Mask2Former模型是在ADE20k语义分割上训练的(基本尺寸版本,Swin骨干网络)。它在 Masked-attention Mask Transformer for Universal Image Segmentation 论文中被介绍,并在 this repository 中首次发布。
免责声明:发布Mask2Former的团队没有为此模型编写模型卡片,因此该模型卡片是由Hugging Face团队编写的。
Mask2Former通过预测一组掩码和相应的标签来处理实例、语义和全景分割这三个任务。因此,所有这三个任务都被视为实例分割。Mask2Former通过(i)用更先进的多尺度可变形注意力Transformer替换像素解码器,(ii)采用具有掩码注意力的Transformer解码器以提高性能而不引入额外的计算量,并且(iii)通过在子采样点计算损失而不是整个掩码来提高训练效率,从而优于以前的SOTA模型 MaskFormer ,无论是性能还是效率方面。
您可以使用此特定检查点进行全景分割。请参阅 model hub 以查找您感兴趣的任务上的其他经过微调版本。
以下是如何使用此模型的方法:
import requests import torch from PIL import Image from transformers import AutoImageProcessor, Mask2FormerForUniversalSegmentation # load Mask2Former fine-tuned on ADE20k semantic segmentation processor = AutoImageProcessor.from_pretrained("facebook/mask2former-swin-base-ade-semantic") model = Mask2FormerForUniversalSegmentation.from_pretrained("facebook/mask2former-swin-base-ade-semantic") url = "http://images.cocodataset.org/val2017/000000039769.jpg" image = Image.open(requests.get(url, stream=True).raw) inputs = processor(images=image, return_tensors="pt") with torch.no_grad(): outputs = model(**inputs) # model predicts class_queries_logits of shape `(batch_size, num_queries)` # and masks_queries_logits of shape `(batch_size, num_queries, height, width)` class_queries_logits = outputs.class_queries_logits masks_queries_logits = outputs.masks_queries_logits # you can pass them to processor for postprocessing predicted_semantic_map = processor.post_process_semantic_segmentation(outputs, target_sizes=[image.size[::-1]])[0] # we refer to the demo notebooks for visualization (see "Resources" section in the Mask2Former docs)
有关更多代码示例,请参阅 documentation 。