| """``transformers`` wrapper of encoder-tvsd: ``AutoModel.from_pretrained(...)`` returns this class, |
| which runs the plain-PyTorch model in ``modeling.py``.""" |
|
|
| from transformers import PreTrainedModel |
|
|
| from .configuration_encoder_tvsd import EncoderTVSDConfig |
| from .modeling import EncoderTVSD, variant_config |
|
|
|
|
| class EncoderTVSDModel(PreTrainedModel): |
| config_class = EncoderTVSDConfig |
| base_model_prefix = "encoder" |
| main_input_name = "images" |
|
|
| @classmethod |
| def from_pretrained(cls, pretrained_model_name_or_path, *args, variant=None, **kwargs): |
| """``variant`` selects a variant (default: the repository's ``default_variant``). |
| ``transformers`` then loads only that variant's weights, ``model.<variant>.safetensors`` |
| (the default variant's are ``model.safetensors``).""" |
|
|
| config = kwargs.get("config") |
| if not isinstance(config, EncoderTVSDConfig): |
| hub = ("revision", "cache_dir", "token", "local_files_only", "force_download") |
| config = cls.config_class.from_pretrained( |
| pretrained_model_name_or_path, **{k: kwargs[k] for k in hub if k in kwargs} |
| ) |
| kwargs["config"] = config |
| name = variant or config.default_variant |
| if name not in config.variants: |
| raise ValueError(f"unknown variant {name!r}; available: {', '.join(config.variants)}") |
| config.selected_variant = name |
| weights = None if name == config.default_variant else name |
| return super().from_pretrained( |
| pretrained_model_name_or_path, *args, variant=weights, **kwargs |
| ) |
|
|
| def __init__(self, config: EncoderTVSDConfig): |
| super().__init__(config) |
| selected = getattr(config, "selected_variant", None) |
| self.encoder = EncoderTVSD(variant_config(config.to_dict(), selected)) |
| self.post_init() |
| self.eval() |
|
|
| @property |
| def variant(self) -> str: |
| return self.encoder.variant |
|
|
| def _init_weights(self, module): |
| """All weights come from the checkpoint.""" |
|
|
| @property |
| def monkeys(self) -> list[str]: |
| return self.encoder.monkeys |
|
|
| @property |
| def areas(self) -> list[str]: |
| return list(self.encoder.areas) |
|
|
| def features(self, images, area: str, batch_size: int = 32): |
| """See ``EncoderTVSD.features``: the readout input of ``area``.""" |
|
|
| return self.encoder.features(images, area, batch_size=batch_size) |
|
|
| def forward(self, images, areas=None, batch_size: int = 32): |
| """See ``EncoderTVSD.forward``: ``{monkey: {area: (images, electrodes)}}``.""" |
|
|
| return self.encoder(images, areas=areas, batch_size=batch_size) |
|
|
|
|
| __all__ = ["EncoderTVSDModel"] |
|
|
