upload aux OSS_fullSFT_stage3_10k_joint30_a30_t30_u10/training_config.yaml

OmniDance-fullSFT/OSS_fullSFT_stage3_10k_joint30_a30_t30_u10/training_config.yaml

@@ -0,0 +1,87 @@
+dataset_base_path: /apdcephfs_gy4/share_302533218/cedricnie/diff/data/omnidance
+dataset_metadata_path: /apdcephfs_gy4/share_302533218/cedricnie/diff/data/omnidance/metadata_train.csv
+dataset_repeat: 1
+dataset_num_workers: 0
+data_file_keys: video,input_audio
+model_paths: null
+model_id_with_origin_paths: Wan2.2-TI2V-5B:diffusion_pytorch_model*.safetensors,Wan2.2-TI2V-5B:models_t5_umt5-xxl-enc-bf16.pth,Wan2.2-TI2V-5B:Wan2.2_VAE.pth
+extra_inputs: input_image,input_audio
+fp8_models: null
+offload_models: null
+learning_rate: 1.0e-05
+audio_learning_rate: 1.0e-05
+dit_learning_rate: 1.0e-05
+plan_learning_rate: null
+lr_warmup_steps: 1500
+lr_warmup_start_factor: 0.0001
+lr_scheduler: linear
+lr_decay_start_step: 10000
+lr_decay_end_step: 12000
+lr_min_factor: 0.2
+num_epochs: 1
+max_steps: 12000
+trainable_models: dit,dit.audio_injector.injector_global_mod_layers
+find_unused_parameters: true
+weight_decay: 0.01
+max_grad_norm: 1.0
+task: sft
+output_path: ./models/train/OSS_fullSFT_stage3_10k_joint30_a30_t30_u10
+remove_prefix_in_ckpt: pipe.dit.
+save_steps: 500
+lora_base_model: null
+lora_target_modules: null
+lora_rank: 32
+lora_checkpoint: null
+preset_lora_path: null
+preset_lora_model: null
+use_gradient_checkpointing: true
+use_gradient_checkpointing_offload: false
+gradient_accumulation_steps: 1
+wandb_project: Wan2.2-TI2V-5B-Audio
+wandb_run_name: OSS_fullSFT_stage3_10k_joint30_a30_t30_u10
+wandb_tags: fullSFT,stage3,S3,10k,TIA2V,joint,audio_only,text_only,uncond
+wandb_init_timeout: 30
+height: 1280
+width: 704
+max_pixels: 1048576
+num_frames: 81
+data_dir: /apdcephfs_gy4/share_302533218/cedricnie/diff/data/omnidance
+tokenizer_path: models/Wan-AI/Wan2.2-TI2V-5B/google/umt5-xxl
+audio_processor_path: null
+dit_variant: audio
+mert_encoder_path: models/MERT-v1-330M
+audio_cache_dir: /apdcephfs_gy4/share_302533218/cedricnie/diff/data/omnidance/audio_cache
+text_prompt_key: null
+motion_prompt_key: null
+motion_prompt_base_key: null
+text_drop_prob: 0.0
+text_drop_prob_start: null
+text_drop_prob_end: null
+text_drop_prob_decay_steps: 0
+audio_drop_prob: 0.0
+condition_quota: joint:0.30,audio_only:0.30,text_only:0.30,uncond:0.10
+condition_quota_start: null
+condition_quota_transition_start_step: 0
+condition_quota_transition_steps: 0
+lora_learning_rate: null
+max_timestep_boundary: 1.0
+min_timestep_boundary: 0.0
+train_checkpoint_path: ./models/train/OSS_fullSFT_stage2_5k_joint45_t45_u10/step-6000.safetensors
+promote_precision_sensitive_params_to_fp32: true
+enable_global_audio_modulation: true
+audio_sink_bias_mode: pass_mass
+audio_sink_enable: null
+enable_plan_schedule: false
+plan_use_planner: false
+plan_use_controller: false
+plan_lambdas: beat:0,sync:0,smooth:0,pose:0
+plan_beat_ramp: ''
+initialize_model_on_cpu: true
+val_dataset_metadata_path: /apdcephfs_gy4/share_302533218/cedricnie/diff/data/omnidance/metadata_val.csv
+val_dataset_base_path: /apdcephfs_gy4/share_302533218/cedricnie/diff/data/omnidance
+val_every_steps: 100000000
+val_num_batches: 16
+val_condition_combos: joint,audio_only,text_only,uncond
+val_seed: 12345
+val_script_path: examples/wanvideo/model_training/validate_loss.py
+val_cuda_visible_devices: null