Henry

@qtw1998 No, you need create folder 'weights' or add `os.makedirs(directory)' to create folder contains 'weights'. You do not need to pretrained the model first. Sorry, I have not added this...

@qtw1998 I see it only takes ~3000 MB per GPU, you try increasing the batch_size?

@qtw1998 Can you train COCO2017 and let me know mAP. I haven't had time to train this dataset now

I updated sell script for download coco2017 at (https://github.com/toandaominh1997/EfficientDet.Pytorch/blob/master/datasets/scripts/COCO2017.sh). Please checkout last commit and follow readme for training coco2017. Thanks @qtw1998

Can you increase the batch size(+4 or +8 ...) to test whether the model for the percentage of gpu? On Mon, Dec 30, 2019 at 5:04 PM Qiao Tianwei wrote:...

@qtw1998 What OS are you using? I have run 2 gpu 2080Ti on ubuntu well with batch size 32 on efficientdet-d0

hi @zhanglaplace, I just trained on VOC at 37 epoch, it really hasn't optimized yet, If you get better for detection, you can try this weights: [weights at epoch 65](https://drive.google.com/open?id=1hWqQf99K49FCeMJ-WSFVR_LFr7VfYCuE)...

@buiduchanh I think there are 2 causes that effect performance: + 1. According to paper, Efficientdet-D0 use backbone B0, Wbifpn = 64, Dbifpn = 2. In my repo, t changed...

@AlexLuya @RayOnFire @shengyuqing I used: OS: Ubuntu 18.04 GPU: 2*2080TI(11GB) When training, I set batch_size 32 for EffficientDet-D0(~20000MB cuda), and batch_size 16 for EfficientDet-Do(~20000MB cuda). At commit #36 , If...