Yongtao Ge

Hi, @pangyyyyy Could you give me an example visualization script for the converted agora dataset? Thanks!

@ZhitingHu could you add an example for using horovod in texar-torch?

Got it!

Just another question. fddb_test.py's output is [x1, y1, w, h, score], which is different from the file format in "detections/fddb/DPMDiscROC.txt" like: ``` 0.890737 35595 -0.999989 0.890737 35594 -0.999972 0.890737 35593...

It really helps. Thank you!


@KaleidoZhouYN 1.Could you explain why FN requires large batch size to train the network? 2.To what extend the batch size will be different using FN or not?

Hi! I would release a pretrained MobileNet0.25 model recently. Maybe more models like MNASNet in the future cause currently I am focusing on other tasks and the code still have...

1. yes, you are right! 2. I just couldn't understand your second question.

Hi, @xizero00 pls use the --cfg-options model.filp_fuse_type=\'type2\' when evaluation, e.g. ``` ./tools/dist_test.sh configs/body/2d_kpt_sview_rgb_img/poseur/coco/poseur_r50_coco_256x192.py \ poseur_256x192_r50_6dec_coco.pth 8 \ --eval mAP \ --cfg-options model.filp_fuse_type=\'type2\' ``` If the results are still lower than...