DeKUT-DSAIL/convnextv2-huge-cifar10-native

convnextv2-huge-native-exp1

ConvNeXt V2 Huge fine-tuned on CIFAR-10 at native 32x32 resolution (adapted stem).

Fine-tuned from ImageNet-pretrained weights (timm name convnextv2_huge) on CIFAR-10 as part of a native-resolution vs upsampling study.

Evaluation results (CIFAR-10 test set, 10,000 images)

Metric	Value
Top-1 accuracy	99.18%
Top-5 accuracy	99.91%
F1 (macro)	0.9918
AUC (macro)	0.9991
ECE	0.1082

• Parameters: 657.50M
• Input: native 32x32 CIFAR images (adapted 3x3 stride-1 stem)
• Preprocessing: resize to 32x32, then normalize with CIFAR-10 mean [0.4914, 0.4822, 0.4465] and std [0.247, 0.2435, 0.2616].

Usage

from cifar_classifier import CIFAR10Classifier

clf = CIFAR10Classifier.from_pretrained("DeKUT-DSAIL/convnextv2-huge-cifar10-upsample")
label, probs = clf.predict("image.jpg")
print(label, probs)

Get cifar_classifier.py from the hf_deployment/ folder of the training repository. It only needs torch, torchvision, timm, Pillow and huggingface_hub.

Classes

airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck