Model card for nf_seresnet50.anfr_in1k
ANFR-50 (SE): a normalization-free ResNet-50 with Squeeze-and-Excitation channel attention in every block, pre-trained on ImageNet-1k. It is one of the three ANFR backbones of the paper An Architecture Built for Federated Learning: Addressing Data Heterogeneity through Adaptive Normalization-Free Feature Recalibration (TMLR, 2025, arXiv:2410.02006), where the federated experiments start from these weights.
Model details
| Architecture | nf_seresnet50 (scaled weight standardization, no activation normalization, SE attention) |
| Parameters | 28.09 M |
| ImageNet-1k top-1 | 80.4 (Table 10 of the paper) |
| Input | 224 x 224, bicubic interpolation, ImageNet mean and standard deviation |
| Weights | The exponential moving average of the training run, which is what the experiments of the paper loaded |
Usage
import timm
model = timm.create_model("hf_hub:adcw447/nf_seresnet50.anfr_in1k", pretrained=True)
The architecture is part of timm, which until now shipped it without weights.
Training
Trained with timm's training script: RMSProp (TensorFlow variant), learning rate 0.09 at batch size 1,024, step decay, weight decay 7e-6, RandAugment (rand-m8-n2-mstd1), mixup 0.2, drop path 0.1, mixed precision, and an exponential moving average of the weights with decay 0.9999. The full command is in the code repository.
Code
https://github.com/siomvas/ANFR
Citation
@article{siomos2025anfr,
title = {An Architecture Built for Federated Learning: Addressing Data Heterogeneity through Adaptive Normalization-Free Feature Recalibration},
author = {Siomos, Vasilis and Passerat-Palmbach, Jonathan and Tarroni, Giacomo},
journal = {Transactions on Machine Learning Research},
year = {2025},
url = {https://openreview.net/forum?id=GtdYFLsblb}
}