Tune-A-VideKO-anything / config.yaml
kyujinpy's picture
Upload config.yaml
3c68167
raw
history blame
1.23 kB
pretrained_model_path: ./checkpoints/kyujinpy/KO-anything-v4-5
output_dir: outputs/bear_guitar
train_data:
video_path: data/bear_guitar.mp4
prompt: A bear is playing guitar
n_sample_frames: 14
width: 512
height: 512
sample_start_idx: 0
sample_frame_rate: 2
validation_data:
prompts:
- ์ž˜์ƒ๊ธด ์™•์ž๊ฐ€ ๊ธฐํƒ€๋ฅผ ์—ฐ์ฃผํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค, ๋ชจ๋˜ํ•œ ๋””์ฆˆ๋‹ˆ ์Šคํƒ€์ผ
- ์—˜ํ”„๊ฐ€ ๊ธฐํƒ€๋ฅผ ์—ฐ์ฃผํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค, ๋ชจ๋˜ํ•œ ๋””์ฆˆ๋‹ˆ ์Šคํƒ€์ผ
- ๋นจ๊ฐ„์ƒ‰ ๋จธ๋ฆฌ๋ฅผ ํ•œ ๊ณต์ฃผ๊ฐ€ ๊ธฐํƒ€๋ฅผ ์—ฐ์ฃผํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค, ๋ชจ๋˜ํ•œ ๋””์ฆˆ๋‹ˆ ์Šคํƒ€์ผ
video_length: 14
width: 512
height: 512
num_inference_steps: 20
guidance_scale: 12.5
use_inv_latent: true
num_inv_steps: 50
validation_steps: 100
trainable_modules:
- attn1.to_q
- attn2.to_q
- attn_temp
train_batch_size: 1
max_train_steps: 400
learning_rate: 3.0e-05
scale_lr: false
lr_scheduler: constant
lr_warmup_steps: 0
adam_beta1: 0.9
adam_beta2: 0.999
adam_weight_decay: 0.01
adam_epsilon: 1.0e-08
max_grad_norm: 1.0
gradient_accumulation_steps: 1
gradient_checkpointing: true
checkpointing_steps: 1000
resume_from_checkpoint: null
mixed_precision: fp16
use_8bit_adam: false
enable_xformers_memory_efficient_attention: true
seed: 33