#cwd: ${hydra:runtime.output_dir}
name: adroit
log_frequency: 100
load_path: null
resume_name: null
auto_resume: false
num_node: 1
node_rank: null
dist_usrl: null
gpu: 0
sync_bn: false
tensorboard: false
timestamp: false
seed: null
cudnn_deterministic: false
amp: false
debug: false
opts: null
solver:
  base_lr: 3.0e-06
  adjust_lr: none
  max_epochs: 100
  save_epochs: 20
  validation_epochs: 1
  sample_iterations: 400
  print_specific_things: true
  ema:
    decay: 0.99
    update_interval: 25
    device: cpu
  clip_grad_norm:
    target: diffusion_reward.models.video_models.vqdiffusion.engine.clip_grad_norm.ClipGradNorm
    params:
      start_iteration: 0
      end_iteration: 5000
      max_norm: 0.5
  optimizers_and_schedulers:
  - name: none
    optimizer:
      target: torch.optim.AdamW
      params:
        betas:
        - 0.9
        - 0.96
        weight_decay: 0.045
    scheduler:
      step_iteration: 1
      target: diffusion_reward.models.video_models.vqdiffusion.engine.lr_scheduler.ReduceLROnPlateauWithWarmup
      params:
        factor: 0.5
        patience: 100000
        min_lr: 1.0e-06
        threshold: 0.1
        threshold_mode: rel
        warmup_lr: 0.00045
        warmup: 5000
save_dir: ???
dist_url: ???
ngpus_per_node: ???
world_size: ???
local_rank: ???
global_rank: ???
distributed: false
model:
  target: diffusion_reward.models.video_models.vqdiffusion.modeling.models.frame_conditional_dalle.FC_DALLE
  params:
    content_info:
      key: image
    condition_info:
      key: frame
    frame_skip: 1
    content_codec_config:
      target: diffusion_reward.models.video_models.vqdiffusion.modeling.codecs.image_codec.vqgan.MiniVQGAN
      params:
        args:
          latent_dim: 64
          device: cuda
          image_channels: 3
          num_codebook_vectors: 1024
          beta: 0.25
          channels:
          - 128
          - 128
          - 256
          - 256
          resolution: 64
          latent_size: 8
        trainable: false
        token_shape:
        - 8
        - 8
        ckpt_path: /exp_local/codec_models/vqgan/adroit/results/checkpoints/vqgan.pt
    diffusion_config:
      target: diffusion_reward.models.video_models.vqdiffusion.modeling.transformers.diffusion_transformer.DiffusionTransformer
      params:
        diffusion_step: 100
        alpha_init_type: alpha1
        auxiliary_loss_weight: 0.001
        adaptive_auxiliary_loss: true
        mask_weight:
        - 1
        - 1
        transformer_config:
          target: diffusion_reward.models.video_models.vqdiffusion.modeling.transformers.transformer_utils.Text2ImageTransformer
          params:
            diffusion_step: ???
            content_emb_config: ???
            attn_type: selfcross
            n_layer: 16
            condition_seq_len: 128
            content_seq_len: 64
            content_spatial_size:
            - 8
            - 8
            n_embd: 128
            condition_dim: 1024
            n_head: 16
            attn_pdrop: 0.0
            resid_pdrop: 0.0
            block_activate: GELU2
            timestep_type: adalayernorm
            mlp_hidden_times: 2
            mlp_type: conv_mlp
        condition_emb_config:
          target: diffusion_reward.models.video_models.vqdiffusion.modeling.embeddings.frame_embedding.FrameEmbedding
          params:
            num_embed: 1024
            embed_dim: 1024
            identity: false
            trainable: true
            num_cond_frames: 2
        content_emb_config:
          target: diffusion_reward.models.video_models.vqdiffusion.modeling.embeddings.dalle_mask_image_embedding.DalleMaskImageEmbedding
          params:
            num_embed: 1024
            spatial_size:
            - 8
            - 8
            embed_dim: 128
            trainable: true
            pos_emb_type: embedding
dataloader:
  data_root: /video_dataset/adroit/
  batch_size: 4
  num_workers: 4
  train_datasets:
  - target: diffusion_reward.models.video_models.vqdiffusion.data.dataset.VideoDataset
    params:
      data_root: ${dataloader.data_root}
      phase: train
      frame_skip: ${model.params.frame_skip}
      frames_per_sample: 3
      size: 64
  validation_datasets:
  - target: diffusion_reward.models.video_models.vqdiffusion.data.dataset.VideoDataset
    params:
      data_root: ${dataloader.data_root}
      phase: test
      frame_skip: ${model.params.frame_skip}
      frames_per_sample: 3
      size: 64