ResNet-18 on CIFAR-10 @ 224×224 - Vision GNN Baseline

This model is ResNet-18 trained from scratch on CIFAR-10 (upsampled to 224×224) as a baseline for Vision GNN (ViG) reproduction study.

Model Description

Architecture: ResNet-18
Parameters: ~11.7M
Dataset: CIFAR-10 (32×32 upsampled to 224×224)
Training: From scratch (no pretraining)
Purpose: Baseline for validating ViG paper claims

Training Details

Optimizer: SGD (lr=0.1, momentum=0.9, weight_decay=1e-4)
Scheduler: MultiStepLR (milestones=[60, 120, 160], gamma=0.2)
Epochs: 150
Batch Size: 128
Normalization: CIFAR-10 statistics (not ImageNet)

Results

Best Test Accuracy: 93.09%
Final Test Accuracy: 92.92%
Training Time: 3.74 hours

Methodology

We follow the ImageNet training protocol adapted for CIFAR-10 to ensure fair comparison with ViG. All models in the comparison are trained under identical conditions:

Same resolution (224×224)
Same training recipe
Same data augmentation
No pretrained weights

Available Checkpoints

best_model.pth - Best performing checkpoint (93.09% accuracy)
final_model.pth - Final model after all epochs
checkpoint_epoch_X.pth - Saved every 50 epochs

Usage

import torch
from torchvision import models

# Load model
model = models.resnet18(weights=None)
model.fc = torch.nn.Linear(model.fc.in_features, 10)

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

Citation

Training protocol adapted from:

ResNet paper (He et al., 2015)
Vision GNN paper (Han et al., 2022)

Downloads last month: -; Downloads are not tracked for this model. How to track