| { |
| "_comment": "EgaraNet: DINOv3 ViT-L backbone + StyleNet (Transposed Attention Transformer) composite model config.", |
|
|
| "model_type": "egara_net", |
| "architectures": ["EgaraNetModel"], |
| "torch_dtype": "float32", |
| "transformers_version": "4.56.0.dev0", |
|
|
| "auto_map": { |
| "AutoConfig": "configuration_egara_net.EgaraNetConfig", |
| "AutoModel": "modeling_egara_net.EgaraNetModel" |
| }, |
|
|
| "_section_backbone": "--- DINOv3 ViT Backbone (nested sub-config) ---", |
| "backbone_config": { |
| "model_type": "dinov3_vit", |
| "architectures": ["DINOv3ViTModel"], |
|
|
| "hidden_size": 1024, |
| "num_hidden_layers": 24, |
| "num_attention_heads": 16, |
| "intermediate_size": 4096, |
| "hidden_act": "gelu", |
|
|
| "image_size": 224, |
| "patch_size": 16, |
| "num_channels": 3, |
| "num_register_tokens": 4, |
|
|
| "attention_dropout": 0.0, |
| "drop_path_rate": 0.0, |
| "layer_norm_eps": 1e-05, |
| "layerscale_value": 1.0, |
|
|
| "pos_embed_rescale": 2.0, |
| "pos_embed_jitter": null, |
| "pos_embed_shift": null, |
| "rope_theta": 100.0, |
|
|
| "key_bias": false, |
| "query_bias": true, |
| "value_bias": true, |
| "proj_bias": true, |
| "mlp_bias": true, |
| "use_gated_mlp": false, |
|
|
| "initializer_range": 0.02, |
| "torch_dtype": "float32", |
| "transformers_version": "4.56.0.dev0" |
| }, |
|
|
| "_section_stylenet": "--- StyleNet: Transposed Attention Transformer (TAT) head ---", |
|
|
| "tat_input_dim": null, |
| "_tat_input_dim_note": "null = auto-inferred from backbone_config.hidden_size at model init", |
|
|
| "tat_hidden_dim": 1024, |
| "_tat_hidden_dim_note": "Internal channel dimension of TAT layers. Matches backbone hidden_size for ViT-L.", |
|
|
| "tat_output_dim": 1024, |
| "_tat_output_dim_note": "Final L2-normalised style vector dimension.", |
|
|
| "tat_num_layers": 3, |
| "_tat_num_layers_note": "Number of stacked TransposedAttentionTransformer layers.", |
|
|
| "tat_num_heads": 16, |
| "_tat_num_heads_note": "Number of attention heads in TAT. Must divide tat_hidden_dim evenly. (1024 / 16 = 64)", |
|
|
| "_section_tat_internals": "--- TAT internals (derived / documented for reference) ---", |
| "tat_rms_norm_eps": 1e-05, |
| "tat_swiglu_multiple": 64, |
| "_tat_swiglu_note": "SwiGLU hidden = round_up(floor(hidden_dim * 8/3), multiple). E.g. 768->2048, 1024->2752.", |
|
|
| "_section_attnpool": "--- Attention Pooling ---", |
| "attn_pool_num_heads": 8, |
| "_attn_pool_note": "nn.MultiheadAttention heads used in AttentionPooling. Must divide tat_hidden_dim.", |
|
|
| "_section_head": "--- Projection Head (hidden -> output) ---", |
| "head_act": "silu", |
| "_head_note": "Linear(hidden) -> SiLU -> Linear(output). Output is L2-normalised.", |
|
|
| "_section_preprocessing": "--- Default inference preprocessing ---", |
| "image_size": 512, |
| "keep_aspect_ratio": true, |
| "_keep_aspect_ratio_note": "true = MaxResizeMod16(image_size); false = square resize.", |
| "image_mean": [0.485, 0.456, 0.406], |
| "image_std": [0.229, 0.224, 0.225], |
| "_image_stats_note": "ImageNet stats; must match backbone preprocessor_config.json." |
| } |
|
|
