PaddlePaddle/PaddleDetection

PAFNet训练loss难以下降

Open

#4,521 opened on Nov 9, 2021

View on GitHub
 (5 comments) (0 reactions) (1 assignee)Python (11,414 stars) (2,731 forks)batch import
help wantedtraining

Description

在自己的实验过程中更换了PAFNet的Backbone进行相关实验，由于此前没有使用dcn_head导致精度可能还是存在差距，所以在近期使用CSPDarknet+dcn_head的配置进行实验，但是发现在实验过程中模型的loss难以下降，使用COCO数据集跑了12个epoch后可以说是什么也没检测到，这是因为DCN的问题所导致的吗？ image

Contributor guide

Tech stack
python
Domain
machine learning
Issue type
bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
PaddlePaddle frameworkObject detection conceptsDCN (Deformable Conv)
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
The issue suspects DCN (Deformable Convolution) in dcn head may cause training instability. To investigate, compare loss curves with a standard head (e.g., without DCN) using the same CSPDarknet backbone. Check the PAFNet configuration files and the DCN implementation in ppdet/modeling/heads/. Also verify that the DCN module is correctly initialized and used during training. If the issue persists, open a minimal reproduction script.