EveryDream-trainer/configs/stable-diffusion/v1-finetune_huge.yaml

model:
  base_learning_rate: 1.0e-6
  target: ldm.models.diffusion.ddpm.LatentDiffusion
  params:
    linear_start: 0.00085
    linear_end: 0.0120
    num_timesteps_cond: 1
    log_every_t: 300
    timesteps: 1000
    first_stage_key: image
    cond_stage_key: caption
    image_size: 64
    channels: 4
    cond_stage_trainable: true
    conditioning_key: crossattn
    monitor: val/loss_simple_ema
    scale_factor: 0.18215
    use_ema: False
    unfreeze_model: True
    model_lr: 1.0e-6

    unet_config:
      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
      params:
        image_size: 32 # unused
        in_channels: 4
        out_channels: 4
        model_channels: 320
        attention_resolutions: [ 4, 2, 1 ]
        num_res_blocks: 2
        channel_mult: [ 1, 2, 4, 4 ]
        num_heads: 8
        use_spatial_transformer: True
        transformer_depth: 1
        context_dim: 768
        use_checkpoint: True
        legacy: False

    first_stage_config:
      target: ldm.models.autoencoder.AutoencoderKL
      params:
        embed_dim: 4
        monitor: val/rec_loss
        ddconfig:
          double_z: true
          z_channels: 4
          resolution: 512
          in_channels: 3
          out_ch: 3
          ch: 128
          ch_mult:
          - 1
          - 2
          - 4
          - 4
          num_res_blocks: 2
          attn_resolutions: []
          dropout: 0.0
        lossconfig:
          target: torch.nn.Identity

    cond_stage_config:
      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

data:
  target: main.DataModuleFromConfig
  params:
    batch_size: 6  # prefer highest possible without getting CUDA Out of Memory error, A100 40GB =~20 80GB= ~48
    num_workers: 8
    wrap: falsegit
    train:
      target: ldm.data.every_dream.EveryDreamBatch
      params:
        repeats: 1   # suggest 1 for 10k+ images
        debug_level: 1   # 1 to print if images are dropped due to multiple-aspect ratio image batching
        conditional_dropout: 0.04   # experimental, likelihood to drop the caption, may help with poorly captioned images
        resolution: 512   # use 512 for 24GB, can use 576, 640, 704, 768, on higher VRAM cards only..
        seed: -1  #  use -1 for random seed, affects ordering of images and shuffling
    validation:
      target: ldm.data.ed_validate.EDValidateBatch
      params:
        repeats: 0.2  # suggest low fractions for 10k+ images
    test:
      target: ldm.data.ed_validate.EDValidateBatch
      params:
        repeats: 0.2

lightning:
  modelcheckpoint:
    params:
      every_n_epochs: 1  # produce a ckpt every epoch, leave 1!
      #every_n_train_steps: 1400 # can only use epoch or train step checkpoints
      save_top_k: 6   # save the best N ckpts according to loss, can reduce to save disk space but suggest at LEAST 2, more if you have max_epochs below higher!
      save_last: True
      filename: "{epoch:02d}-{step:05d}"
  callbacks:
    image_logger:
      target: main.ImageLogger
      params:
        batch_frequency: 500
        max_images: 16
        increase_log_steps: False

  trainer:
    benchmark: True
    max_epochs: 7   # better to run several epochs and test your checkpoints!  Try 4-5, you get a checkpoint every epoch to test! 
    max_steps: 99000   # better to end on epochs not steps, especially with >500 images to ensure even distribution, but you can set this if you really want...
    check_val_every_n_epoch: 1
    gpus: 0,