You can deploy this model and start earning money today!

iamgarvit/
vig-tiny-cifar100-224-2

Name: vig-tiny-cifar100-224-2
Availability: InStock

Views

No views yet

ViG-Tiny on CIFAR-100 @ 224×224

Vision GNN (Graph Neural Network) model trained from scratch on CIFAR-100 upsampled to 224×224. ViG represents images as graphs and uses graph convolutions instead of traditional convolutions or attention mechanisms.

Model Details

Architecture: ViG-Tiny (Vision GNN)
Parameters: ~7.6M
Dataset: CIFAR-100 (100 classes, 32×32→224×224)
Training: From scratch, no pretraining
Key Innovation: Graph-based image representation with k-NN dynamic graphs

Architecture

ViG-Tiny uses:

Channels: 192
Blocks: 12 (Grapher + FFN pairs)
k-NN: 9 neighbors (increases to 18)
Graph Conv: Max-Relative (MR) convolution
Stem: Multi-scale CNN stem (224→14×14)
Head: 1024-dim classifier

Training Setup

CIFAR-friendly hyperparameters with light augmentation and regularization:

Optimizer: AdamW (lr=0.001, wd=0.0005)
Scheduler: Cosine with 5 epoch warmup
Epochs: 100
Batch Size: 128
Augmentation: RandAugment + Light Mixup (0.2)
Regularization: No label smoothing, No drop path
Normalization: CIFAR-100 statistics
Mixed Precision: Enabled

Results

Best Test Acc@1: 76.98%
Best Test Acc@5: 93.66%
Final Test Acc@1: 76.98%
Final Test Acc@5: 93.40%
Training Time: 5.60 hours

Available Checkpoints

best_model.pth - Best performing model (76.98% Acc@1)
final_model.pth - Final model after 100 epochs
checkpoint_epoch_X.pth - Saved every 20 epochs

Usage

python

import torch

# Load model (copy model definition from notebook)
model = vig_tiny(num_classes=100)

# Load trained weights
checkpoint = torch.load('best_model.pth')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Inference
with torch.no_grad():
    output = model(image_tensor)
    probs = torch.softmax(output, dim=1)

Citation

bibtex

@inproceedings{han2022vision,
  title={Vision GNN: An Image is Worth Graph of Nodes},
  author={Han, Kai and Wang, Yunhe and Guo, Jianyuan and Tang, Yehui and Wu, Enhua},
  booktitle={NeurIPS},
  year={2022}
}

Notes

This model uses CIFAR-friendly hyperparameters optimized for the smaller dataset, with reduced augmentation and regularization compared to ImageNet training.

iamgarvit/vig-tiny-cifar100-224-2