로컬 드림부스

1차

데이터셋 12장
Training Steps Per Image   100
Batch Size 1
Gradient Accumulation Steps  1
Class Batch Size 4
Learning Rate Scheduler  linear
Learning Rate 0.0000045
Torch AdamW
Mixed Precision bf16

2차

데이터셋 12장 

Classification 데이터셋 720장
Training Steps Per Image   100
Batch Size 1
Gradient Accumulation Steps  1
Class Batch Size 1
Learning Rate Scheduler  constant
Learning Rate 0.0000045
Torch AdamW
Mixed Precision bf16

3차 

데이터셋 55장

Training Steps Per Image   88
Batch Size 1
Gradient Accumulation Steps  1
Class Batch Size 1
Learning Rate Scheduler  constant
Learning Rate 0.0000035
Torch AdamW
Mixed Precision bf16