I am trying to understand the code, and it seems like during training, the entire clip is labelled as violent or non-violent?

However, during testing, there are test annotations for each frame. I was wondering if this is how the model works, or are there also training annotations being used during training?

Interesting work! Recently, I would like to use the I3D extracted features to conduct some experiments on your dataset. But I found that some errors happened when loading the features from training set, such as:

ValueError: cannot reshape array of size 6558518 into shape (6409,1024)

So I wonder whether these feature files are cropped. I would appreciate it if you can check this issue. Thank you.

Hi, I found the metric mentioned in the paper is a bit confusing, you said corresponding area under the curve (average precision, AP), and in your code you used def precision_recall_curve(y_true, probas_pred, *, pos_label=None, sample_weight=None)to calculate precision and recall then further use auc to get pr_auc. What about average_precision_score, is the function the same as your AP calculation? Thanks

Hi, Thanks for generous sharing of your splendid work! Though the codes are available, I cannot find the requirement.txt or any dockerfile, which makes me difficult to build the same training environment(pytorch version, cuda version, etc) to reproduce or conduct further optimization. So I wonder if you could share the requirement.txt or provide the cuda/pytorch version of the training environment? Thanks!

I am following the paper, and I was able to get the frame level annotations for the testing data, but not the training data. Is this file also made available in the resources? If so, where can I find it? Thanks.

Hello, I find some videoes in 1-1004.zip cannot be palyed completely by the local video player ? For an example, Mission.Impossible.II.2000__#01-49-35_01-51-52_label_B1-0-0.mp4 has more than 10 mins, but only about 2 mins can be played? What's wrong with this ？Can you give me some suggestions to deal with this problem ?

NewAdd.NBA-2017.12.25_CLE@GSW__#01-08-34_01-40-09_label_A.mp4 - CRC 校验错误。 Saving.Private.Ryan.1998__#02-29-31_02-30-55_label_B2-G-0.mp4 - CRC 校验错误。 v=8cTqh9tMz_I__#1_label_A.mp4 - 数据错误 - 该文件已损坏。 v=9eME1y6V-T4__#01-12-00_01-18-00_label_A.mp4 - CRC 校验错误。 出现错误

已经重复下载过2次了，都报一样的错误。其中CRC校验错误的视频仍能播放，但是文件损坏的视频就无法播放了

使用的软件是Bandizip，Win10平台

其他压缩包没有问题

I can't unzip 1-1004.zip and other zip files from trainning videos, which is download from ondrive.when I google it, I got one solution through jar -xvf 1-1004.zip,it can't work,and show this. Maybe someone else get the same problem,did it means there are something wrong with zip files java.util.zip.ZipException: only DEFLATED entries can have EXT descriptor at java.util.zip.ZipInputStream.readLOC(ZipInputStream.java:310) at java.util.zip.ZipInputStream.getNextEntry(ZipInputStream.java:122) at sun.tools.jar.Main.extract(Main.java:979) at sun.tools.jar.Main.run(Main.java:311) at sun.tools.jar.Main.main(Main.java:1288)

Hi, I am trying to extract audio feature on my own, but it does not reproduce the feature you gave. I am trying to run Vggish model, both implemented in tensorflow and pytorch, with any changes.

Are there any changes from the original Vggish code when you extract the features? (ex. sampling rate, hop size... etc.)