Training and evaluation on MS-COCO dataset

[KunpengLi1994]

Thanks for sharing the code! It seems that config files, pre-train models and evaluation script are only provided for Flickr30K dataset. Could you please share these files for MS-COCO dataset?

If they are not available, could you please just share the test yaml file, evaluation script for both 5-folder average and the whole 5 folder on MS-COCO for now? Thanks a lot.

[ZihaoWang-CV]

Thank you for your interest in our paper! Currently, I have left CUHK and SenseTime, may not provide those evaluation scripts and pre-train models quickly. But the hyper-parameters for both two datasets are similar.

The codes for 5-folds average evaluation is not uploaded here, but we could add several lines to test_modules.py to implement it. Refer to line 202-232 in evaluation.py, we could add a similar loop to line 202 intest_modules.py to apply the 5-folds evaluation.

[KunpengLi1994]

Thanks for the response. Such modifications make sense for evaluation on MS-COCO.

I also find training and evaluation of CAMP require large GPU memories and multi-GPUs training. Did you remember how long it took for training on MS-COCO and how many GPUs were used.