Question about DAgger training epochs

[xu-yang16]

Hi, thanks for sharing the work!

I have a question about the DAgger training setup. The paper mentions 75k training steps with a batch size of 720. Assuming the data is sampled at 30 Hz, 10 hours of data corresponds to about 1.08M samples, or ~1.5k steps per epoch.

Does this mean the initial 75k-step finetuning effectively goes through the 10-hour dataset about 50 epochs?

Also, for DAgger, should the 80:10:10 ratio be interpreted as a sampling ratio between previous data / interventions / remaining rollout data, rather than their actual dataset-size ratio?

Thanks!

[ArthurAllshire]

I have a question about the DAgger training setup. The paper mentions 75k training steps with a batch size of 720. Assuming the data is sampled at 30 Hz, 10 hours of data corresponds to about 1.08M samples, or ~1.5k steps per epoch.

Hi, yes this is correct. I think we over-trained for this as training for less epochs didn't substantially change the performance (you could probably get away with 1/4 as many epochs with similar behaviour, IIRC the behaviour of 25k steps isn't that different). But generally as a rule for these post-training things training for longer can often help hence the large number of epochs.

Also, for DAgger, should the 80:10:10 ratio be interpreted as a sampling ratio between previous data / interventions / remaining rollout data, rather than their actual dataset-size ratio?

Yes.