data-archetype commited on
Commit
0f964b6
·
verified ·
1 Parent(s): cc82ac6

Upload technical_report_p32.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. technical_report_p32.md +7 -5
technical_report_p32.md CHANGED
@@ -37,11 +37,13 @@ The patchification is applied as an outer wrapper:
37
  Running stats (mean/variance for whitening) are tracked in the patchified
38
  512-channel space.
39
 
40
- Semantic alignment and the VP posterior variance expansion loss operate
41
- **before** patchification, in the 128-channel space. The latent scale
42
- penalty (log-variance L2) operates **after** patchification, in the
43
- 512-channel space. Running stats for whitening are also tracked in the
44
- patchified space.
 
 
45
 
46
  ## Training
47
 
 
37
  Running stats (mean/variance for whitening) are tracked in the patchified
38
  512-channel space.
39
 
40
+ Semantic alignment operates **before** patchification, in the 128-channel
41
+ space. The VP posterior variance expansion loss and the latent scale penalty
42
+ (log-variance L2) operate **after** patchification, in the 512-channel
43
+ space. The variance expansion loss is per-element, so this is equivalent to
44
+ operating before patchification (the 2x2 reshape is lossless and the mean
45
+ over all elements is invariant to the rearrangement). Running stats for
46
+ whitening are also tracked in the patchified space.
47
 
48
  ## Training
49