描述

Swi Trasformer v2 模型在 ImageNet-1k 上以 256x256 的分辨率进行预训练。 Liu 等人在论文《Swi Trasformer V2：Scalig UpCapacity ad Resolutio》中对此进行了介绍。并首次在GitHub中发布。

本版本来自 Huggigface 仓库，用以方便因各种原因无法在原仓库中下载的情况。

如何使用

依赖安装

# 安装 modelscope
!pip istall modelscope

# 使用 sapshot_dowload 下载模型可能需要（这几个包可能需要重启 Rutime）
!pip istall urllib3 --upgrade
!pip istall requests --upgrade
!pip istall spotipy --upgrade

特征提取

from modelscope import sapshot_dowload
from trasformers import AutoImageProcessor, Swiv2Model
from PIL import Image
import requests, torch

model_dir = sapshot_dowload('Alie1996/Mirror_swiv2-base-patch4-widow8-256')
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.ope(requests.get(url, stream=True).raw)

image_processor = AutoImageProcessor.from_pretraied(model_dir)
model = Swiv2Model.from_pretraied(model_dir)

iputs = image_processor(image, retur_tesors="pt")

with torch.o_grad():
    outputs = model(**iputs)

last_hidde_states = outputs.last_hidde_state
list(last_hidde_states.shape)

分类任务

from modelscope import sapshot_dowload
from trasformers import AutoImageProcessor, Swiv2ForImageClassificatio
from PIL import Image
import requests, torch

model_dir = sapshot_dowload('Alie1996/Mirror_swiv2-base-patch4-widow8-256')
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.ope(requests.get(url, stream=True).raw)

image_processor = AutoImageProcessor.from_pretraied(model_dir)
model = Swiv2ForImageClassificatio.from_pretraied(model_dir)

iputs = image_processor(image, retur_tesors="pt")
outputs = model(**iputs)
logits = outputs.logits
# model predicts oe of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
prit("Predicted class:", model.cofig.id2label[predicted_class_idx])

引用

@article{DBLP:jourals/corr/abs-2111-09883,
  author    = {Ze Liu ad
               Ha Hu ad
               Yutog Li ad
               Zhuliag Yao ad
               Zheda Xie ad
               Yixua Wei ad
               Jia Nig ad
               Yue Cao ad
               Zheg Zhag ad
               Li Dog ad
               Furu Wei ad
               Baiig Guo},
  title     = {Swi Trasformer {V2:} Scalig Up Capacity ad Resolutio},
  joural   = {CoRR},
  volume    = {abs/2111.09883},
  year      = {2021},
  url       = {https://arxiv.org/abs/2111.09883},
  eprittype = {arXiv},
  eprit    = {2111.09883},
  timestamp = {Thu, 02 Dec 2021 15:54:22 +0100},
  biburl    = {https://dblp.org/rec/jourals/corr/abs-2111-09883.bib},
  bibsource = {dblp computer sciece bibliography, https://dblp.org}
}

描述 Swin Transformer v2 模型在 ImageNet-1k 上以 256x256 的分辨率进行预训练。 Liu 等人在论文《Swin Transformer V2：Scaling U