questions about implement and deploy details

[math-yyj]

Nice work ! and I have two questions:

1. How many GPUs needed for full-training ?
2. Is there any guide for ONNX export ?

many thanks !

[wanghao9610]

Thanks for your attention to our work. (1) OV-DINO is pre-trained on 32 A100 GPUs with batch size 128, then finetuned on 8 A100 GPUs with batch size 32. (2) Good suggestion, I will add it to the road map. I'm the only one coding, please stay tuned.

[math-yyj]

Thanks for your attention to our work. (1) OV-DINO is pre-trained on 32 A100 GPUs with batch size 128, then finetuned on 8 A100 GPUs with batch size 32. (2) Good suggestion, I will add it to the road map. I'm the only one coding, please stay tuned.

Thank you for your kindly reply.

And, how many hours did pretrain and finetune cost ?

[wanghao9610]

OV-DINO¹ pre-trained on O365 takes about 1 day, OV-DINO² pre-trained on (O365, GoldG) takes about 3 days, and OV-DINO³ pre-trained on (O365, GoldG, CC1M‡) takes about 1 week, the fine-tuninging on COCO takes about 0.5 day. Note: The above training costs are not accurate exactly, just provided for clarification.