jpata commited on
Commit
61853d5
·
verified ·
1 Parent(s): dece87d

Zoom CMS loss scales to show late training convergence

Browse files

Use independent loss axes for pileup and no-pileup groups. Retain full-history panels and add zoom panels for steps 60000 to 100000, preserving the existing data and smoothing.

README.md CHANGED
@@ -40,6 +40,7 @@ CMS uses `group/physical_nopu/loss_Total_unweighted` and
40
  `group/physical_pu/loss_Total_unweighted`. Training points average each group over
41
  the preceding 1,000 steps before smoothing. The panels separate the differing
42
  mixtures: training is 91% no pileup; validation weights the groups equally.
 
43
 
44
  ![CMS training and validation loss by pileup group versus optimizer step](images/v3.2.0/cms_loss_history.png)
45
 
 
40
  `group/physical_pu/loss_Total_unweighted`. Training points average each group over
41
  the preceding 1,000 steps before smoothing. The panels separate the differing
42
  mixtures: training is 91% no pileup; validation weights the groups equally.
43
+ Each panel has its own loss scale; the bottom row zooms into steps 60k–100k.
44
 
45
  ![CMS training and validation loss by pileup group versus optimizer step](images/v3.2.0/cms_loss_history.png)
46
 
images/v3.2.0/cms_loss_history.png CHANGED

Git LFS Details

  • SHA256: 9ce66b9c530fdc1ce92d722292dc3521702afd3a42fd04d0b781ce2ee38a5059
  • Pointer size: 131 Bytes
  • Size of remote file: 123 kB

Git LFS Details

  • SHA256: aaf13e361caa3c2f2d52bde45c1ac2a0fbf4ba01e681977e86b331b7e5219fef
  • Pointer size: 131 Bytes
  • Size of remote file: 282 kB
images/v3.2.0/cms_loss_history.svg CHANGED