config_imagenet.yaml

model:
  params:
    condition_info:
      key: label
    content_codec_config:
      params:
        ckpt_path: OUTPUT/pretrained_model/taming_dvae/vqgan_imagenet_f16_16384.pth
        config_path: OUTPUT/pretrained_model/taming_dvae/vqgan_imagenet_f16_16384.yaml
        mapping_path: ./help_folder/statistics/taming_vqvae_974.pt
        num_tokens: 16384
        quantize_number: 974
        token_shape:
        - 16
        - 16
        trainable: false
      target: image_synthesis.modeling.codecs.image_codec.taming_gumbel_vqvae.TamingVQVAE
    content_info:
      key: image
    diffusion_config:
      params:
        adaptive_auxiliary_loss: true
        alpha_init_type: alpha1
        auxiliary_loss_weight: 0.001
        condition_emb_config:
          params:
            embed_dim: 512
            identity: true
            num_embed: 1000
          target: image_synthesis.modeling.embeddings.class_embedding.ClassEmbedding
        content_emb_config:
          params:
            embed_dim: 512
            num_embed: 974
            pos_emb_type: embedding
            spatial_size: !!python/tuple
            - 32
            - 32
            trainable: true
          target: image_synthesis.modeling.embeddings.dalle_mask_image_embedding.DalleMaskImageEmbedding
        diffusion_step: 100
        mask_weight:
        - 1
        - 1
        transformer_config:
          params:
            attn_pdrop: 0.0
            attn_type: selfcondition
            block_activate: GELU2
            class_number: 1000
            class_type: adalayernorm
            content_seq_len: 256
            content_spatial_size:
            - 16
            - 16
            mlp_hidden_times: 4
            mlp_type: conv_mlp
            n_embd: 512
            n_head: 16
            n_layer: 24
            resid_pdrop: 0.0
            timestep_type: adalayernorm
          target: image_synthesis.modeling.transformers.transformer_utils.Condition2ImageTransformer
      target: image_synthesis.modeling.transformers.diffusion_transformer.DiffusionTransformer
  target: image_synthesis.modeling.models.conditional_dalle.C_DALLE