Model card for cspnext_l.rsb_a1_in1k

A CSPNeXt image classification model. Pretrained on ImageNet-1k by OpenMMLab (RTMDet) and converted to timm format.

Name disambiguation: This is the CSPNeXt backbone of RTMDet (OpenMMLab), as implemented in MMDetection. A separate paper, CSPNeXt: A new efficient token hybrid backbone (Chen et al., EAAI 2024, doi:10.1016/j.engappai.2024.107886), uses the same name for a different architecture. These weights do not implement that paper; please cite RTMDet (below) for this model.

Model Details

• Model Type: Image classification / feature backbone
• Model Stats:
  • Params (M): 27.16
  • GMACs: 4.18
  • Activations (M): 8.68
  • Image size: 224 x 224
• Papers:
  • RTMDet: An Empirical Study of Designing Real-Time Object Detectors: https://arxiv.org/abs/2212.07784
• Original:
  • https://github.com/open-mmlab/mmdetection/tree/main/configs/rtmdet/classification
  • https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-l_8xb256-rsb-a1-600e_in1k-6a760974.pth
• Dataset: ImageNet-1k

Model Usage

Image Classification

from urllib.request import urlopen
from PIL import Image
import timm
import torch

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model('hf_hub:munehiro-k/cspnext_l.rsb_a1_in1k', pretrained=True)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # unsqueeze single image into batch of 1

top5_probabilities, top5_class_indices = torch.topk(output.softmax(dim=1) * 100, k=5)

Feature Map Extraction

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'hf_hub:munehiro-k/cspnext_l.rsb_a1_in1k',
    pretrained=True,
    features_only=True,
)
model = model.eval()

data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # unsqueeze single image into batch of 1

for o in output:
    # print shape of each feature map in output
    # e.g.:
#  torch.Size([1, 64, 112, 112])
#  torch.Size([1, 128, 56, 56])
#  torch.Size([1, 256, 28, 28])
#  torch.Size([1, 512, 14, 14])
#  torch.Size([1, 1024, 7, 7])

    print(o.shape)

Image Embeddings

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'hf_hub:munehiro-k/cspnext_l.rsb_a1_in1k',
    pretrained=True,
    num_classes=0,  # remove classifier nn.Linear
)
model = model.eval()

data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # output is (batch_size, num_features) shaped tensor

# or equivalently (without needing to set num_classes=0)
output = model.forward_features(transforms(img).unsqueeze(0))
# output is unpooled, a (1, 1024, 7, 7) shaped tensor

output = model.forward_head(output, pre_logits=True)
# output is a (1, num_features) shaped tensor

Model Comparison

ImageNet-1k validation accuracy at 224x224 as reported by OpenMMLab (source); not re-evaluated for this conversion. Params and GMACs are measured on the timm model. The evaluation transform of the source is: resize shorter edge to 236 (bicubic), center crop 224 (crop_pct=0.949).

Provenance and License

• Weights were converted to timm format from the OpenMMLab checkpoint listed under Original above (key renaming only, no change to the tensor values). The model definition is a plain PyTorch re-implementation adapted from MMDetection / MMPretrain.
• SHA-256 of the original checkpoint (cspnext-l_8xb256-rsb-a1-600e_in1k-6a760974.pth): 6a7609745272569ba9a693f5c8b452cd355f5676c9e45521209e7deecaba375f
• The OpenMMLab checkpoints were trained on ImageNet-1k with the MMPretrain rsb-a1 configs (configs/rtmdet/classification/cspnext-*_8xb256-rsb-a1-600e_in1k.py; 600 epochs, Lamb optimizer).
• License: the MMDetection repository is released under Apache-2.0 (LICENSE, Copyright 2018-2023 OpenMMLab) and its RTMDet classification README makes no separate license statement for the checkpoints. They are therefore treated as Apache-2.0 here. This has not been confirmed with OpenMMLab: the same question (open-mmlab/mmdetection#11484, opened 2024-02-20) had no reply from the maintainers as of 2026-10-04. The training data, ImageNet-1k, is subject to its own terms of access.

Citation

@misc{lyu2022rtmdet,
      title={RTMDet: An Empirical Study of Designing Real-Time Object Detectors},
      author={Chengqi Lyu and Wenwei Zhang and Haian Huang and Yue Zhou and Yudong Wang and Yanyi Liu and Shilong Zhang and Kai Chen},
      year={2022},
      eprint={2212.07784},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
}