BC03, 1 Sphere, Ablations

(1) No skip connections, (2) pointwise before SVD, (3) using dense loss
Created on June 5|Last edited on June 10
Comment
Contents:
Overview(06/05/2022) (06/07/2022) (06/08/2022)(06/10/2022) Now the dense lossResultsGIFs, no skip connectionsGIFs, pointwise before SVDGIFs, 6d flow, dense loss in addition to pointwise and consistencyGIFs, 3d flow, dense loss in addition to pointwise and consistency
﻿
Overview
(06/05/2022) Adding ablation of no skip connection.
	https://github.com/Xingyu-Lin/softagent_rpad/commit/5066037b7ee24e83aa35a5d038a7b3f09e6259d4﻿
	Let's see how that does. Compare with 6d flow with keeping skip connections https://wandb.ai/mooey5775/mixed_media/reports/BC03-4DoF-1-Sphere-PN-SVD-Pointwise-6D-flow-ee2flow-100-demos--VmlldzoyMDM5ODgy﻿
(06/07/2022) OK here it seems to be reasonable (strangely with the pouring https://wandb.ai/mooey5775/mixed_media/reports/Pouring-Ablations--VmlldzoyMTI4ODE3 we don't get this). What do the MSEs look like?
Here, train/MSE_loss goes to 0.019 ish, eval/MSE_loss to 0.03 ish.
Before, train/MSE_loss goes to 0.015ish, eval/MSE_loss to 0.02 ish.
Why are these much higher as compared to the pouring env? I do a similar scaling as with pouring right?
ALSO today, running some pointwise before SVD ablations:
﻿https://github.com/Xingyu-Lin/softagent_rpad/commit/aa13215848167768de090f1d58ab1cb2b6f4317c﻿
# 6D flow, but using pointwise loss before SVD.
SVD_PRE_POINTWISE_6D_EE2FLOW_SEP_RT_SVD = dict(
    obs_type='point_cloud',
    act_type='ee2flow_sep_rt',
    encoder_type='pointnet_svd_pointwise_6d_flow',
    method_flow2act='svd',
    use_consistency_loss=False,
    scale_pcl_flow=True,
    scale_pcl_val=250,
    separate_MLPs_R_t=False,
    gaussian_noise_PCL=0.0,  # normally 0
)
(06/08/2022)Putting some results together for the pointwise before SVD. I have 3X on the cluster and Eddie's run 2 (or a third, since one might have crashed).
(06/10/2022) Now the dense lossThis is the new dense loss so we have 3 losses now. TODO -- I need to actually check if this is done correctly again with scaling considered.
We have 3D and 6D flow here. ﻿﻿
Results﻿
Ablation: no skip connections5
Ablation: pointwise before SVD5
Ablation: dense loss (6d flow)5
Ablation: dense loss (3d flow)5
﻿
GIFs, no skip connectionsDo we see a similar thing as with the pouring where skip connections are essential to produce rotations? https://wandb.ai/mooey5775/mixed_media/reports/Pouring-Ablations--VmlldzoyMTI4ODE3﻿
TODO
﻿
GIFs, pointwise before SVDTODO
﻿
GIFs, 6d flow, dense loss in addition to pointwise and consistency﻿
﻿
GIFs, 3d flow, dense loss in addition to pointwise and consistency﻿
Add a comment