LLaVA – Vision Model

众所周知，LLaVA的框架十分简洁，如下：

故构建一个对应的vision Encoder和Projection也相对来说比较简单。

1 Projection

vision-text projector是一层或者N层激活函数为gelu的MLP。

• projector_type: linear
  • nn.Linear()
• projector_type: mlp2x_gelu()
  • nn.Linear()
  • nn.gelu()
  • nn.Linear()

def build_vision_projector(config, delay_load=False, **kwargs):
    projector_type = getattr(config, 'mm_projector_type', 'linear')

    if projector_type == 'linear':
        return nn.Linear(config.mm_hidden_size, config.hidden_size)

    mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
    if mlp_gelu_match:
        mlp_depth = int(mlp_gelu_match.group(1))
        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
        for _ in range(1, mlp_depth):
            modules.append(nn.GELU())
            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
        return nn.Sequential(*modules)

    if projector_type == 'identity':
        return IdentityMap()

    raise ValueError(f'Unknown projector type: {projector_type}')

2 vision Encoder

vision model最重要的是CLIP-VIT Model和CLIP-VIT Processor(切分图片patch)。

• 1 load_model: 设置Model的requires_grad_ = False
• 2 feature_select(): 选择不同的feature，即是否携带cls
• 3 forward(): 切分patch，编码图像。

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

        self.is_loaded = False

        self.vision_tower_name = vision_tower
        self.select_layer = args.mm_vision_select_layer
        self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')

        if not delay_load:
            self.load_model()
        elif getattr(args, 'unfreeze_mm_vision_tower', False):
            self.load_model()
        else:
            self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_tower_name)

    def load_model(self, device_map=None):
        if self.is_loaded:
            print('{} is already loaded, `load_model` called again, skipping.'.format(self.vision_tower_name))
            return

        self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_tower_name)
        self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name, device_map=device_map)
        self.vision_tower.requires_grad_(False)

        self.is_loaded = True

    def feature_select(self, image_forward_outs):
        image_features = image_forward_outs.hidden_states[self.select_layer]
        if self.select_feature == 'patch':
            image_features = image_features[:, 1:]
        elif self.select_feature == 'cls_patch':
            image_features = image_features
        else:
            raise ValueError(f'Unexpected select feature: {self.select_feature}')
        return image_features

    @torch.no_grad()
    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