search

google-research / ravens

Train robotic agents to learn pick and place with deep learning for vision-based manipulation in PyBullet. Transporter Nets, CoRL 2020.
https://transporternets.github.io
Apache License 2.0
538 stars 93 forks source link

Question: batch_size=1 for Transporter models? #7

Closed MohitShridhar closed 3 years ago

MohitShridhar commented 3 years ago

Thanks for open-sourcing this great work!

I noticed that all the Transporter models are trained with a batch size of 1. Is there anything preventing the use of larger batches other than memory?

andyzeng commented 3 years ago

Thanks Mohit! Other than memory, there shouldn't be anything preventing the use of larger batches. Note that the input rotations expand the batch size to the number of rotations (default 36). So a batch size of 2 should be expanded to 2 * n_rotations. To support this, you may need to make a few changes to the data preprocessing and loss function.

MohitShridhar commented 3 years ago

Awesome, thanks!

MohitShridhar commented 3 years ago

Hi Andy, sorry, just a quick follow-up: for the evaluations in Table 2 of the paper, were the agents trained for a fixed number of iterations (1k? 10k?). Or did you pick the agents with the best validation loss for each run?

andyzeng commented 3 years ago

Hi Mohit, I believe we trained the agents for multiple fixed iterations (1K, 2K, 5K, 10K, 20K, 40K), then picked the iteration with the best average validation loss between seeds.