gradient accumulation

2026-06-09 00:58:25 +02:00
parent 2551d8dd40
commit ca026ffecb
3 changed files with 30 additions and 25 deletions
@@ -1,5 +1,5 @@
 data:
-  batch_size: 512
+  batch_size: 16
  color_jitter_strength: 0.0
  crop_scale:
    - 1.0
@@ -40,6 +40,7 @@ meta:
  use_bfloat16: true
  use_gradient_checkpointing: true
 optimization:
+  gradient_accumulation_steps: 32
  ema:
    - 0.996
    - 1.0