Model Card for hornet_tiny_7x7_danube-delta

A HorNet image classification model. This model was trained on the danube-delta dataset (all the relevant bird species found int the Danube Delta region).

The species list is derived from data available at https://www.discoverdanubedelta.com/wp-content/uploads/2023/01/BirdsList-ian-2023.pdf.

Note: this is a subset of the eu-common dataset.

Model Details

• Model Type: Image classification and detection backbone

• Model Stats:

  • Params (M): 22.1
  • Input image size: 256 x 256

• Dataset: danube-delta (368 classes)

• Papers:

  • HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions: https://arxiv.org/abs/2207.14284

Model Usage

Image Classification

import birder
from birder.inference.classification import infer_image

(net, model_info) = birder.load_pretrained_model("hornet_tiny_7x7_danube-delta", inference=True)

# Get the image size the model was trained on
size = birder.get_size_from_signature(model_info.signature)

# Create an inference transform
transform = birder.classification_transform(size, model_info.rgb_stats)

image = "path/to/image.jpeg"  # or a PIL image, must be loaded in RGB format
(out, _) = infer_image(net, image, transform)
# out is a NumPy array with shape of (1, 368), representing class probabilities.

Image Embeddings

import birder
from birder.inference.classification import infer_image

(net, model_info) = birder.load_pretrained_model("hornet_tiny_7x7_danube-delta", inference=True)

# Get the image size the model was trained on
size = birder.get_size_from_signature(model_info.signature)

# Create an inference transform
transform = birder.classification_transform(size, model_info.rgb_stats)

image = "path/to/image.jpeg"  # or a PIL image
(out, embedding) = infer_image(net, image, transform, return_embedding=True)
# embedding is a NumPy array with shape of (1, 512)

Detection Feature Map

from PIL import Image
import birder

(net, model_info) = birder.load_pretrained_model("hornet_tiny_7x7_danube-delta", inference=True)

# Get the image size the model was trained on
size = birder.get_size_from_signature(model_info.signature)

# Create an inference transform
transform = birder.classification_transform(size, model_info.rgb_stats)

image = Image.open("path/to/image.jpeg")
features = net.detection_features(transform(image).unsqueeze(0))
# features is a dict (stage name -> torch.Tensor)
print([(k, v.size()) for k, v in features.items()])
# Output example:
# [('stage1', torch.Size([1, 64, 64, 64])),
#  ('stage2', torch.Size([1, 128, 32, 32])),
#  ('stage3', torch.Size([1, 256, 16, 16])),
#  ('stage4', torch.Size([1, 512, 8, 8]))]

Citation

@misc{rao2022hornetefficienthighorderspatial,
      title={HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions},
      author={Yongming Rao and Wenliang Zhao and Yansong Tang and Jie Zhou and Ser-Nam Lim and Jiwen Lu},
      year={2022},
      eprint={2207.14284},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2207.14284},
}