Zoom CMS loss scales to show late training convergence
Browse filesUse 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 +1 -0
- images/v3.2.0/cms_loss_history.png +2 -2
- images/v3.2.0/cms_loss_history.svg +1514 -676
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 |

|
| 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 |

|
| 46 |
|
images/v3.2.0/cms_loss_history.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|
images/v3.2.0/cms_loss_history.svg
CHANGED
|
|
|
|