What is the training flow?

Question

What is the training flow?

Closed this issue 2 years ago · 3 comments

Hello, we are happy to follow this good job. We met some problems with the training flow. Is the model pre-trained on coco dataset first and then finetuned on the tao_ow_train dataset?

Answer 1 · 2022-08-24T15:09:16.000Z

Hi, the model (the detector to be precise) is ONLY trained on COCO, and there is no fine-tuning nor any kind of training on the TAO-OW dataset.
The model we were using is here: https://github.com/YangLiu14/detectron2-OWT

Answer 2 · 2022-08-25T07:29:12.000Z

Hi, the model (the detector to be precise) is ONLY trained on COCO, and there is no fine-tuning nor any kind of training on the TAO-OW dataset. The model we were using is here: https://github.com/YangLiu14/detectron2-OWT

Thanks for your quick reply! Since the model does not need to train on TAO-OW dataset, what is the train_split mentioned in paper used for ?

Answer 3 · 2022-08-31T02:18:28.000Z

Because the object categories in train_split is aligned with COCO-classes. So actually we only trained on COCO (namely using the pretrained models from Detectron2). You can also train/finetune on TAO-OW train-set. It's just that we did not do this fine-tuning step.