About Pretrain

[ysysys666]

Thank you for your excellent work!

I have a few questions I would like to ask and look forward to your response.

1. Was the model initialized from scratch, or were the weights from GroundingDINO loaded? It seems like the weights were loaded based on the configuration.

2. The paper mentions that the model was trained using 4 A100 GPUs, with a batch size of 2 for 180k steps. I calculated that this corresponds to approximately 12 epochs. However, the configuration suggests 32 epochs. Could you please clarify the actual training setup?

[jaychempan]

1. LAE-DINO uses pretrained GroundingDINO for faster convergence—you may also try training from scratch.

2. Convergence is judged by the validation set, with a large training epoch setting. Using 32 epochs and a batch size of 4 as reference, optimal performance is usually reached around 20+ epochs.