I wonder train on a mixture ? or train bridge dataset , then train on fractal?

[LukeLIN-web]

19k gradient steps with bridge (~12 epochs) and 30k with fractal (~8 epochs

I wonder train on a mixture ? or train bridge dataset , then train on fractal?

[LukeLIN-web]

I think it is [ only. train bridge then. test on Simpler windowX] , or [ only. train fractal then test on simpler google robot]. , right ?

[LukeLIN-web]

And how to use bridge data is 5Hz and fractal data is 3Hz. information in simpler? did we neeed to change some hyperparameter in simpler?

[LukeLIN-web]

I trained 12hours with 8 h100, per_device_batch_size:32 it still bad on windowX robot . Don't know how to achieve readme said 12hours achieve good results , maybe because I don't use torch copmile.

[LukeLIN-web]

I have trained 48 hours in fractal with 4 h100. But it It only has 1% in pick coke can.