Class QwenVL2ImageTransform

data/transforms.py:151–178

        return img
    
class QwenVL2ImageTransform:
    def __init__(
        self, 
        image_size_h, 
        image_size_w, 
        image_stride=14, 
        max_pixels=14*14*9*1024,
        image_mean=OPENAI_CLIP_MEAN, 
        image_std=OPENAI_CLIP_STD,
    ):
        self.processor = Qwen2VLImageProcessor.from_pretrained('InternRobotics/G2VLM-2B-MoT')
        self.img_h = image_size_h
        self.img_w = image_size_w
        self.stride = image_stride

        # self.to_tensor_transform = transforms.ToTensor()
        # self.normalize_transform = transforms.Normalize(mean=image_mean, std=image_std, inplace=True)

    def __call__(self, img, img_num=1):
        # if self.img_h is not None:
        target_size = (self.img_h, self.img_w )
        img = [ii.resize(target_size,3) for ii in img]

        out = self.processor(img, return_tensors='pt')
        pixel_values =  out['pixel_values'] # this is flattened. 
        image_grid_thw =  out['image_grid_thw']

        return pixel_values,  image_grid_thw
        
class ImageTransform:
    def __init__(