boheumd / a2summ Goto Github PK

I saw in the paper that you reported both F1, rho and tau. However, when I tried to reproduce the results, I can not find any function that calculate rho and tau result. How could I produce the value of rho and tau?

Dear author, when I run the program according to the running code you provided, I always report an error showing that there is no pywrapknapsack_solver in ortools.algorithms, may I ask where the error occurred, sincerely looking forward to your reply

Thanks for this great work and congratulations for the paper being accepted in CVPR 2023. Can you please provide the inference code for a single video? It will be extremely helpful.

Hi author, I would like to ask how you got the transcribed text corresponding to the videos in the SumMe and TVSum datasets? Was it created manually or did you use an existing model?I am very much looking forward to your answer.

hello, how does the evaluation metrics of Summe and TVSum datasets τ and ρ calculate? I only see the calculation method of fscore in the code.
Looking forward to getting your reply, thank you very much!

When I ran the code, the Terminal got the error "ModuleNotFoundError: No module named 'ortools.algorithms.pywrapknapsack_solver'". And I have installed "ortools", could you please help me?
And how can I get the F-socre of each video in SumMe/TVSum?
Thanks.

Hi!

thanks for your contribution for open source. I just read A2Summ and very interested in your model, Saddly, I found the repo only contains README, and my question is when will complete code of A2Summ be committed?

Best wishes!

When will the code be made public?

For each split (0-4), did you select the highest f-score of all epochs and average score of the 5 splits as result?

Thanks for sharing your open-source code!
I have run the project and got F-score on the SumMe and TVSum datasets. But the results seem different from those in the paper. The following are the results I got.

For SumMe:
F1_results: {'split0': 0.4879799819599918, 'split1': 0.5798566120133917, 'split2': 0.5529651338830718, 'split3': 0.5264276927288803, 'split4': 0.4501540520029801}
F1-score: 0.5195

For TVSum:
F1_results: {'split0': 0.6302811067187952, 'split1': 0.5925235378291761, 'split2': 0.6476801856701864, 'split3': 0.6284219094681587, 'split4': 0.6346191839437079}
F1-score: 0.6267

The results on the SumMe dataset are quite different from those in the paper.
The parameters in the code have no other modification except setting num_workers to 0 due to the code running error. I completed the experiment with the following settings, which are inconsistent with the settings in the readme file.

• pytorch 1.13.1, cuda 11.6, and torchversion 0.14.1
  I think the difference in these settings may not be the main reason for so much difference.
  Could you tell me what causes this?
  Looking forward to your answer.

Hi, I wanted to know what version of CLIP you used and how did you get the features for Roberta ? Is it through average pooling or you took the features of the [CLS] token ?

Thx for your great work！But I am confused by the labels in the multimodel datasets(CNN and Daily Mail).How did you get the labels for the summary?I have read the aritle for many times, but I can't find anything about it. Thx for your reply!

running a new dataset for your model
getting error "h5py objects cannot be pickled"
All requirements are satisfied and are taken special care

Dear author, if I want to test my own dataset, how can I generate h5 files and please provide sample code

Hi Bo,
when I choose a CNN data set training model, it is impossible to generate the model_best_video.pt, and only the model_best_text.pt. However, when using the Daliy_mail dataset training model, the model_best_text.pt and model_best_video.pt can be generated. I checked that the code did not find the reason. Do you know what the reason is?
Looking forward to your reply, thank you very much, I wish you all the best!

Thanks for your contribution for open source!
I just reviewed logs using tensorboard. It seems that before using SummaryWriter.add_scalar, may be you need to add some prefix like split_{index} for datasets like SumMe.

image above generated from training A2summ using SumMe dataset.