low map issue...

[Yang]

After getting 10th epoch model. The program evaluated this model and returned a series of results. But it seems that the results are very low...

[Abdullah Dursun]

Hi, thank you for your interest.

Unfortunately, the training is not fully deterministic, so it might return low mAP sometimes. Can you please try re-training?

[Yang]

yeah I try again. This time I tested the 20th epoch model but the result still very low....
By the way, when I tested the model, I still used trainval data cuz I felt that maybe the performance will be better rather than test dataset.

[Yang]

And this is my training process. I don't know whether it's correct...

[Abdullah Dursun]

Yes, your training output looks good.

Can you please download and put this file into states directory and run:

python src/evaluate.py --base_net vgg --state_path states/vgg_epoch_20.pt

Its output should be as follows:

Avg AP: [0.38831317 0.45263229 0.24562751 0.26572207 0.12294847 0.59961043
 0.47957328 0.24194355 0.09629139 0.29096157 0.15507882 0.26510262
 0.31636181 0.53462855 0.06115122 0.16652659 0.33271927 0.29181557
 0.47336772 0.51627529]
Avg mAP: 0.3148325589240723

[Yang]

thanks author! The results I run nearly same as the one that you give.
So... it looks strange and I think something wrong in training.
Maybe the number of epoch is too small?

[Abdullah Dursun]

You are welcome! The pretrained state file I've sent is a result of 20 epochs training, so I don't think epoch count is a problem.

On the other hand, the training might stuck at a local minimum. Can you please try with the following:

python src/train.py --base_net vgg --lr 1e-4

[Yang]

thanks author! it can work!

[Abdullah Dursun]

Thank you too!