search

real-stanford / im2Flow2Act

[CoRL 2024] Im2Flow2Act: Flow as the Cross-domain Manipulation Interface
https://im-flow-act.github.io/
MIT License
62 stars 3 forks source link

The recommended GPU configuration to train the flow generation #6

Closed ZorAttC closed 1 month ago

ZorAttC commented 1 month ago

Dear Author, Thank you for your amazing work and generou open source behaviour. And I want to train my own flow generation model using the real world dataset.How many demonstrations for each task do you think that will make the training to converge? And as a greenhand to DL, I wanna know how much GPU RAM is needed to train the model,any specific info? Again Thank you so much.

mengdaxu commented 1 month ago

Hi,

Thanks for your kind words. We train our model on 4 L40s for roughly 1 or 2 days.

In terms of number of demonstrations, it really depends on your specific tasks. I would say roughly 100 to 200 demonstrations is a good starting point.

ZorAttC commented 1 month ago

Thank you for the supportive response.