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

PP—ShiTuV2 RPC训练数据问题确认 #3174

Open
xiahc opened this issue Jun 28, 2024 · 10 comments
Open

PP—ShiTuV2 RPC训练数据问题确认 #3174

xiahc opened this issue Jun 28, 2024 · 10 comments
Assignees

Comments

@xiahc
Copy link

xiahc commented Jun 28, 2024

数据集
1,数据集中的数据量,是图片数,还是分类数?
2,PRC的数据,官方给出的是200分类,训练集数据近6万。与PaddleClas给出数据不符。
请帮忙确认图片中,RPC 3K,是什么意思。如果我们只用RPC数据做迁移学习,训练数据集应如何调整?
3,我们目前用RPC全量数据做迁移学习,loss停留在0.4无法收敛,请给出优化建议。
配置如下:
BASE: [
'../../../runtime.yml',
'../../base/picodet_esnet.yml',
'../../base/optimizer_100e.yml',
'../../base/picodet_640_reader.yml',
]

pretrain_weights: https://paddledet.bj.bcebos.com/models/picodet_lcnet_x2_5_640_mainbody.pdparams
weights: output/picodet_lcnet_x2_5_640_mainbody/model_final
find_unused_parameters: True
use_ema: true
cycle_epoch: 10
snapshot_epoch: 1

PicoDet:
backbone: LCNet
neck: CSPPAN
head: PicoHead

LCNet:
scale: 2.5
feature_maps: [3, 4, 5]

metric: COCO
num_classes: 1

LearningRate:
base_lr: 0.0005
schedulers:

  • !CosineDecay
    max_epochs: 100
  • !LinearWarmup
    start_factor: 1
    steps: 1

TrainDataset:
!COCODataSet
#image_dir: ./
image_dir: train2019/
anno_path: instances_train2019_main.json
dataset_dir: dataset/dataset_rpc/
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']

EvalDataset:
!COCODataSet
image_dir: val2019/
anno_path: instances_val2019_main.json
dataset_dir: dataset/dataset_rpc/
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']

TestDataset:
!ImageFolder
anno_path: ./dataset/dataset_rpc/instances_test2019_main.json

@xiahc
Copy link
Author

xiahc commented Jun 28, 2024

batch size=32

@cuicheng01
Copy link
Collaborator

你好,感谢提问

  1. 是图片数量
  2. 使用的是RPC的子集
  3. mAP的指标如何呢?是否还在增长呢?

@xiahc
Copy link
Author

xiahc commented Jul 8, 2024

您好,感谢回复。
mAP约0.4。
另外,我使用官方的主体检测模型(picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer)进行测试。存在常规目标(非小目标)丢失的问题,请问,该如何调整,以改进漏检的问题(如图片中中间最明显的牛奶瓶)。
0

@cuicheng01
Copy link
Collaborator

该主体检测模型其实是一个框的召回模型,具体的识别效果取决于识别模型和你的数据库,所以可以从调整识别模型和数据库上入手调试

@zhanghaowei01
Copy link

我训练rpc数据集精度很差怎么办,我感觉是因为验证集和训练集有差距的问题,你是怎么训练到map0.4的?

@cuicheng01
Copy link
Collaborator

你的精度有多少呢

@zhanghaowei01
Copy link

我就是用他单个目标的训练集,验证集是多个目标的,map95顶多到0.24

@zhanghaowei01
Copy link

PP—ShiTuV2训练主体检测的时候rpc数据集是怎么用的?多个目标的数据有没有加入到训练集里?

@cuicheng01
Copy link
Collaborator

cuicheng01 commented Jul 11, 2024

多个目标也是按单类别来训练的

@zhanghaowei01
Copy link

P—ShiTuV2训练主体检测的时候训练集里加了多个目标的数据吗?还是只有单个目标

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

No branches or pull requests

4 participants