vision_tower.py

'''
CopyRight @DeepGlint 2025
'''


import torch
import torch.nn as nn
from transformers import CLIPVisionModel, CLIPImageProcessor, CLIPVisionConfig


def build_vision_tower(model_cfg, **kwargs):
    vision_tower = getattr(model_cfg, "vision_tower_config", getattr(model_cfg, "vision_tower", None))
    return CLIPVisionTower(vision_tower, args=model_cfg, **kwargs)


class CLIPVisionTower(nn.Module):
    def __init__(self, vision_tower, args, delay_load=False):
        super().__init__()

        self.is_loaded = False

        self.vision_tower_cfg = vision_tower
        self.vision_tower_processor = args.vision_tower_processor
        self.select_layer = args.mm_vision_select_layer
        self.select_feature = getattr(args, "mm_vision_select_feature", "patch")

        if not delay_load:
            self.init_model()
        elif getattr(args, "unfreeze_mm_vision_tower", False):
            # TODO: better detector is needed.
            self.init_model()
        elif hasattr(args, "mm_tunable_parts") and "mm_vision_tower" in args.mm_tunable_parts:
            self.init_model()
        else:
            raise RuntimeError("Not support now, please check config.json or contact us")

    def init_model(self, device_map=None):
        if self.is_loaded:
            return
        vision_tower_config = CLIPVisionConfig().from_dict(self.vision_tower_cfg)
        self.image_processor = CLIPImageProcessor(**self.vision_tower_processor)
        self.vision_tower = CLIPVisionModel(config=vision_tower_config)
        self.vision_tower.requires_grad_(False)

        self.is_loaded = True

    def feature_select(self, image_forward_outs):
        select_feature_type = self.select_feature

        if self.select_feature in ["slicefour_patch", "slicefour_cls_patch"]:
            select_every_k_layer = len(image_forward_outs.hidden_states) // 4
            image_features = torch.cat([image_forward_outs.hidden_states[i] for i in range(select_every_k_layer + self.select_layer, len(image_forward_outs.hidden_states), select_every_k_layer)], dim=-1)
            select_feature_type = select_feature_type.replace("slicefour_", "")
        elif self.select_feature in ["slice_m25811_f6_patch", "slice_m25811_f6_cls_patch"]:
            select_layers = [-2, -5, -8, -11, 6]
            image_features = torch.cat([image_forward_outs.hidden_states[i] for i in select_layers], dim=-1)
            select_feature_type = select_feature_type.replace("slice_m25811_f6_", "")
        else:
            image_features = image_forward_outs.hidden_states[self.select_layer]

        if select_feature_type == "patch":
            image_features = image_features[:, 1:]
        elif select_feature_type == "cls_patch":
            image_features = image_features
        else:
            raise ValueError(f"Unexpected select feature: {select_feature_type}")
        return image_features

    def forward(self, images):
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = self.feature_select(image_forward_out).to(image.dtype)
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
            image_features = self.feature_select(image_forward_outs).to(images.dtype)

        return image_features

    @property
    def dummy_feature(self):
        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)

    @property
    def dtype(self):
        return self.vision_tower.dtype

    @property
    def device(self):
        return self.vision_tower.device

    @property
    def config(self):
        if self.is_loaded:
            return self.vision_tower.config
        else:
            return self.cfg_only

    @property
    def hidden_size(self):
        _hidden_size = self.config.hidden_size
        if "slicefour" in self.select_feature:
            _hidden_size *= 4
        if "slice_m25811_f6" in self.select_feature:
            _hidden_size *= 5
        return _hidden_size

    @property
    def num_patches_per_side(self):
        return self.config.image_size // self.config.patch_size

    @property
    def num_patches(self):
        _num_patches = (self.config.image_size // self.config.patch_size) ** 2
        if "cls_patch" in self.select_feature:
            _num_patches += 1
        return _num_patches

    @property
    def image_size(self):
        return self.config.image_size