Maybe there is an error in the loss_indices calculation?

[Koruvika]

line 152 in trainer_spectre.py

In the paper, batch_size equals to 1. With that, loss_indices equals [], is there any mistakes here?

[filby89]

Hey,
if you check https://github.com/filby89/spectre/blob/master/src/trainer_spectre.py#L111 the original batch is 5 D:

1 x T x 3 x H x W

and then we resize it to:

T x 3 x H x W

so these loss indices actually show that the loss applies to all frames in the sequence but the first two and the last two !