Error on Stage 1 training with FSCD147 dataset

[briliantnugraha]

Hi VinAI research, great paper there for counting unknown objects!

Issue: I have tried the following script to test the stage 1 with FSCD147 dataset.

1. This script is successfully run
2. this script raise "key error whs"

The error happens on stage 1, point 2 (as in the link above), where I get

KeyError: 'whs'

when trying to run this script on stage 1.

CMIIW, Could I know, is it a bug, unfinished label sharing, or something else?

Here I also attach the snapshot of the error, hope it helps.

Expected output: I could run the whole script in the weakly_supervise_fscd_147.sh without any error.

Thanks in advance~

[briliantnugraha]

I've just reviewed the code a little bit more, it seems that the "whs" key is available when using --dataset_file "fscd_147", but not on the "fscd_147_point".

However, this key is needed in your training pipeline as shown in here, and here where I encounter the keyerror "whs".

new update

So, I've added "whs" key in here, and then I get another error in the loss calculation. Where the number of wh in tgt_whs is only (3,2), but the shape of tgt_point and tgt_whs are (57,2) as below.

So, I guess it is due to the number of gt and density which do not match (but this error shouldn't happen in your experiments). Hence, could i know what is the purpose of the stage 1, point 2/this script?

Thanks~

[NguyenVanThanhHust]

Sorry for late reply. Can you copy error into here? Instead of screenshot?

[NguyenVanThanhHust]

@briliantnugraha I will check ASAP. Currently, I'm quite busy so I might do it in the weekend. It could be the scripts training, not the source code.

[briliantnugraha]

Hi @NguyenVanThanhHust , will there be a fix for this issue? Or should I move to the LVIS script directly (because all annotations is available in the dataset)? thanks