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