search

tuyunbin / SCORER

[ICCV 2023] This is the Pytorch code for our paper "Self-Supervised Cross-View Representation Reconstruction for Change Captioning".
14 stars 1 forks source link

Trying to reproduce on IER-like dataset #5

Open Gautier29 opened 5 months ago

Gautier29 commented 5 months ago

First thanks for this great work, it is very valuable to see your code and be able to run it,

I have reproduced your work on CLEVR-Change. Now I am trying to do it on a smaller dataset and as i see the results on Image Editing Request I think you could have some valuable insights if you are okay with sharing :)

  1. Will you release the code for training on Image Editing Request ? Most of my questions could be answered with that I guess !
  2. What was the training time with your GPU on Image Editing Request ?
  3. What was the hyperparameters (learning rate, batch_size, other thing you can think of and you struggled with) ?

For now, i am training on a dataset of the size of IER. SCORER seems to struggle and I want to be sure I am not doing anything wrong to be able to compare my thing with it fairly.

Again, thanks for this work and for your time

tuyunbin commented 5 months ago

Hi, thank you for your high recognition to our work. I will organize and release the code for training on Image Editing Request. Besides, for the training time and other hyperparameters, I have already reported them in supp. https://openaccess.thecvf.com/content/ICCV2023/supplemental/Tu_Self-supervised_Cross-view_Representation_ICCV_2023_supplemental.pdf.

Gautier29 commented 5 months ago

Thanks a lot for your answer, I'm looking forward to see your code and the supplementary material is great :)

Gautier29 commented 5 months ago

Hi, I guess the most important files I need are just the train.py and the dataset file suited for an IER training, if you have it somewhere I would be happy to take a look at it.

Gautier29 commented 5 months ago

Right now the main issue is that it always output the same caption after few epochs.

tuyunbin commented 5 months ago

You can give me the email, and I will sent both files to you. 

Yunbin Tu @.***

 

------------------ 原始邮件 ------------------ 发件人: "Gautier @.>; 发送时间: 2024年1月18日(星期四) 晚上11:31 收件人: @.>; 抄送: "Yunbin @.>; @.>; 主题: Re: [tuyunbin/SCORER] Trying to reproduce on IER-like dataset (Issue #5)

Hi, I guess the most important files I need are just the train.py and the dataset file suited for an IER training, if you have it somewhere I would be happy to take a look at it.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>