Upload spatial_temporal_tmin040_smin015/config.yaml with huggingface_hub
Browse files
spatial_temporal_tmin040_smin015/config.yaml
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
generator_ckpt: checkpoints/ode_init.pt
|
| 2 |
+
generator_fsdp_wrap_strategy: size
|
| 3 |
+
real_score_fsdp_wrap_strategy: size
|
| 4 |
+
fake_score_fsdp_wrap_strategy: size
|
| 5 |
+
real_name: Wan2.1-T2V-14B
|
| 6 |
+
fake_name: Wan2.1-T2V-1.3B
|
| 7 |
+
text_encoder_fsdp_wrap_strategy: size
|
| 8 |
+
denoising_step_list:
|
| 9 |
+
- 1000
|
| 10 |
+
- 750
|
| 11 |
+
- 500
|
| 12 |
+
- 250
|
| 13 |
+
warp_denoising_step: true
|
| 14 |
+
ts_schedule: false
|
| 15 |
+
num_train_timestep: 1000
|
| 16 |
+
timestep_shift: 5.0
|
| 17 |
+
guidance_scale: 3.0
|
| 18 |
+
denoising_loss_type: flow
|
| 19 |
+
mixed_precision: true
|
| 20 |
+
seed: 0
|
| 21 |
+
wandb_host: WANDB_HOST
|
| 22 |
+
wandb_key: WANDB_KEY
|
| 23 |
+
wandb_entity: WANDB_ENTITY
|
| 24 |
+
wandb_project: WANDB_PROJECT
|
| 25 |
+
sharding_strategy: hybrid_full
|
| 26 |
+
lr: 2.0e-06
|
| 27 |
+
lr_critic: 4.0e-07
|
| 28 |
+
beta1: 0.0
|
| 29 |
+
beta2: 0.999
|
| 30 |
+
beta1_critic: 0.0
|
| 31 |
+
beta2_critic: 0.999
|
| 32 |
+
data_path: prompts/vidprom_filtered_extended.txt
|
| 33 |
+
batch_size: 1
|
| 34 |
+
ema_weight: 0.99
|
| 35 |
+
ema_start_step: 200
|
| 36 |
+
gradient_accumulation_steps: 8
|
| 37 |
+
log_iters: 200
|
| 38 |
+
negative_prompt: 色调艳丽,过曝,静态,细节模糊不清,字幕,风格,作品,画作,画面,静止,整体发灰,最差质量,低质量,JPEG压缩残留,丑陋的,残缺的,多余的手指,画得不好的手部,画得不好的脸部,畸形的,毁容的,形态畸形的肢体,手指融合,静止不动的画面,杂乱的背景,三条腿,背景人很多,倒着走
|
| 39 |
+
dfake_gen_update_ratio: 5
|
| 40 |
+
image_or_video_shape:
|
| 41 |
+
- 1
|
| 42 |
+
- 21
|
| 43 |
+
- 16
|
| 44 |
+
- 60
|
| 45 |
+
- 104
|
| 46 |
+
distribution_loss: rewardeddmd
|
| 47 |
+
trainer: rewarded_distillation
|
| 48 |
+
gradient_checkpointing: true
|
| 49 |
+
full_training_steps: 1000
|
| 50 |
+
num_frame_per_block: 3
|
| 51 |
+
load_raw_video: false
|
| 52 |
+
beta: 2.0
|
| 53 |
+
model_kwargs:
|
| 54 |
+
timestep_shift: 5.0
|
| 55 |
+
local_attn_size: 9
|
| 56 |
+
sink_size: 3
|
| 57 |
+
checkpoint_path: checkpoints
|
| 58 |
+
spatial_reward: true
|
| 59 |
+
spatial_reward_method: gradient
|
| 60 |
+
spatial_reward_pixel_grad: true
|
| 61 |
+
spatial_reward_metrics:
|
| 62 |
+
- MQ
|
| 63 |
+
- VQ
|
| 64 |
+
- TA
|
| 65 |
+
spatial_reward_strength: 1.0
|
| 66 |
+
spatial_reward_min_weight: 0.15
|
| 67 |
+
spatial_reward_combination: adaptive
|
| 68 |
+
spatial_reward_temperature: 1.0
|
| 69 |
+
reward_mode: Overall
|
| 70 |
+
temporal_saliency_weighting: true
|
| 71 |
+
temporal_saliency_strength: 1.0
|
| 72 |
+
temporal_saliency_min_weight: 0.4
|