towhee.models.visualization.transformer_visualization.generate_attention

towhee.models.visualization.transformer_visualization.generate_attention(model: Module, image_tensor: Tensor, method='transformer_attribution', class_index: Optional[int] = None, img_size: int = 224, device: Optional[str] = None) → ndarray

Generate attention visualization of the model. https://openaccess.thecvf.com/content/CVPR2021/papers/ Chefer_Transformer_Interpretability_Beyond_Attention_Visualization_CVPR_2021_paper.pdf :param model: Model to visualize. :type model: nn.Module :param image_tensor: Image tensor. :type image_tensor: torch.Tensor :param method: full, rollout or transformer_attribution. Default is transformer_attribution. :type method: str :param class_index: Class index. :type class_index: int :param img_size: Image size. :type img_size: int :param device: Model device, cpu or cuda :type device: str

Returns:

Attention map.

Return type:

(numpy.ndarray)