模型:
optimum/segformer-b0-finetuned-ade-512-512
SegFormer模型在分辨率为512x512的ADE20k上进行了微调。这个模型是由Xie等人在 SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers 论文中提出的,并在 this repository 中首次发布。
注意:发布SegFormer的团队并没有为这个模型编写模型卡片,因此这个模型卡片是由Hugging Face团队编写的。
SegFormer由一个分层Transformer编码器和一个轻量级的全MLP解码头组成,可以在ADE20K和Cityscapes等语义分割基准上取得很好的结果。分层Transformer首先在ImageNet-1k上进行预训练,然后添加解码头并在下游数据集上进行全面微调。
您可以使用原始模型进行语义分割。可以查看 model hub 以查找您感兴趣的任务的微调版本。
以下是如何使用此模型对COCO 2017数据集的图像进行分类为1,000个ImageNet类别之一的方法:
from transformers import SegformerImageProcessor from PIL import Image import requests from optimum.onnxruntime import ORTModelForSemanticSegmentation image_processor = SegformerImageProcessor.from_pretrained("optimum/segformer-b0-finetuned-ade-512-512") model = ORTModelForSemanticSegmentation.from_pretrained("optimum/segformer-b0-finetuned-ade-512-512") url = "http://images.cocodataset.org/val2017/000000039769.jpg" image = Image.open(requests.get(url, stream=True).raw) inputs = image_processor(images=image, return_tensors="pt").to(device) outputs = model(**inputs) logits = outputs.logits # shape (batch_size, num_labels, height/4, width/4)
如果使用pipeline:
from transformers import SegformerImageProcessor, pipeline from optimum.onnxruntime import ORTModelForSemanticSegmentation image_processor = SegformerImageProcessor.from_pretrained("optimum/segformer-b0-finetuned-ade-512-512") model = ORTModelForSemanticSegmentation.from_pretrained("optimum/segformer-b0-finetuned-ade-512-512") url = "http://images.cocodataset.org/val2017/000000039769.jpg" pipe = pipeline("image-segmentation", model=model, feature_extractor=image_processor) pred = pipe(url)
有关更多代码示例,请参阅 Optimum documentation 。
有关此模型的许可证,请查看 here 。
@article{DBLP:journals/corr/abs-2105-15203, author = {Enze Xie and Wenhai Wang and Zhiding Yu and Anima Anandkumar and Jose M. Alvarez and Ping Luo}, title = {SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers}, journal = {CoRR}, volume = {abs/2105.15203}, year = {2021}, url = {https://arxiv.org/abs/2105.15203}, eprinttype = {arXiv}, eprint = {2105.15203}, timestamp = {Wed, 02 Jun 2021 11:46:42 +0200}, biburl = {https://dblp.org/rec/journals/corr/abs-2105-15203.bib}, bibsource = {dblp computer science bibliography, https://dblp.org} }