpu default parameters
```python
# GPU单卡训练示例
python train.py \
--config_path=./config/imagenet_config_gpu.yaml \
--dataset_name=imagenet --train_data_dir=../dataset/imagenet_original/train --device_target=GPU
OR
cd scripts
bash run_distribute_train_gpu.sh [DEVICE_ID] [TRAIN_DATA_DIR] [cifar10 | imagenet]
# GPU多卡训练示例
export RANK_SIZE=8
mpirun --allow-run-as-root -n $RANK_SIZE --output-filename log_output --merge-stderr-to-stdout \
python train.py \
--config_path=./config/imagenet_config_gpu.yaml \
--dataset_name=imagenet \
--train_data_dir=../dataset/imagenet_original/train \
--device_target=GPU
OR
bash scripts/run_distribute_train_gpu.sh [RANK_SIZE] [TRAIN_DATA_DIR] [cifar10 | imagenet]
# GPU评估示例
python eval.py -device_target=GPU --val_data_dir=../dataset/imagenet_original/val --dataset_name=imagenet --config_path=./config/imagenet_config_gpu.yaml \
--checkpoint_path=$PATH2
OR
bash scripts/run_train_gpu.sh [VAL_DATA_DIR] [cifar10|imagenet] [checkpoint_path]
```
- 在ModelArts上运行
如果你想在modelarts上运行，可以参考以下文档 [modelarts](https://support.huaweicloud.com/modelarts/)
@ -162,12 +191,20 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的
├── README.md // Tiny-Darknet英文说明
├── README_CN.md // Tiny-Darknet中文说明
├── ascend310_infer // 用于310推理
├── config
├── imagenet_config.yaml // imagenet参数配置
├── imagenet_config_gpu.yaml // imagenet参数配置
├── cifar10_config.yaml // cifar10参数配置
├── cifar10_config_gpu.yaml // cifar10参数配置
├── scripts
├── run_standalone_train.sh // Ascend单卡训练shell脚本
├── run_standalone_train_gpu.sh // GPU单卡训练shell脚本
├── run_distribute_train.sh // Ascend分布式训练shell脚本
├── run_distribute_train_gpu.sh // GPU分布式训练shell脚本
├── run_train_cpu.sh // CPU训练shell脚本
├── run_eval.sh // Ascend评估shell脚本
├── run_eval_cpu.sh // CPU评估shell脚本
├── run_eval_gpu.sh // GPU评估shell脚本
└── run_infer_310.sh // Ascend310推理shell脚本
├── src
├── lr_scheduler // 学习率策略
@ -186,8 +223,6 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的
├── train.py // 训练脚本
├── eval.py // 评估脚本
├── export.py // 导出checkpoint文件
├── imagenet_config.yaml // imagenet参数配置
├── cifar10_config.yaml // cifar10参数配置
├── mindspore_hub_conf.py // hub配置文件
└── postprocess.py // 310推理后处理脚本
@ -259,6 +294,29 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的
模型checkpoint文件将会保存在当前文件夹下.
<!-- The model checkpoint will be saved in the current directory. -->
- 在GPU资源上运行：
```python
cd scripts
bash run_standalone_train_gpu.sh [DEVICE_ID] [TRAIN_DATA_DIR] [cifar10|imagenet]
```
上述的命令将运行在后台中，可以通过 `train_single_gpu/train.log` 文件查看运行结果.
训练完成后,默认情况下,可在script文件夹下得到一些checkpoint文件. 训练的损失值将以如下的形式展示:
<!-- After training, you'll get some checkpoint files under the script folder by default. The loss value will be achieved as follows: -->
```python
# grep "loss is " train.log
epoch: 498 step: 1251, loss is 2.7798953
Epoch time: 130690.544, per step time: 104.469
epoch: 499 step: 1251, loss is 2.9261637
Epoch time: 130511.081, per step time: 104.325
epoch: 500 step: 1251, loss is 2.69412
Epoch time: 127067.548, per step time: 101.573
...
```
- 在CPU资源上运行：
```python
@ -273,16 +331,35 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的
bash scripts/run_distribute_train.sh [RANK_TABLE_FILE]
```
上述的脚本命令将在后台中进行分布式训练，可以通过`train_parallel[X]/log`文件查看运行结果. 训练的损失值将以如下的形式展示:
上述的脚本命令将在后台中进行分布式训练，可以通过`distribute_train/nohup.out`文件查看运行结果. 训练的损失值将以如下的形式展示:
```python
# grep "result: " train_parallel*/log
epoch: 498 step: 1251, loss is 2.7798953
Epoch time: 130690.544, per step time: 104.469
epoch: 499 step: 1251, loss is 2.9261637
Epoch time: 130511.081, per step time: 104.325
epoch: 500 step: 1251, loss is 2.69412
Epoch time: 127067.548, per step time: 101.573
# grep "result: " distribute_train/nohup.out
epoch: 498 step: 1251, loss is 2.7825122
epoch time: 200066.210 ms, per step time: 159.925 ms
epoch: 499 step: 1251, loss is 2.799798
epoch time: 199098.258 ms, per step time: 159.151 ms
epoch: 500 step: 1251, loss is 2.8718748
epoch time: 197784.661 ms, per step time: 158.101 ms
...
```
- 在GPU资源上运行：
```python
bash scripts/run_distribute_train_gpu.sh [RANK_SIZE] [TRAIN_DATA_DIR] [cifar10|imagenet]
```
上述的脚本命令将在后台中进行分布式训练，可以通过`distribute_train_gpu/nohup.out`文件查看运行结果. 训练的损失值将以如下的形式展示:
```python
# grep "result: " distribute_train_gpu/nohup.out
epoch: 498 step: 1251, loss is 2.7825122
epoch time: 200066.210 ms, per step time: 159.925 ms
epoch: 499 step: 1251, loss is 2.799798
epoch time: 199098.258 ms, per step time: 159.151 ms
epoch: 500 step: 1251, loss is 2.8718748
epoch time: 197784.661 ms, per step time: 158.101 ms
...
```
@ -314,12 +391,34 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的
accuracy: {'top_1_accuracy': 0.5871979166666667, 'top_5_accuracy': 0.8175280448717949}
```
- 在GPU资源上进行评估:
在运行如下命令前,请确认用于评估的checkpoint文件的路径.checkpoint文件须包含在tinydarknet文件夹内.请将checkpoint路径设置为相对于 eval.py文件 的路径,例如:"./ckpts/train_tinydarknet.ckpt"(ckpts 与 eval.py 同级).
```python
bash scripts/run_train_gpu.sh [VAL_DATA_DIR] [cifar10|imagenet] [checkpoint_path]
```
上述的python命令将运行在后台中，可以通过"eval.log"文件查看结果. 测试数据集的准确率将如下面所列:
```python
# grep "accuracy: " eval.log
accuracy: {'top_1_accuracy': 0.5896033653846153, 'top_5_accuracy': 0.8176482371794872}
```
请注意在并行训练后,测试请将checkpoint_path设置为最后保存的checkpoint文件的路径,准确率将如下面所列:
```python
# grep "accuracy: " eval.log
accuracy: {'top_1_accuracy': 0.5896033653846153, 'top_5_accuracy': 0.8176482371794872}
```
- 在CPU资源上进行评估
在运行如下命令前,请确认用于评估的checkpoint文件的路径.checkpoint文件须包含在tinydarknet文件夹内.请将checkpoint路径设置为相对于 eval.py文件 的路径,例如:"./ckpts/train_tinydarknet.ckpt"(ckpts 与 eval.py 同级).
```python
bash scripts/run_eval.sh [VAL_DATA_DIR] [imagenet|cifar10] [CHECKPOINT_PATH]
bash scripts/run_eval_cpu.sh [VAL_DATA_DIR] [imagenet|cifar10] [CHECKPOINT_PATH]
```
可以通过"eval.log"文件查看结果. 测试数据集的准确率将如下面所列:
@ -395,34 +494,36 @@ bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [LABEL_PATH] [DVPP] [DEVICE_ID]
### [训练性能](#目录)
| 参数 | Ascend |
| -------------------------- | ----------------------------------------------------------- |
| 模型版本 | V1 |
| 资源 | Ascend 910；CPU 2.60GHz，56cores；内存 314G；系统 Euler2.8 |
| 上传日期 | 2020/12/22 |
| MindSpore版本 | 1.1.0 |
| 数据集 | 1200k张图片 |
| 训练参数 | epoch=500, steps=1251, batch_size=128, lr=0.1 |
| 优化器 | Momentum |
| 损失函数 | Softmax Cross Entropy |
| 速度 | 8卡: 104 ms/step |
| 总时间 | 8卡: 17.8小时 |
| 参数(M) | 4.0 |
| 脚本 | [Tiny-Darknet脚本](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/tinydarknet) |
#### Tinydarknet on ImageNet 2012
| 参数 | Ascend | GPU |
| -------------------------- | ------------------------------------------------------------| ----------------------------------------------------|
| 模型版本 | V1 | V1 |
| 资源 | Ascend 910；CPU 2.60GHz，56cores；内存 314G；系统 Euler2.8 | PCIE V100-32G |
| 上传日期 | 2020/12/22 | 2021/07/15 |
| MindSpore版本 | 1.1.0 | 1.3.0 |
| 数据集 | 1200k张图片 | 1200k张图片 |
| 训练参数 | epoch=500, steps=1251, batch_size=128, lr=0.1 | epoch=500, steps=1251, batch_size = 128, lr=0.005 |
| 优化器 | Momentum | Momentum |
| 损失函数 | Softmax Cross Entropy | Softmax Cross Entropy |
| 速度 | 8卡: 104 ms/step | 8卡: 255 ms/step |
| 总时间 | 8卡: 17.8小时 | 8卡: 46.9小时 |
| 参数(M) | 4.0; | 4.0; |
| 脚本 | [Tiny-Darknet脚本](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/tinydarknet)
### [评估性能](#目录)
| 参数 | Ascend |
| ------------------- | --------------------------- |
| 模型版本 | V1 |
| 资源 | Ascend 910；系统 Euler2.8 |
| 上传日期 | 2020/12/22 |
| MindSpore版本 | 1.1.0 |
| 数据集 | 200k张图片 |
| batch_size | 128 |
| 输出 | 分类概率 |
| 准确率 | 8卡 Top-1: 58.7%; Top-5: 81.7% |
| 推理模型 | 11.6M (.ckpt文件) |
| 参数 | Ascend | GPU |
| ------------------- | ----------------------------------| ----------------------------------|
| 模型版本 | V1 | V1 |
| 资源 | Ascend 910；系统 Euler2.8 | NV SMX2 V100-32G |
| 上传日期 | 2020/12/22 | 2021/7/15 |
| MindSpore版本 | 1.1.0 | 1.3.0 |
| 数据集 | 200k张图片 | 200k张图片 |
| batch_size | 128 | 128 |
| 输出 | 分类概率 | 分类概率 |
| 准确率 | 8卡 Top-1: 58.7%; Top-5: 81.7% | 8卡 Top-1: 58.9%; Top-5: 81.7% |
| 推理模型 | 11.6M (.ckpt文件) | 10.06M (.ckpt文件) |
### [推理性能](#目录)

0
model_zoo/official/cv/tinydarknet/cifar10_config.yaml → model_zoo/official/cv/tinydarknet/config/cifar10_config.yaml
View File

57
model_zoo/official/cv/tinydarknet/config/cifar10_config_gpu.yaml Normal file
View File

@ -0,0 +1,57 @@
# Builtin Configurations(DO NOT CHANGE THESE CONFIGURATIONS unless you know exactly what you are doing)
enable_modelarts: False
# Url for modelarts
data_url: ""
train_url: ""
checkpoint_url: ""
# Path for local
data_path: "/cache/data"
output_path: "/cache/train"
load_path: "/cache/checkpoint_path"
device_target: "GPU"
enable_profiling: False
modelarts_dataset_unzip_name: ''
# ==============================================================================
#train-eval-export related
dataset_name : cifar10
ckpt_save_dir: checkpoints
pre_trained: False
device_id: 0
num_classes: 10
lr_init: 0.1
batch_size: 32
epoch_size: 120
momentum: 0.9
weight_decay: 0.0001
image_height: 227
image_width: 227
train_data_dir: './data/cifar10_train/'
val_data_dir: './data/cifar10_val/'
keep_checkpoint_max: 1
checkpoint_path: './scripts/train_parallel4/ckpt_4/train_tinydarknet_imagenet-300_1251.ckpt'
onnx_filename: 'tinydarknet.onnx'
air_filename: 'tinydarknet.air'
# optimizer and lr related
lr_scheduler: 'exponential'
lr_epochs: [70, 140, 210, 280]
lr_gamma: 0.1
eta_min: 0.0
T_max: 150
warmup_epochs: 0
# loss related
is_dynamic_loss_scale: False
loss_scale: 1024
label_smooth_factor: 0.1
use_label_smooth: True
---
# Help description for each configuration
enable_modelarts: "Whether training on modelarts, default: False"
data_url: "Url for modelarts"
train_url: "Url for modelarts"
data_path: "The location of the input data."
output_path: "The location of the output file."
device_target: "Running platform, choose from Ascend, GPU or CPU, and default is Ascend."
enable_profiling: 'Whether enable profiling while training, default: False'

0
model_zoo/official/cv/tinydarknet/imagenet_config.yaml → model_zoo/official/cv/tinydarknet/config/imagenet_config.yaml
View File

61
model_zoo/official/cv/tinydarknet/config/imagenet_config_gpu.yaml Normal file
View File

@ -0,0 +1,61 @@
# Builtin Configurations(DO NOT CHANGE THESE CONFIGURATIONS unless you know exactly what you are doing)
enable_modelarts: False
# Url for modelarts
data_url: ""
train_url: ""
checkpoint_url: ""
# Path for local
data_path: "/cache/data"
output_path: "/cache/train"
load_path: "/cache/checkpoint_path"
device_target: "GPU"
enable_profiling: False
modelarts_dataset_unzip_name: ''
# ==============================================================================
#train-eval-export related
dataset_name: imagenet
ckpt_save_dir: checkpoints
pre_trained: False
device_id: 0
num_classes: 1000
lr_init: 0.1
batch_size: 128
epoch_size: 500
momentum: 0.9
weight_decay: 0.0001
image_height: 224
image_width: 224
train_data_dir: './dataset/imagenet_original/train/'
val_data_dir: './dataset/imagenet_original/val/'
keep_checkpoint_max: 1
checkpoint_path: './scripts/train_parallel4/ckpt_4/train_tinydarknet_imagenet-300_1251.ckpt'
file_name: 'tinydarknet'
file_format: 'MINDIR'
# optimizer and lr related
lr_scheduler: 'exponential'
lr_epochs: [70, 140, 210, 280]
lr_gamma: 0.3
eta_min: 0.0
T_max: 150
warmup_epochs: 0
# loss related
is_dynamic_loss_scale: False
loss_scale: 1024
label_smooth_factor: 0.1
use_label_smooth: True
#310infer postprocess
result_path: ''
label_file: ''
---
# Help description for each configuration
enable_modelarts: "Whether training on modelarts, default: False"
data_url: "Url for modelarts"
train_url: "Url for modelarts"
data_path: "The location of the input data."
output_path: "The location of the output file."
device_target: "Running platform, choose from Ascend, GPU or CPU, and default is Ascend."
enable_profiling: 'Whether enable profiling while training, default: False'
file_format: '["MINDIR", "AIR"]'

2
model_zoo/official/cv/tinydarknet/scripts/run_distribute_train.sh
View File

@ -57,7 +57,7 @@ do
mkdir ./train_parallel$i
cp -r ../src ./train_parallel$i
cp ../train.py ./train_parallel$i
cp ../*.yaml ./train_parallel$i
cp -r ../config ./train_parallel$i
echo "start training for rank $RANK_ID, device $DEVICE_ID, $dataset_type"
cd ./train_parallel$i || exit
env > env.log

75
model_zoo/official/cv/tinydarknet/scripts/run_distribute_train_gpu.sh Normal file
View File

@ -0,0 +1,75 @@
#!/usr/bin/env bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# 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.
# ============================================================================
if [ $# != 3 ]; then
echo "Usage: sh run_distribute_train_gpu.sh [RANK_SIZE] [TRAIN_DATA_DIR] [cifar10|imagenet]"
exit 1
fi
get_real_path() {
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
dataset_type='imagenet'
if [ $# == 3 ]
then
if [ $3 != "cifar10" ] && [ $3 != "imagenet" ]
then
echo "error: the selected dataset is neither cifar10 nor imagenet"
exit 1
fi
dataset_type=$3
fi
export RANK_SIZE=$1
PROJECT_DIR=$(cd ./"`dirname $0`" || exit; pwd)
TRAIN_DATA_DIR=$(get_real_path $2)
if [ ! -d $TRAIN_DATA_DIR ]; then
echo "error: TRAIN_DATA_DIR=$TRAIN_DATA_DIR is not a directory"
exit 1
fi
if [ -d "distribute_train_gpu" ]; then
rm -rf ./distribute_train_gpu
fi
mkdir ./distribute_train_gpu
cp ./*.py ./distribute_train_gpu
cp -r ./config ./distribute_train_gpu
cp -r ./src ./distribute_train_gpu
cd ./distribute_train_gpu || exit
if [ $3 == 'imagenet' ]; then
CONFIG_FILE="$PROJECT_DIR/../config/imagenet_config_gpu.yaml"
elif [ $3 == 'cifar10' ]; then
CONFIG_FILE="$PROJECT_DIR/../config/cifar10_config_gpu.yaml"
else
echo "error: the selected dataset is neither cifar10 nor imagenet"
exit 1
fi
mpirun --allow-run-as-root -n $RANK_SIZE --output-filename log_output --merge-stderr-to-stdout \
nohup python train.py \
--config_path=$CONFIG_FILE \
--dataset_name=$dataset_type \
--train_data_dir=$TRAIN_DATA_DIR \
--device_target=GPU > log.txt 2>&1 &
cd ..

2
model_zoo/official/cv/tinydarknet/scripts/run_eval.sh
View File

@ -22,7 +22,7 @@ rm -rf ./eval
mkdir ./eval
cp -r ../src ./eval
cp ../eval.py ./eval
cp ../*.yaml ./eval
cp -r ../config ./eval
cd ./eval || exit
env >env.log
python ./eval.py > ./eval.log 2>&1 &

6
model_zoo/official/cv/tinydarknet/scripts/run_eval_cpu.sh
View File

@ -43,9 +43,9 @@ fi
BASE_PATH=$(dirname "$(dirname "$(readlink -f $0)")")
if [ $2 == 'imagenet' ]; then
CONFIG_FILE="${BASE_PATH}/imagenet_config.yaml"
CONFIG_FILE="${BASE_PATH}/config/imagenet_config.yaml"
elif [ $2 == 'cifar10' ]; then
CONFIG_FILE="${BASE_PATH}/cifar10_config.yaml"
CONFIG_FILE="${BASE_PATH}/config/cifar10_config.yaml"
else
echo "error: the selected dataset is neither cifar10 nor imagenet"
exit 1
@ -55,7 +55,7 @@ rm -rf ./eval
mkdir ./eval
cp -r ./src ./eval
cp ./eval.py ./eval
cp ./*.yaml ./eval
cp -r ./config ./eval
env >env.log
echo "start evaluation for device CPU"
cd ./eval || exit

64
model_zoo/official/cv/tinydarknet/scripts/run_eval_gpu.sh Normal file
View File

@ -0,0 +1,64 @@
#!/usr/bin/env bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# 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.
# ============================================================================
if [ $# != 1 ] && [ $# != 2 ] && [ $# != 3 ]
then
echo "Usage bash scripts/run_train_gpu.sh [VAL_DATA_DIR] [cifar10|imagenet] [checkpoint_path]"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
PATH1=$(get_real_path $1)
if [ ! -d $PATH1 ]
then
echo "error: VAL_DATA_DIR=$PATH1 is not a directory"
exit 1
fi
PATH2=$(get_real_path $3)
if [ ! -f $PATH2 ]
then
echo "error: CHECKPOINT_PATH=$PATH2 is not a file"
exit 1
fi
BASE_PATH=$(dirname "$(dirname "$(readlink -f $0)")")
if [ $2 == 'imagenet' ]; then
CONFIG_FILE="${BASE_PATH}/config/imagenet_config_gpu.yaml"
elif [ $2 == 'cifar10' ]; then
CONFIG_FILE="${BASE_PATH}/config/cifar10_config_gpu.yaml"
else
echo "error: the selected dataset is neither cifar10 nor imagenet"
exit 1
fi
rm -rf ./eval
mkdir ./eval
cp -r ./src ./eval
cp ./eval.py ./eval
cp -r ./config ./eval
env >env.log
echo "start evaluation for device GPU"
cd ./eval || exit
python ./eval.py --device_target=GPU --val_data_dir=$PATH1 --dataset_name=$2 --config_path=$CONFIG_FILE \
--checkpoint_path=$PATH2 > ./eval.log 2>&1 &
cd ..

2
model_zoo/official/cv/tinydarknet/scripts/run_standalone_train.sh
View File

@ -55,7 +55,7 @@ rm -rf ./train_single
mkdir ./train_single
cp -r ../src ./train_single
cp ../train.py ./train_single
cp ../*.yaml ./train_single
cp -r ../config ./train_single
echo "start training for rank $RANK_ID, device $DEVICE_ID, $dataset_type"
cd ./train_single || exit
python ./train.py --dataset_name=$dataset_type --train_data_dir=$train_data_dir> ./train.log 2>&1 &

73
model_zoo/official/cv/tinydarknet/scripts/run_standalone_train_gpu.sh Normal file
View File

@ -0,0 +1,73 @@
#!/usr/bin/env bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# 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.
# ============================================================================
echo "$1 $2 $3"
if [ $# != 2 ] && [ $# != 3 ]
then
echo "Usage: bash run_distribute_train_gpu.sh [DEVICE_ID] [TRAIN_DATA_DIR] [cifar10|imagenet]"
exit 1
fi
expr $1 + 6 &>/dev/null
if [ $? != 0 ]
then
echo "error:DEVICE_ID=$1 is not a integer"
exit 1
fi
if [ ! -d $2 ]
then
echo "error:TRAIN_DATA_DIR=$2 is not a folder"
exit 1
fi
train_data_dir=$2
PROJECT_DIR=$(cd ./"`dirname $0`" || exit; pwd)
CONFIG_FILE="$PROJECT_DIR/../config/imagenet_config_gpu.yaml"
dataset_type='imagenet'
if [ $# == 3 ]
then
if [ $3 != "cifar10" ] && [ $3 != "imagenet" ]
then
echo "error: the selected dataset is neither cifar10 nor imagenet"
exit 1
fi
dataset_type=$3
fi
if [ $3 == 'imagenet' ]; then
CONFIG_FILE="$PROJECT_DIR/../config/imagenet_config_gpu.yaml"
elif [ $3 == 'cifar10' ]; then
CONFIG_FILE="$PROJECT_DIR/../config/cifar10_config_gpu.yaml"
else
echo "error: the selected dataset is neither cifar10 nor imagenet"
exit 1
fi
export DEVICE_ID=$1
export RANK_ID=0
export DEVICE_NUM=1
export RANK_SIZE=1
rm -rf ./train_single_gpu
mkdir ./train_single_gpu
cp -r ../src ./train_single_gpu
cp ../train.py ./train_single_gpu
cp -r ../config ./train_single_gpu
echo "start training for rank $RANK_ID, device $DEVICE_ID, $dataset_type"
cd ./train_single_gpu || exit
python ./train.py --config_path=$CONFIG_FILE \
--dataset_name=$dataset_type --train_data_dir=$train_data_dir --device_target=GPU> ./train.log 2>&1 &

2
model_zoo/official/cv/tinydarknet/scripts/run_train_cpu.sh
View File

@ -49,7 +49,7 @@ rm -rf ./train_cpu
mkdir ./train_cpu
cp ./train.py ./train_cpu
cp -r ./src ./train_cpu
cp ./*.yaml ./train_cpu
cp -r ./config ./train_cpu
echo "start training for device CPU"
cd ./train_cpu || exit
env > env.log

11
model_zoo/official/cv/tinydarknet/src/dataset.py
View File

@ -40,14 +40,10 @@ def create_dataset_cifar(dataset_path,
Returns:
dataset
"""
if target == "Ascend":
device_num, rank_id = _get_rank_info()
elif target == "CPU":
if target == "CPU":
device_num = 1
else:
init()
rank_id = get_rank()
device_num = get_group_size()
device_num, rank_id = _get_rank_info()
if device_num == 1:
data_set = ds.Cifar10Dataset(dataset_path,
@ -165,7 +161,8 @@ def _get_rank_info():
rank_size = int(os.environ.get("RANK_SIZE", 1))
if rank_size > 1:
from mindspore.communication.management import get_rank, get_group_size
from mindspore.communication.management import init, get_rank, get_group_size
init()
rank_size = get_group_size()
rank_id = get_rank()
else:

2
model_zoo/official/cv/tinydarknet/src/model_utils/config.py
View File

@ -117,7 +117,7 @@ def get_config():
"""
parser = argparse.ArgumentParser(description="default name", add_help=False)
current_dir = os.path.dirname(os.path.abspath(__file__))
parser.add_argument("--config_path", type=str, default=os.path.join(current_dir, "../../{}".format(_config)),
parser.add_argument("--config_path", type=str, default=os.path.join(current_dir, "../../config/{}".format(_config)),
help="Config file path")
path_args, _ = parser.parse_known_args()
default, helper, choices = parse_yaml(path_args.config_path)

14
model_zoo/official/cv/tinydarknet/train.py
View File

@ -21,7 +21,7 @@ import time
from mindspore import Tensor
from mindspore import context
from mindspore.communication.management import init
from mindspore.communication.management import init, get_rank
from mindspore.nn.optim.momentum import Momentum
from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor
from mindspore.train.loss_scale_manager import DynamicLossScaleManager, FixedLossScaleManager
@ -36,7 +36,7 @@ from src.tinydarknet import TinyDarkNet
from src.CrossEntropySmooth import CrossEntropySmooth
from src.model_utils.config import config
from src.model_utils.moxing_adapter import moxing_wrapper
from src.model_utils.device_adapter import get_device_id, get_device_num, get_rank_id
from src.model_utils.device_adapter import get_device_id, get_device_num
set_seed(1)
@ -132,11 +132,11 @@ def run_train():
else:
context.set_context(device_id=get_device_id())
if device_num > 1:
init()
context.reset_auto_parallel_context()
context.set_auto_parallel_context(device_num=device_num, parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True)
init()
rank = get_rank_id()
rank = get_rank()
if config.dataset_name == "imagenet":
dataset = create_dataset_imagenet(config.train_data_dir, 1)
@ -204,10 +204,12 @@ def run_train():
if device_target == "CPU":
model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'}, loss_scale_manager=loss_scale_manager)
else:
elif device_target == "Ascend":
model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'},
amp_level="O3", loss_scale_manager=loss_scale_manager)
elif device_target == "GPU":
model = Model(net, loss_fn=loss, optimizer=opt, metrics={'acc'},
amp_level="O2", loss_scale_manager=loss_scale_manager)
config_ck = CheckpointConfig(save_checkpoint_steps=batch_num * 50, keep_checkpoint_max=config.keep_checkpoint_max)
time_cb = TimeMonitor(data_size=batch_num)
ckpt_save_dir = os.path.join(config.ckpt_save_dir, str(rank))