Issues when training

[tqtrunghnvn]

Hi,

Thank you so much for your contributions!

I am trying to train Openpifpaf model considering Temporal part. The command I used is as follows.

# train
time python3 -m torch.distributed.launch --nproc_per_node=4 \
  -m openpifpaf.train --ddp \
  --lr=0.0003 --momentum=0.95 --b-scale=10.0 \
  --epochs=50 --lr-decay 40 45 --lr-decay-epochs=5 \
  --batch-size=8 \
  --weight-decay=1e-5 \
  --dataset=cocokpst --dataset-weights 1 1 --stride-apply=2 \
  --posetrack-upsample=2 \
  --cocokp-upsample=2 --cocokp-orientation-invariant=0.1 --cocokp-blur=0.1 \
  --checkpoint outputs/tshufflenetv2k30-210217-075056-cocokp-o10s-6f9daa84.pkl

The only difference is I do NOT use posetrack dataset. I met an error:

assert x.shape[2] == 1 + self.n_confidences + self.n_vectors * 2 + self.n_scales
AssertionError

Delving into your code, I found two lines in openpifpaf/network/losses/composite.py:

assert x.shape[2] == 1 + self.n_confidences + self.n_vectors * 2 + self.n_scales
assert t.shape[2] == self.n_confidences + self.n_vectors * 3 + self.n_scales

I am so confused with the first line. For instance, Caf module has 9 fields: 1 confidence, 2 vectors with 3 components each, and 2 scaling factors. As a result, why do you have the first line?

Could you please give me an explanation?

I am looking forward to seeing your answer soon!
Thank you!