config_prior.yaml

params_path:
clip_mean_std_path:
clip_name: ViT-L/14
num_epochs: 2
save_every: 1000
save_name: model
save_path: test
inpainting: false
device: cuda

model_config:
    model:
      type: prior
      diffusion_sampler: uniform
      hparams:
        text_ctx: 77
        xf_width: 2048
        xf_layers: 20
        xf_heads: 32
        xf_final_ln: true
        xf_padding: false
        text_drop: 0.2
        clip_dim: 768
        clip_xf_width: 768

    diffusion:
      steps: 1000
      learn_sigma: false
      sigma_small: true
      noise_schedule: cosine
      use_kl: false
      predict_xstart: true
      rescale_learned_sigmas: false
      timestep_respacing: ''

  
optim_params:
    name: transformers.Adafactor
    params:
      lr: 0.000005
      weight_decay: 0
      scale_parameter: false
      relative_step: false
        
data:
  train:
    df_path: 
    clip_image_size: 224
    drop_text_prob: 0.1
    batch_size: 1
    shuffle: true
    num_workers: 4