configs/train/train_temporal_adapter.yaml

output_dir: "outputs"

pretrained_model_path: ""

motion_module_path: "models/mm_sd_v15_v2.ckpt"

train_data:
  csv_path: "./curated.csv"
  audio_fps: 48000
  audio_size: 480000

validation_data:
  prompts:
    - "./data/input/lighthouse.png"
    - "./data/input/guitar.png"
    - "./data/input/lion.png"
    - "./data/input/gun.png"
  num_inference_steps: 25
  guidance_scale: 7.5
  sample_size: 512

trainable_modules:
  - 'time_conv_in.'
  - 'conv_in.'

video_unet_checkpoint_path: "models/vggsound_unet.ckpt"
audio_unet_checkpoint_path: ""

learning_rate:    5.0e-5
train_batch_size: 1 # max for mixed
gradient_accumulation_steps: 1

max_train_epoch:      -1
max_train_steps:      500000
checkpointing_epochs: 4000
checkpointing_steps:  500

validation_steps:       3000
validation_steps_tuple: [2, 300, 1000]

global_seed: 42
mixed_precision_training: true

is_debug: False

resume_ckpt: ""

zero_no_label_mel: false