Questions on the dataset for DHN and training process about deepmot HOT 3 CLOSED

Hi, thank for testing our work!
(1) the printed number "1998" and "11870" are actually # of batches. Because the matrices are of varying, I randomly packed matrices of a same size and saved it to a same .npy (of <=10 in a batch). If you count the # of elements in the batches. You should get 1998 and 11870.

(2) I tested with my singularity image (like docker), it printed the wa correctly:

It was a problem of pytorch version of line 506 and 507 in deepmot/train_DHN/utils.py:

 num_tp = float(((predicted[b, :, :] == target[b, :, :]) + (target[b, :, :] == 1.0)).eq(2).sum())
 num_tn = float(((predicted[b, :, :] == target[b, :, :]) + (target[b, :, :] == 0.0)).eq(2).sum())

in higher version of pytorch, it seems that it will not convert bool to float automatically, so ".eq(2)" will always be False and num_tp and num_tn will be always zero.

The bug can be fixed by adding ".float()":

 num_tp = float(((predicted[b, :, :] == target[b, :, :]).float() + (target[b, :, :] == 1.0).float()).eq(2).sum())
 num_tn = float(((predicted[b, :, :] == target[b, :, :]).float() + (target[b, :, :] == 0.0).float()).eq(2).sum())

Thank you for pointing out this bug.

from deepmot.

Thanks for your reply @yihongXU !

1. I see, so there's at most 10 in each of the batches, that makes sense. Did you hand-pick the matrices so the dataset would not use too much GPU memory?

2. Thank you for debugging! It was my bad, I didn't use singularity but followed tracking_wo_bw's instructions and forgot to use an older version of pytorch.

from deepmot.

Hi，
I randomly pick ~10 and made them a batch. Why 10, for GPU memory issue, but you can reorganize them according to your gpus. And you can also use your own data since the given one are all from MOT.

from deepmot.