[重庆理工大学][高校贡献][Pytorch][Vgg19]-第一次提交 · Pull Request !8 · Ascend/modelzoo

#Vgg19
任务：torch1.5->1.8
主要修改：
1.在maim.py 加入了
if torch.__version__>= '1.8.1':
      import torch_npu
2../scripts/train_8p.sh 将默认训练epoch修改为240
3.重新训练后，更新了readme.md中精度和推理速度指标。

## VGG-19 training result

| Acc@1  | FPS  | Npu_nums | Epochs | AMP_Type |
|:------:|:----:|:--------:|:------:|:--------:|
|   -    | 719  |    1     |  240   |    O2    |
| 75.553 | 3238 |    8     |  240   |    O2    |

# 自验报告
## Training

To train a model, run `main.py` with the desired model architecture and the path to the ImageNet dataset:

# 1p training 1p
bash ./scripts/train_1p.sh          # training accuracy

cd test
bash ./train_performance_1p.sh   # training performance

# 8p training 8p
bash ./scripts/train_8p.sh    #240epoch     # training accuracy

cd test
bash ./train_performance_8p.sh    # training performance

# eval default 8p， should support 1p
bash ./test/train_eval_8p.sh  --data_path=xxx

## Traing log 
cat output_1p.log ; output_8p.log    # training detail log

see terminal print    # training performance result

test/output/devie_id/Vgg19_${device_id}_bs_8p_acc.log

Ascend/modelzoo暂停 .gitee-modal { width: 500px !important; }

搜索帮助

Ascend/modelzoo
暂停