search

daodaofr / hypergraph_reid

Code for CVPR 2020 paper Learning Multi-Granular Hypergraphs for Video-Based Person Re-Identification
72 stars 23 forks source link

loss converges to 4.4, map only reaches 60.0%, rank1 only reaches 73.3%, learning rate is set to 0.000075, and the step size is 100. #6

Closed Seven-gcc closed 3 years ago

Seven-gcc commented 4 years ago

I use GeForce RTX 2080 with 7G memory, loss converges to 4.4, map only reaches 60.0%, rank1 only reaches 73.3%, batchsize is set to 8,learning rate is set to 0.000075, and the step size is 100. Because there was not enough memory during the test, only a part of it was loaded. Do you think it is the parameter setting or too little test load data? I adjusted it several times, but the effect was not satisfactory. Thank you ！

daodaofr commented 4 years ago

I think the combination of batch size and learning rate may impact the performance. Sorry, I didn't tried batchsize=8, so I cannot give much suggestion.

About the memory issue, you may extract the features part by part, and then evaluate it.

Seven-gcc commented 4 years ago

Thank you for your reply! I will try again,hoping it will be ok!

Seven-gcc commented 3 years ago

Hello！I want to confirm a few questions. To what extent can your loss function finally converge? What is the accuracy of the code that you published?map 84.1,rank1 88.7? I want to confirm that whether the code excludes Lmi and Att. The loss I reproduced can only converge to about 4. Thank you very much.

daodaofr commented 3 years ago

Hi, yes, this code does not contain Lmi and Att, these two terms have a limited impact on the overall performance. Normally, this code can achieve map ~85, rank1 ~89. I will check and upload a trained model soon, for you to reproduce the performance.

Seven-gcc commented 3 years ago

It would be great if you can upload the trained model. I tried for a long time but still didn't reproduce it.Thank you again!!

daodaofr commented 3 years ago

Don't worry, I will use this code to train the model, normally it will finish training in 10 days. I will get back to you as soon as possible.

Seven-gcc commented 3 years ago

Thank you very much!! That will help me a lot!!

daodaofr commented 3 years ago

I've trained 200 epochs, the result seems normal. I get mAP 80% and rank-1 85% for now. You may find the trained model helpful: https://drive.google.com/file/d/164jD6fxQISl1HFkkjKmCCwPOTDaQT_gu/view?usp=sharing

Seven-gcc commented 3 years ago

Thank you for your excellent work!! Maybe the batch size has an important effect on the performance. All in all, thank you very much for your help！

khawar-islam commented 3 years ago

@Seven-gcc Do you get this error? "Will redo. Don't worry. Just chill". I just put the dataset inside mars folder and align all the folders but this error happened.

Seven-gcc commented 3 years ago

Sorry, I did not meet the problem. It is my regret that I cannot help you.

在 2021-03-18 20:24:54，"Khawar Islam" @.***> 写道：

@Seven-gcc Do you get this error? "Will redo. Don't worry. Just chill". I just put the dataset inside mars folder and align all the folders but this error happened.

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

khawar-islam commented 3 years ago

Did you get this error? RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED

Seven-gcc commented 3 years ago

I think that belongs to your environmental problems. Sorry I have not encountered.

Seven-gcc commented 3 years ago

Sorry,I gave up the frame work and I did'nt reproduce the performance.

At 2021-03-20 20:13:09, "Khawar Islam" @.***> wrote:

@Seven-gcc I solved this error. Still gets CUDA out of memory when my batch size in 1. I have 16GB RAM and training on RTX 2070. Help required please

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