| """PatchSVAE configuration for HuggingFace AutoModel.""" |
|
|
| from transformers import PretrainedConfig |
|
|
|
|
| class PatchSVAEConfig(PretrainedConfig): |
| """Configuration for PatchSVAE — Fresnel geometric compression lens. |
| |
| Args: |
| matrix_v: Number of rows per patch matrix (vocabulary size) |
| D: Embedding dimension (number of singular values per patch) |
| patch_size: Spatial patch size in pixels |
| hidden: MLP hidden width for encoder/decoder |
| depth: Number of residual blocks in encoder and decoder |
| n_cross_layers: Number of spectral cross-attention layers |
| max_alpha: Maximum coordination strength per spectral mode |
| alpha_init: Initial alpha logit (sigmoid(alpha_init) * max_alpha) |
| target_cv: Soft hand CV target for training |
| image_size: Expected input image size (H=W) |
| """ |
| model_type = "patchsvae" |
|
|
| def __init__( |
| self, |
| matrix_v=256, |
| D=16, |
| patch_size=16, |
| hidden=768, |
| depth=4, |
| n_cross_layers=2, |
| max_alpha=0.2, |
| alpha_init=-2.0, |
| target_cv=0.125, |
| image_size=128, |
| **kwargs, |
| ): |
| self.matrix_v = matrix_v |
| self.D = D |
| self.patch_size = patch_size |
| self.hidden = hidden |
| self.depth = depth |
| self.n_cross_layers = n_cross_layers |
| self.max_alpha = max_alpha |
| self.alpha_init = alpha_init |
| self.target_cv = target_cv |
| self.image_size = image_size |
|
|
| |
| self.n_patches = (image_size // patch_size) ** 2 |
| self.patch_dim = 3 * patch_size * patch_size |
| self.mat_dim = matrix_v * D |
| self.latent_channels = D |
| self.latent_size = image_size // patch_size |
|
|
| super().__init__(**kwargs) |
