ammesatyajit / VideoBERT

Using VideoBERT to tackle video prediction
113 stars 14 forks source link

time for pretraining #3

Open Tortoise17 opened 3 years ago

Tortoise17 commented 3 years ago

I have to ask that do you have any figures how much this takes for pretraining and on which GPU and which dataset_ ? please if you can guide.

harshraj32 commented 3 years ago

where u able to download the dataset?

ammesatyajit commented 3 years ago

So for the dataset, I used the HowTo100M dataset and filtered out the cooking videos. The specific ids for only cooking videos are listed in VideoBERT/data/ids.txt. Here are the steps I took to download the dataset: go to this link: https://www.di.ens.fr/willow/research/howto100m/ On the bottom of the page there should be a download section. To get access to the private server to download it easily, you will need to fill out the form that it links. The creator of the dataset should respond within 2-3 days giving you a login for the server. Using that, you can download the videos with the instructions given along with the ids.txt file.

@Tortoise17 Unfortunately I do not have the pretraining figures for this model. I do not recall which GPU I was using at the time since I was using a VM. However, this should run on a google colab equivalent GPU if the model size is decreased a bit. Hope that helps.

harshraj32 commented 3 years ago

yeah there are over 47k ids, what would be the size of all 47k videos?

ammesatyajit commented 3 years ago

It was about 400-500 GB