Method __init__

generation/tokenizer/vtp_tokenizer.py:15–54

Initialize VTP Tokenizer. Args: hf_model_path: Path to HuggingFace VTPModel directory img_size: Input image size horizon_flip: Horizontal flip probability for data augmentation fp16: Whether to use FP16 precision normalize_type: No

(
        self,
        hf_model_path,
        img_size=256,
        horizon_flip=0.5,
        fp16=True,
        normalize_type="imagenet"
    )

class VTP_Tokenizer:
    def __init__(
        self,
        hf_model_path,
        img_size=256,
        horizon_flip=0.5,
        fp16=True,
        normalize_type="imagenet"
    ):
        """Initialize VTP Tokenizer.

        Args:
            hf_model_path: Path to HuggingFace VTPModel directory
            img_size: Input image size
            horizon_flip: Horizontal flip probability for data augmentation
            fp16: Whether to use FP16 precision
            normalize_type: Normalization type, one of "half" (0.5 mean/std) or "imagenet"
        """
        self.img_size = img_size
        self.horizon_flip = horizon_flip
        self.fp16 = fp16
        self.normalize_type = normalize_type

        # Setup normalization transforms
        self._setup_normalization(normalize_type)

        # Load HuggingFace model
        from vtp.models.vtp_hf import VTPModel
        self.model = VTPModel.from_pretrained(hf_model_path)
        self.model = self.model.cuda().eval()

        config = self.model.config
        self.patch_size = config.vision_patch_size
        self.embed_dim = config.vision_feature_bottleneck

        self.downsample_ratio = self.patch_size
        self.latent_size = img_size // self.downsample_ratio

        print(f"VTP Tokenizer: patch_size={self.patch_size}, embed_dim={self.embed_dim}, "
              f"downsample_ratio={self.downsample_ratio}, latent_size={self.latent_size}, "
              f"normalize={self.normalize_type}")

    def _setup_normalization(self, normalize_type):
        """Setup normalization and inverse normalization transforms."""