KeyError: 'feat_map' in dataset.py

Question

KeyError: 'feat_map' in dataset.py

Closed this issue 4 years ago · 9 comments

Thanks for releasing your wonderful work!

When I run the training commend, I got this error. Then I print all keys of info and I found that there is no feat_map item. I'm sure the program load info from db_trainval_with_pool.pkl file. Is there any thing wrong with my operation?

File "/home/x/Desktop/wsl/_HOI/IDN/dataset.py", line 290, in __getitem__
uni_vec.append(f['R'][info['feat_map'][cand_id], :])
KeyError: 'feat_map'

All keys in info:

dict_keys(['obj_gt_classes', 'obj_scores', 'is_gt', 'labels_ro', 'cand_id', 'pair_ids', 'obj_id', 'labels_r', 'height', 'width', 'boxes', 'is_gt_pair', 'pair_iou', 'obj_classes', 'filename', 'labels_sr', 'pool', 'H_mapping'])

Looking forward to your replay!

Answer 1 · 2020-11-21T07:50:04.000Z

BTW, when I run eval.py, it told me that:
RuntimeError: exp/IDN_IPT_hico/epoch_30.pth is a zip archive (did you mean to use torch.jit.load()?)
According to web, this error is due to the version of pytorch. However, my env is torch==1.3.1 same to requirements.txt. How can I deal with this situation? Thanks a lot!

Answer 2 · 2020-11-23T03:07:15.000Z

Thanks for releasing your wonderful work!

When I run the training commend, I got this error. Then I print all keys of info and I found that there is no feat_map item. I'm sure the program load info from db_trainval_with_pool.pkl file. Is there any thing wrong with my operation?

File "/home/x/Desktop/wsl/_HOI/IDN/dataset.py", line 290, in __getitem__
uni_vec.append(f['R'][info['feat_map'][cand_id], :])
KeyError: 'feat_map'

All keys in info:

dict_keys(['obj_gt_classes', 'obj_scores', 'is_gt', 'labels_ro', 'cand_id', 'pair_ids', 'obj_id', 'labels_r', 'height', 'width', 'boxes', 'is_gt_pair', 'pair_iou', 'obj_classes', 'filename', 'labels_sr', 'pool', 'H_mapping'])

Looking forward to your replay!

We have fixed it in the latest commit.

Answer 3 · 2020-11-23T03:07:37.000Z

BTW, when I run eval.py, it told me that:
RuntimeError: exp/IDN_IPT_hico/epoch_30.pth is a zip archive (did you mean to use torch.jit.load()?)
According to web, this error is due to the version of pytorch. However, my env is torch==1.3.1 same to requirements.txt. How can I deal with this situation? Thanks a lot!

We will update the requirements.txt

Answer 4 · 2020-11-23T03:49:38.000Z

Thank you very much for your reply!

When I train AE again, I print loss every batch. And I found that the Loss explode....Here is the information

I didn't change the code....It seems that this problem is caused by semi_hard loss.

Looking forward to your replay. Thank you very much!

Answer 5 · 2020-11-23T11:44:45.000Z

The semi-hard loss did result in loss exploding sometimes. Using the probability could help a lot in making the loss more stable with little performance drop (less than 0.1 in our experiments). We have updated the code to the more stable version.

Answer 6 · 2020-11-23T13:06:41.000Z

Thank you for your remanding! Now this code can work.
But the loss changes negative. It's this right?

Thank you very much!

Answer 7 · 2020-11-23T13:54:14.000Z

This could be fine, though tuning down the loss scale by weighting the semi-hard loss might bring better performance.

Answer 8 · 2020-11-23T14:13:08.000Z

Thank you very much for your reply. We are looking forward to your improvement of this project, which is very enlightening!

Answer 9 · 2020-11-24T09:07:03.000Z

Thanks for your support! We will keep going to enrich this project.