search

WangWenhao0716 / Adapted-Center-and-Scale-Prediction

Pytorch implementation of "Adapted Center and Scale Prediction: More stable and More Accurate"
https://arxiv.org/abs/2002.09053
64 stars 12 forks source link

The losses were low at the beginning of the training #13

Open s775980299 opened 3 years ago

s775980299 commented 3 years ago

I retrained your code. Why does the loss fluctuate from around 0.

WangWenhao0716 commented 3 years ago

What about the accuracy? I think the loaded pre-trained model is False.

s775980299 commented 3 years ago

I have set the pre-training model to False, but the training loss value is still even 0.Because I only have one GPU and the batchsize can only be set to 1 for training, so my accuracy is not as good as yours, your loss value is also from 0 when training？

WangWenhao0716 commented 3 years ago

I see. Please give up training this project. Sorry, the minimum requirement has not been achieved.

发自我的iPhone

------------------ Original ------------------ From: s775980299 <notifications@github.com> Date: Sun,Jan 24,2021 3:21 PM To: WangWenhao0716/Adapted-Center-and-Scale-Prediction <Adapted-Center-and-Scale-Prediction@noreply.github.com> Cc: Wenhao Wang <2609667791@qq.com>, Comment <comment@noreply.github.com> Subject: Re: [WangWenhao0716/Adapted-Center-and-Scale-Prediction] The losses were low at the beginning of the training (#13)

I have set the pre-training model to False, but the training loss value is still even 0.Because I only have one GPU and the batchsize can only be set to 1 for training, so my accuracy is not as good as yours, your loss value is also from 0 when training？

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

s775980299 commented 3 years ago

Excuse me, is it because I only have one GPU that I cannot meet the minimum requirements？ Should I give up your code? I've been looking at it for a long time.

WangWenhao0716 commented 3 years ago

Please refer the discussion of gpu number and batch size in the original paper. I give the suggestion is because this setting may (very possible) not bring you a convergent result.

发自我的iPhone

------------------ Original ------------------ From: s775980299 <notifications@github.com> Date: Sun,Jan 24,2021 3:49 PM To: WangWenhao0716/Adapted-Center-and-Scale-Prediction <Adapted-Center-and-Scale-Prediction@noreply.github.com> Cc: Wenhao Wang <2609667791@qq.com>, Comment <comment@noreply.github.com> Subject: Re: [WangWenhao0716/Adapted-Center-and-Scale-Prediction] The losses were low at the beginning of the training (#13)

Excuse me, is it because I only have one GPU that I cannot meet the minimum requirements？ Should I give up your code? I've been looking at it for a long time.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.