taming-transformers/configs/coco_scene_images_transformer.yaml

model:
  base_learning_rate: 4.5e-06
  target: taming.models.cond_transformer.Net2NetTransformer
  params:
    cond_stage_key: objects_bbox
    transformer_config:
      target: taming.modules.transformer.mingpt.GPT
      params:
        vocab_size: 8192
        block_size: 348  # = 256 + 92 = dim(vqgan_latent_space,16x16) + dim(conditional_builder.embedding_dim)
        n_layer: 40
        n_head: 16
        n_embd: 1408
        embd_pdrop: 0.1
        resid_pdrop: 0.1
        attn_pdrop: 0.1
    first_stage_config:
      target: taming.models.vqgan.VQModel
      params:
        ckpt_path: /path/to/coco_epoch117.ckpt  # https://heibox.uni-heidelberg.de/f/78dea9589974474c97c1/
        embed_dim: 256
        n_embed: 8192
        ddconfig:
          double_z: false
          z_channels: 256
          resolution: 256
          in_channels: 3
          out_ch: 3
          ch: 128
          ch_mult:
          - 1
          - 1
          - 2
          - 2
          - 4
          num_res_blocks: 2
          attn_resolutions:
          - 16
          dropout: 0.0
        lossconfig:
          target: taming.modules.losses.DummyLoss
    cond_stage_config:
      target: taming.models.dummy_cond_stage.DummyCondStage
      params:
        conditional_key: objects_bbox

data:
  target: main.DataModuleFromConfig
  params:
    batch_size: 6
    train:
      target: taming.data.annotated_objects_coco.AnnotatedObjectsCoco
      params:
        data_path: data/coco_annotations_100  # substitute with path to full dataset
        split: train
        keys: [image, objects_bbox, file_name, annotations]
        no_tokens: 8192
        target_image_size: 256
        min_object_area: 0.00001
        min_objects_per_image: 2
        max_objects_per_image: 30
        crop_method: random-1d
        random_flip: true
        use_group_parameter: true
        encode_crop: true
    validation:
      target: taming.data.annotated_objects_coco.AnnotatedObjectsCoco
      params:
        data_path: data/coco_annotations_100  # substitute with path to full dataset
        split: validation
        keys: [image, objects_bbox, file_name, annotations]
        no_tokens: 8192
        target_image_size: 256
        min_object_area: 0.00001
        min_objects_per_image: 2
        max_objects_per_image: 30
        crop_method: center
        random_flip: false
        use_group_parameter: true
        encode_crop: true