Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

RT-DETR的N和tiny配置,以及学习率等配置 #8379

Open
1 task done
sdreamforchen opened this issue Jun 19, 2023 · 4 comments
Open
1 task done

RT-DETR的N和tiny配置,以及学习率等配置 #8379

sdreamforchen opened this issue Jun 19, 2023 · 4 comments
Assignees
Labels
feature request New feature or request rtdetr

Comments

@sdreamforchen
Copy link

是否愿意提交PR Are you willing to submit a PR?

  • Yes I'd like to help by submitting a PR!
@sdreamforchen sdreamforchen added the feature request New feature or request label Jun 19, 2023
@sdreamforchen
Copy link
Author

您好!很多应用场景其实N和Tiny级别的算法就够用了。大佬可否分享下,这两个级别的参数配置多少呢?
我自己的N级别的配置如下:
ResNet:
depth: 18
variant: d
return_idx: [1, 2, 3]
freeze_at: -1
freeze_norm: false
norm_decay: 0.

HybridEncoder:
hidden_dim: 192
use_encoder_idx: [2]
num_encoder_layers: 1
encoder_layer:
name: TransformerLayer
d_model: 192
nhead: 8
dim_feedforward: 512
dropout: 0.
activation: 'gelu'
expansion: 0.5
depth_mult: 1.0

RTDETRTransformer:
eval_idx: -1
num_decoder_layers: 3
另手动将backbone的resnet部分最后输出为128通道。经测试精度不算很好,应该和学习率、epoch(80)也有关系。
麻烦大佬分享一下配置,万分感谢!

@lyuwenyu
Copy link
Collaborator

lyuwenyu commented Jun 19, 2023

目前这个是最小的 再小的没有了 https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/rtdetr/rtdetr_r18vd_6x_coco.yml

再小的有几个位置可以调一调

  1. HybridEncoder.depth_mult 可以改成0.67或者 0.33试一下
  2. 修理完成之后,推理时候可以调一下 eval_idx 比如 -2

@sdreamforchen
Copy link
Author

感谢大佬回复! HGnet或者ResNet部分不需要调整吗? 我想获取传统yolo系列N和tiny级别的算法。
另外,因为目前训练COCO是用了预训练模型的,我理解,如果您能分享下训练object365的学习率等配置,是不是从0开始训练COCO也会更好

@lyuwenyu
Copy link
Collaborator

backbone有预训练一般不调整 更小的话去尝试换更小的backbone 比如mobilenet等等

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
feature request New feature or request rtdetr
Projects
None yet
Development

No branches or pull requests

2 participants