ValueError: not enough values to unpack (expected7, got6)

[WUHE-art]

Hello, when I run python coco_scripts/train.py-- exp_name captioning_model-- batch_size 100-- lr 5e-4, the following questions arise: ValueError: not enough values to unpack (expected7, got6) have you ever encountered such a problem? Looking forward to your reply. Thanks.

[mad-red]

change "from speaksee.data import ImageDetectionsField" into "from data import ImageDetectionsField"

[WUHE-art]

Hello, the error didn't disappear when I modified "from speaksee.data import TextField,ImageDetectionsField", but when I changed "for it, (detections, _, ctrl_det_seqs, ctrl_det_gts, ctrl_det_seqs_test, _, caps_gt)" to "for it, (detections,ctrl_det_seqs, ctrl_det_gts, ctrl_det_seqs_test, _, caps_gt)"", the program worked fine. I don't know why.

[mad-red]

the deleted "_" stands for imgids, and the training doesn't need it

[WUHE-art]

Hello, I have tried to reproduce your code, and all parameters used have not been modified, but the cross-entropy loss function of training is only about 170, which is far from your paper. Looking forward to your reply. Thanks.

[mad-red]

Do you mean the score of the whole framework? Or only validation score during the training procedure of the captioning module?

[WUHE-art]

The final score for the entire framework is very low and far from your paper score.

[mad-red]

Sorry for the problem you meet, I'll retrain the model using the uploaded codes to find why and fix it.

[WUHE-art]

Hello, I used the 2080s training model for a total of 13 days, and here are the results of the final test. It doesn't look good when I run it. I don't know if it's the graphics card.

[mad-red]

It seems strange. Are you using the captioning model after Reinforcement Learning?

[WUHE-art]

Hello, I am very interested in your article, and recently I read the code of the paper. There were a few problems in the code that I hadn't thought through.
In the train.py file, in line 99 of code, detections stands for the object characteristics detected by Faster R-CNN, and ctrl_det_seqs stands for the ordered set of regions. I don't know if I understand this correctly.
Besides, I don't know what ctrl_det_gts stands for. And the meaning represented by ctrl_det_seqs_test and caps_gt in 148 lines of code.
Thank you very much

[mad-red]

Hi, thanks for your interest.
ctrl_det_gts is a sequence of 0 and 1, 1 means the end of describing a region and begin of next regions in next time step, 0 means not end.
Compared to ctrl_det_seqs, ctrl_det_seqs_test has a process of deduplication of regions with same class. If we only use detections of Faster-RCNN to generate captions with the captioning module, we don't know how many times a region with same class appears in the caption. So the deduplication is used, when only testing the caption module's performance.
caps_gt is the ground-truth captions.