Skip to content

【生态活动】元碁智汇·定义未来-团队名称-模型训练-在PyTorch框架上支持Deformable-DETR模型在VOC2007数据集上的训练【PR模版】#5

Open
Arrmsgt wants to merge 1 commit into
Tecorigin:mainfrom
Arrmsgt:contrib/jiangnan_university_ailab/Deformable-DETR
Open

【生态活动】元碁智汇·定义未来-团队名称-模型训练-在PyTorch框架上支持Deformable-DETR模型在VOC2007数据集上的训练【PR模版】#5
Arrmsgt wants to merge 1 commit into
Tecorigin:mainfrom
Arrmsgt:contrib/jiangnan_university_ailab/Deformable-DETR

Conversation

@Arrmsgt

@Arrmsgt Arrmsgt commented Jul 15, 2026

Copy link
Copy Markdown

【请勿合入,仅作为PR模板进行展示】
当前软件栈版本:
--------------+----------------------------------------------
Host IP | 10.71.12.24
PyTorch | 2.7.1
Torch-SDAA | 3.2.0b0+git20c7eca
--------------+----------------------------------------------
SDAA Driver | 3.2.0 (N/A)
SDAA Runtime | 3.2.0b0 (/opt/tecoai/lib64/libsdaart.so)
SDPTI | 1.7.0 (/opt/tecoai/lib64/libsdpti.so)
TecoDNN | 3.2.0b0 (/opt/tecoai/lib64/libtecodnn.so)
TecoBLAS | 3.2.0b0 (/opt/tecoai/lib64/libtecoblas.so)
CustomDNN | 3.2.0b0 (/opt/tecoai/lib64/libtecocustom.so)
TecoRAND | 3.1.0 (/opt/tecoai/lib64/libtecorand.so)
TCCL | 3.2.0b0 (/opt/tecoai/lib64/libtccl.so)
--------------+----------------------------------------------

源码参考:
fork from:https://github.com/Tecorigin/Deformable-DETR
commit id:
工作目录:
PyTorch/contrib/Detection/Deformable-DETR
适配内容:
龙芯环境:使用1张T100芯片,在PyTorch框架上支持Deformable-DETR模型在VOC2007数据集上的训练。

模型运行脚本如下:
python main.py
--dataset_file coco
--coco_path ./data/voc_coco
--output_dir exps/voc_finetune
--num_classes 21
--resume ./r50_deformable_detr_single_scale-checkpoint.pth
--num_feature_levels 1
--lr 2e-5
--lr_backbone 2e-6
--lr_drop 30
--epochs 50
--batch_size 2
--clip_max_norm 0.1
--weight_decay 1e-4

执行测试:
cd PyTorch/contrib/Detection/Deformable-DETR/
pip install -r requirements.txt
cd run_scripts
bash test.sh

训练结果:
表格的形式展示结果(同README.md中的训练结果),例如:

加速卡数量 模型 混合精度 Batch size Shape 性能(imgs/s) AP@0.50:0.95 sdaa训练耗时
1 Deformable-DETR 2 动态 254 imgs/s 0.586 4h

备注:

  1. 训练结果的 Batch size 和 性能(imgs/s)均为,单卡的结果(bs和性能在tcap logger打印中为单核组, 太初卡单卡为四核组)
  2. sdaa训练耗时为完整端到端精度耗时
  3. 该PR仅展示提交PR的信息说明,请勿参考修改的文件

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant