Views
No views yet
arabian-peninsula dataset (all the relevant bird species found in the Arabian peninsula inc. rarities).
The training followed RSB procedure A2.1import birder
2from birder.inference.classification import infer_image
3
4(net, model_info) = birder.load_pretrained_model("resnet_v1_50_arabian-peninsula", inference=True)
5
6# Get the image size the model was trained on
7size = birder.get_size_from_signature(model_info.signature)
8
9# Create an inference transform
10transform = birder.classification_transform(size, model_info.rgb_stats)
11
12image = "path/to/image.jpeg" # or a PIL image, must be loaded in RGB format
13(out, _) = infer_image(net, image, transform)
14# out is a NumPy array with shape of (1, 735), representing class probabilities.1import birder
2from birder.inference.classification import infer_image
3
4(net, model_info) = birder.load_pretrained_model("resnet_v1_50_arabian-peninsula", inference=True)
5
6# Get the image size the model was trained on
7size = birder.get_size_from_signature(model_info.signature)
8
9# Create an inference transform
10transform = birder.classification_transform(size, model_info.rgb_stats)
11
12image = "path/to/image.jpeg" # or a PIL image
13(out, embedding) = infer_image(net, image, transform, return_embedding=True)
14# embedding is a NumPy array with shape of (1, 2048)1from PIL import Image
2import birder
3
4(net, model_info) = birder.load_pretrained_model("resnet_v1_50_arabian-peninsula", inference=True)
5
6# Get the image size the model was trained on
7size = birder.get_size_from_signature(model_info.signature)
8
9# Create an inference transform
10transform = birder.classification_transform(size, model_info.rgb_stats)
11
12image = Image.open("path/to/image.jpeg")
13features = net.detection_features(transform(image).unsqueeze(0))
14# features is a dict (stage name -> torch.Tensor)
15print([(k, v.size()) for k, v in features.items()])
16# Output example:
17# [('stage1', torch.Size([1, 256, 64, 64])),
18# ('stage2', torch.Size([1, 512, 32, 32])),
19# ('stage3', torch.Size([1, 1024, 16, 16])),
20# ('stage4', torch.Size([1, 2048, 8, 8]))]1@misc{he2015deepresiduallearningimage,
2 title={Deep Residual Learning for Image Recognition},
3 author={Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun},
4 year={2015},
5 eprint={1512.03385},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV},
8 url={https://arxiv.org/abs/1512.03385},
9}
10
11@misc{wightman2021resnetstrikesbackimproved,
12 title={ResNet strikes back: An improved training procedure in timm},
13 author={Ross Wightman and Hugo Touvron and Hervé Jégou},
14 year={2021},
15 eprint={2110.00476},
16 archivePrefix={arXiv},
17 primaryClass={cs.CV},
18 url={https://arxiv.org/abs/2110.00476},
19}