search

WentaoTan / MLLM4Text-ReID

Code for Harnessing the Power of MLLMs for Transferable Text-to-Image Person ReID (CVPR 2024)
39 stars 0 forks source link

Regarding the issue of three checkpoint models #9

Open HuidiXie opened 3 months ago

HuidiXie commented 3 months ago

Thank you for your open source.I have a question for you. I selected best2.pth from the Pretrain Model Checkpoints you provided, and it performed well. However, I still want to know the differences between the three Pretrain Model Checkpoints, and which model would yield the best results for replication. (I would like to ask this sincerely.)

WentaoTan commented 3 months ago

The difference in direct transfer performance of these three pretrained models on a single dataset is about 1% - that is, the three .pth files should have similar effects. 

In our experiments, we use best0.pth for the CUHK-PEDES fine-tuning, best1.pth for the ICFG, and best2.pth for the RSTP.

Hope this helps you!

------------------ 原始邮件 ------------------ 发件人: "WentaoTan/MLLM4Text-ReID" @.>; 发送时间: 2024年8月15日(星期四) 下午3:29 @.>; @.***>; 主题: [WentaoTan/MLLM4Text-ReID] Regarding the issue of three checkpoint models (Issue #9)

Thank you for your open source.I have a question for you. I selected best2.pth from the Pretrain Model Checkpoints you provided, and it performed well. However, I still want to know the differences between the three Pretrain Model Checkpoints, and which model would yield the best results for replication. (I would like to ask this sincerely.)

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

HuidiXie commented 3 months ago

Thank you for clarifying.

huzaifarashid3 commented 2 weeks ago

have a miissing config yaml file in this , any solutions?

Kuangdd01 commented 1 week ago

Same question. If I don't misunderstand, the 3 pretrained models were trained with different hyper-parameters? Could you share your different configurations during pretraining? @WentaoTan
It will be very helpful for us to reproduce this pretraining or just train models on the MLLM-REID sub-dataset(0.1M).

WentaoTan commented 1 week ago

As mentioned in README, the entire pre-training setting is to directly run "bash run.sh". When you run the code, you will find that each time the model tests the performance of three downstream data, three models will appear in the log folder, corresponding to the three models we released. So these three models are three different models run under the same setting, and they are saved in different epochs according to their performance in downstream tasks. For example, best0.pth may be the model of epoch=20, and best1.pth is epoch=25.