README.md
4.5 KB · 106 lines · markdown Raw
1 ---
2 license: apache-2.0
3 tags:
4 - object-detection
5 - vision
6 datasets:
7 - coco
8 widget:
9 - src: https://huggingface.co/datasets/mishig/sample_images/resolve/main/savanna.jpg
10 example_title: Savanna
11 - src: https://huggingface.co/datasets/mishig/sample_images/resolve/main/football-match.jpg
12 example_title: Football Match
13 - src: https://huggingface.co/datasets/mishig/sample_images/resolve/main/airport.jpg
14 example_title: Airport
15 ---
16
17 # YOLOS (tiny-sized) model
18
19 YOLOS model fine-tuned on COCO 2017 object detection (118k annotated images). It was introduced in the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Fang et al. and first released in [this repository](https://github.com/hustvl/YOLOS).
20
21 Disclaimer: The team releasing YOLOS did not write a model card for this model so this model card has been written by the Hugging Face team.
22
23 ## Model description
24
25 YOLOS is a Vision Transformer (ViT) trained using the DETR loss. Despite its simplicity, a base-sized YOLOS model is able to achieve 42 AP on COCO validation 2017 (similar to DETR and more complex frameworks such as Faster R-CNN).
26
27 The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
28
29 ## Intended uses & limitations
30
31 You can use the raw model for object detection. See the [model hub](https://huggingface.co/models?search=hustvl/yolos) to look for all available YOLOS models.
32
33 ### How to use
34
35 Here is how to use this model:
36
37 ```python
38 from transformers import YolosImageProcessor, YolosForObjectDetection
39 from PIL import Image
40 import torch
41 import requests
42
43 url = "http://images.cocodataset.org/val2017/000000039769.jpg"
44 image = Image.open(requests.get(url, stream=True).raw)
45
46 model = YolosForObjectDetection.from_pretrained('hustvl/yolos-tiny')
47 image_processor = YolosImageProcessor.from_pretrained("hustvl/yolos-tiny")
48
49 inputs = image_processor(images=image, return_tensors="pt")
50 outputs = model(**inputs)
51
52 # model predicts bounding boxes and corresponding COCO classes
53 logits = outputs.logits
54 bboxes = outputs.pred_boxes
55
56
57 # print results
58 target_sizes = torch.tensor([image.size[::-1]])
59 results = image_processor.post_process_object_detection(outputs, threshold=0.9, target_sizes=target_sizes)[0]
60 for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
61 box = [round(i, 2) for i in box.tolist()]
62 print(
63 f"Detected {model.config.id2label[label.item()]} with confidence "
64 f"{round(score.item(), 3)} at location {box}"
65 )
66 ```
67
68 Currently, both the feature extractor and model support PyTorch.
69
70 ## Training data
71
72 The YOLOS model was pre-trained on [ImageNet-1k](https://huggingface.co/datasets/imagenet2012) and fine-tuned on [COCO 2017 object detection](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
73
74 ### Training
75
76 The model was pre-trained for 300 epochs on ImageNet-1k and fine-tuned for 300 epochs on COCO.
77
78 ## Evaluation results
79
80 This model achieves an AP (average precision) of **28.7** on COCO 2017 validation. For more details regarding evaluation results, we refer to the original paper.
81
82 ### BibTeX entry and citation info
83
84 ```bibtex
85 @article{DBLP:journals/corr/abs-2106-00666,
86 author = {Yuxin Fang and
87 Bencheng Liao and
88 Xinggang Wang and
89 Jiemin Fang and
90 Jiyang Qi and
91 Rui Wu and
92 Jianwei Niu and
93 Wenyu Liu},
94 title = {You Only Look at One Sequence: Rethinking Transformer in Vision through
95 Object Detection},
96 journal = {CoRR},
97 volume = {abs/2106.00666},
98 year = {2021},
99 url = {https://arxiv.org/abs/2106.00666},
100 eprinttype = {arXiv},
101 eprint = {2106.00666},
102 timestamp = {Fri, 29 Apr 2022 19:49:16 +0200},
103 biburl = {https://dblp.org/rec/journals/corr/abs-2106-00666.bib},
104 bibsource = {dblp computer science bibliography, https://dblp.org}
105 }
106 ```