The lack of activitynet dataset.

[hugddygff]

About three weeks ago, I asked some questions about the download of the dataset. Now I have download about 18000 video clips, however about 2000 videos are not available. So if possible, I want to ask for help from you. I have sent an email with the lost video ids to your mailbox.

[tanghaoyu258]

excuse me, i have some trouble when reimplementing your model through Pytorch. Could you please tell me the effects of the function 'group_by_window' you've applied on the training dataset? i've read the official explanation of the function and i guess it might split the training set according to the length of the reference video divided by the bucket_span. Am i right ? But i also noticed that some reference video with length larger than 300 have been droped after such processing, which confused me a lot . i'd be very appreciate if you could help me ...

[fengyang0317]

@tanghaoyu258
group_by_window is used to improve running speed. It is safe to remove it.
dynamic_rnn will be run n steps, where n is the length of the longest sequence in a batch.

[tanghaoyu258]

Thank u for your explanation!