From 31b7c9f62f7a48e1543e6945f278d0aaba171350 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E9=83=91=E7=89=B9=E9=A9=B9?= Date: Tue, 20 May 2025 17:46:18 +0800 Subject: [PATCH] =?UTF-8?q?[contrib][PyTorch][MAE]=20=E5=88=A0=E9=99=A4REA?= =?UTF-8?q?DME=E4=B8=AD=E9=83=A8=E5=88=86=E4=B8=8E=E4=BA=A7=E5=93=81?= =?UTF-8?q?=E7=9B=B8=E5=85=B3=E7=9A=84=E5=86=85=E5=AE=B9?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../cv/classification/MAE_for_PyTorch/README.md | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-) diff --git a/PyTorch/contrib/cv/classification/MAE_for_PyTorch/README.md b/PyTorch/contrib/cv/classification/MAE_for_PyTorch/README.md index 3919fc9430..d0c1281073 100644 --- a/PyTorch/contrib/cv/classification/MAE_for_PyTorch/README.md +++ b/PyTorch/contrib/cv/classification/MAE_for_PyTorch/README.md @@ -172,13 +172,13 @@ MAE的设计虽然简单，但已被证明是一个强大的、可扩展的视 # 输出完整微调日志./output_finetune_full_8p/Atlas_8p_finetune_full.log、总结性日志./output_finetune_full_8p/log.txt bash ./test/finetune_full_8p.sh --data_path=real_data_path --finetune_pth=pretrained_model_path - # fine-tuning_large 8p performance，8p上运行1个epoch，910运行时间约为14min + # fine-tuning_large 8p performance，8p上运行1个epoch bash ./test/finetune_performance_large_8p.sh --data_path=real_data_path --finetune_pth=pretrained_model_path - # fine-tuning_large 8p full，8p上运行50个epoch，910运行时间约为12h + # fine-tuning_large 8p full，8p上运行50个epoch bash ./test/finetune_full_large_8p.sh --data_path=real_data_path --finetune_pth=pretrained_model_path - # fine-tuning_large 16p full，16p上运行50个epoch，910运行时间约为6h + # fine-tuning_large 16p full，16p上运行50个epoch bash ./test/finetune_full_large_16p.sh --data_path=real_data_path --finetune_pth=pretrained_model_path # 8p Base_eval，运行时间约为3min @@ -186,7 +186,7 @@ MAE的设计虽然简单，但已被证明是一个强大的、可扩展的视 bash ./test/finetune_eval_8p.sh --data_path=real_data_path --resume_pth=finetuned_model_path ``` - 说明：MAE-Large配置在910上进行训练，16p脚本需要在启动脚本修改节点IP + 说明：16p脚本需要在启动脚本修改节点IP 模型训练脚本参数说明如下。 @@ -229,9 +229,7 @@ MAE的设计虽然简单，但已被证明是一个强大的、可扩展的视 | 8p-NPU | 83.86 | 1603 | 50 | O2 | 1.8 | | 16p-NPU | 85.97 | 3145 | 50 | O2 | 1.8 | -说明1：MAE-Large配置在910上进行训练 - -说明2：上表为历史数据，仅供参考。2025年5月10日更新的性能数据如下： +说明：上表为历史数据，仅供参考。2025年5月10日更新的性能数据如下： | NAME | 精度类型 | FPS | | :------ |:-------:|:------:| | 8p-竞品 | FP16 | 3538.01 | -- Gitee