Next steps validation

Next steps validation, following discussion with @nik20652:

• train with >20 epochs

• train with window size 800

• validate with SUCCESS bags

• only train with forces

• Create new FrankaData class

• try to narrow best suitable threshold

• validate for more FAILURE bags

• just accuracy score

assigned to @lis07128

created branch 52-next-steps-validation to address this issue

changed the description

marked the checklist item train with >20 epochs as completed

marked the checklist item train with window size 800 as completed

Hi @hanikevi,

I only want to train with force values. What could be the best way to define that? I tried while loading but it wasn't working. My other guess would be in the call function in autoencoder. I am not sure if this would be the best solution if we get the leaves and only hand over the force values to the encoder. What do you think?

I'd suggest you could try making a FrankaData which doesnt have a position attribute. I've been testing the SynthData quite a bit recently, so I think at least all the data loading and training is pretty general over a pytree. You'd need to adjust the franka_data_init here, e.g. take that as an argument in load_dataset, but then should work.

Could also make an autoencoder which ignores jt but IMO we're going to have different data topics as we roll out on different systems so it's probably better that we practice with different pytrees.

changed the description

added backlog label

mentioned in issue #59 (closed)

closed