search

huanghoujing / beyond-part-models

PCB of paper: Beyond Part Models: Person Retrieval with Refined Part Pooling, using Pytorch
331 stars 81 forks source link

Do you evaluate the model with the checkpoint by epoch-60 ? #14

Closed YunpengZhai closed 6 years ago

YunpengZhai commented 6 years ago

Emmmm, I found the rank1 and mAP drop a little, maybe because overfitting.

huanghoujing commented 6 years ago

Yes, I test the model on test set ONLY when the training process ends. Maybe intermediate results are better, I am not sure :-)

Gavin666Github commented 6 years ago

pcb

the results of training and testing are as above ,but set test only ,it turned out pretty bad.see figure below:

pcb_test

is there anyone who knows what happened?

huanghoujing commented 6 years ago

In your log, there is warning about loading model weight. You have to make this correct first.

huanghoujing commented 6 years ago

@Gavin666Github It seems that your problem is related to the difference between model_weight_file and ckpt_file in my code. I think this issue should solve your problem.

Gavin666Github commented 6 years ago

@huanghoujing Thank you for the guidance. I'll try to do it with new commands.

Gavin666Github commented 6 years ago

pcb_test1

@huanghoujing As shown in the picture, the problem has been solved perfectly, and it is very powerful. Thank you very much indeed!