Model card for lowformer_b1.in1k

A LowFormer image classification model. LowFormer targets real measured latency rather than MAC count, combining fused and grouped MBConv stages with an efficient attention block that projects to a lower spatial resolution with a strided depthwise convolution and upsamples the result with a transposed convolution. The B variants are the original models from the WACV paper. This checkpoint was trained on ImageNet-1k by the paper authors and converted to the timm state-dict layout.

Model Notes

• The default preprocessing is encoded in the pretrained configuration: bicubic resize, ImageNet mean and standard deviation, and a center crop with crop_pct=0.95.
• The model can be used for ImageNet-1k classification, image embeddings, or multi-scale feature-map extraction with features_only=True.
• The accuracies above are measured in FP32. lowformer_b0 and lowformer_b1 are the variants most affected by bfloat16 autocast, losing roughly 3.0 and 1.5 Top-1 respectively; the other variants stay within 0.25 of their FP32 result. FP16 autocast matches FP32 for all variants.
• Latency depends on the target hardware, inference runtime, export path, and batch size. See the papers for device-specific measurements and methodology.

Model Details

• Model Type: Image Classification / Feature Encoder
• Model Stats:
  • Params (M): 17.9
  • GMACs: 1.4
  • Activations (M): 3.7
  • Image size: 224 x 224
• Original: https://github.com/altair199797/LowFormer
• License: Apache 2.0
• Dataset: ImageNet-1k
• Papers:
  • LowFormer: Hardware Efficient Design for Convolutional Transformer Backbones: https://arxiv.org/abs/2409.03460
  • PyTorch Image Models: https://github.com/huggingface/pytorch-image-models

Model Usage

Image Classification

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('lowformer_b1.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(
    'lowformer_b1.in1k',
    pretrained=True,
    features_only=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

for o in output:
    # print shape of each feature map in output
    # e.g.:
    #  torch.Size([1, 32, 56, 56])
    #  torch.Size([1, 64, 28, 28])
    #  torch.Size([1, 128, 14, 14])
    #  torch.Size([1, 256, 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(
    'lowformer_b1.in1k',
    pretrained=True,
    num_classes=0,  # remove classifier nn.Linear
)
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))  # 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, 256, 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 in FP32 with bicubic interpolation and a centered crop (crop_pct=0.95). Values are Top-1 / Top-5 percentages; only input resolution changes between columns.

Model	Params (M)	224 Top-1 / Top-5	256 Top-1 / Top-5	288 Top-1 / Top-5
lowformer_b0.in1k	14.10	78.388 / 94.026	79.194 / 94.462	79.306 / 94.444
lowformer_b1.in1k	17.94	79.806 / 94.592	80.260 / 94.914	80.406 / 95.072
lowformer_b15.in1k	33.98	81.102 / 95.258	81.558 / 95.470	81.708 / 95.588
lowformer_b3.in1k	57.09	83.656 / 96.656	83.988 / 96.738	84.066 / 96.834
lowformer_e1.in1k	18.90	78.772 / 94.120	79.366 / 94.450	79.624 / 94.562
lowformer_e2.in1k	22.75	81.612 / 95.714	81.982 / 95.948	82.156 / 96.098
lowformer_e3.in1k	41.32	83.044 / 96.344	83.166 / 96.536	83.402 / 96.552

Citation

@inproceedings{nottebaum2025lowformer,
  title={LowFormer: Hardware Efficient Design for Convolutional Transformer Backbones},
  author={Nottebaum, Moritz and Dunnhofer, Matteo and Micheloni, Christian},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages={7008--7018},
  year={2025}
}

@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}}
}