Views
No views yet
1import birder
2from birder.inference.classification import infer_image
3
4# Option 1: manual setup (more control over preprocessing)
5net, model_info = birder.load_pretrained_model("rope_vit5_reg4_b16_nepa-bio", inference=True)
6
7# Get the image size the model was trained on
8size = birder.get_size_from_signature(model_info.signature)
9
10# Create an inference transform
11transform = birder.classification_transform(size, model_info.rgb_stats)
12
13# Option 2: helper (quick start with default preprocessing)
14net, model_info, transform = birder.load_pretrained_model_and_transform("rope_vit5_reg4_b16_nepa-bio", inference=True)
15
16image = "path/to/image.jpeg" # or a PIL image
17out, embedding = infer_image(net, image, transform, return_embedding=True)
18# embedding is a NumPy array with shape of (1, 768)1from PIL import Image
2import birder
3
4net, model_info, transform = birder.load_pretrained_model_and_transform("rope_vit5_reg4_b16_nepa-bio", inference=True)
5
6image = Image.open("path/to/image.jpeg")
7features = net.detection_features(transform(image).unsqueeze(0))
8# features is a dict (stage name -> torch.Tensor)
9print([(k, v.size()) for k, v in features.items()])
10# Output example:
11# [('stage1', torch.Size([1, 768, 14, 14]))]1@misc{wang2026vit5visiontransformersmid2020s,
2 title={ViT-5: Vision Transformers for The Mid-2020s},
3 author={Feng Wang and Sucheng Ren and Tiezheng Zhang and Predrag Neskovic and Anand Bhattad and Cihang Xie and Alan Yuille},
4 year={2026},
5 eprint={2602.08071},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV},
8 url={https://arxiv.org/abs/2602.08071},
9}
10
11@misc{xu2025nextembeddingpredictionmakesstrong,
12 title={Next-Embedding Prediction Makes Strong Vision Learners},
13 author={Sihan Xu and Ziqiao Ma and Wenhao Chai and Xuweiyi Chen and Weiyang Jin and Joyce Chai and Saining Xie and Stella X. Yu},
14 year={2025},
15 eprint={2512.16922},
16 archivePrefix={arXiv},
17 primaryClass={cs.CV},
18 url={https://arxiv.org/abs/2512.16922},
19}