How many steps to train roma_outdoor?

[Zhimin00]

I notice the length of megadepth_train is 9,903,321, and the number of total training steps is (32*250,000) = 8,000,000

megadepth_train1 = mega.build_scenes( split="train_loftr", min_overlap=0.01, shake_t=32, use_horizontal_flip_aug = use_horizontal_flip_aug, rot_prob = rot_prob, ht=h,wt=w, ) megadepth_train2 = mega.build_scenes( split="train_loftr", min_overlap=0.35, shake_t=32, use_horizontal_flip_aug = use_horizontal_flip_aug, rot_prob = rot_prob, ht=h,wt=w, ) megadepth_train = ConcatDataset(megadepth_train1 + megadepth_train2)

[Parskatt]

Yes, that's correct.

[Zhimin00]

Why do you use less steps than the whole training dataset?

[Parskatt]

Yes, it doesnt matter much, as dataset is pairs, not images.

[Zhimin00]

Thank you for your reply.