Model Card for Yurim0507/resnet18-fashionmnist-unlearning

This repository contains ResNet-18 models retrained on the FashionMNIST dataset with specific classes excluded during training. Each model is trained to study the impact of excluding a class on model performance and generalization on the remaining classes.

Evaluation

• Testing Data: FashionMNIST test set (10,000 images, 1,000 per class)
• Metrics: Top-1 accuracy
  • For the original model: accuracy over all 10 classes.
  • For excluded-class models: accuracy computed on the 9 retained classes’ test samples (9,000 images).

Results

Model File	Excluded Class	FashionMNIST Accuracy
resnet18_fashionmnist_original.pth	None	95.39%
resnet18_fashionmnist_forget0.pth	T-shirt/top	96.54%
resnet18_fashionmnist_forget1.pth	Trouser	95.07%
resnet18_fashionmnist_forget2.pth	Pullover	95.99%
resnet18_fashionmnist_forget3.pth	Dress	95.69%
resnet18_fashionmnist_forget4.pth	Coat	95.92%
resnet18_fashionmnist_forget5.pth	Sandal	94.88%
resnet18_fashionmnist_forget6.pth	Shirt	97.76%
resnet18_fashionmnist_forget7.pth	Sneaker	95.37%
resnet18_fashionmnist_forget8.pth	Bag	94.76%
resnet18_fashionmnist_forget9.pth	Ankle boot	95.33%

Training Details

Training Procedure

• Base Model: ResNet-18
• Dataset: FashionMNIST (28×28 grayscale)
• Excluded Class: varies per model (one of: T-shirt/top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot)
• Loss Function: CrossEntropyLoss
• Optimizer: SGD
  • Learning rate: 0.1
  • Momentum: 0.9
  • Weight decay: 5e-4
  • Nesterov: True
• Scheduler: CosineAnnealingLR (T_max=200)
• Epochs: 200
• Batch Size: 128
• Augmentation: RandomCrop(28, padding=4), RandomHorizontalFlip
• Gradient Clipping: Norm 0.5
• Input Processing:
  • ToTensor() and Normalize with FashionMNIST mean/std (e.g., mean=0.286, std=0.353).
  • If preferring 3-channel input, replicate grayscale to 3 channels and use ImageNet normalization (less common).
• Hardware: Single GPU (NVIDIA GeForce RTX 3090)

Data Preprocessing

• Base Transform (train & test):
  • transforms.ToTensor()
  • transforms.Normalize(mean=(0.286,), std=(0.353,))
• Training Augmentation:
  • transforms.RandomCrop(28, padding=4)
  • transforms.RandomHorizontalFlip(p=0.5)
• Excluded-Class Handling:
  • Remove all training samples of the excluded class from the training split.
  • Evaluate retained-class performance on the remaining 9 classes’ test samples.
  • (Optional) Evaluate excluded-class accuracy separately to confirm near-zero performance.

Related Work

This model is part of the research conducted using the Machine Unlearning Comparator. The tool was developed to compare various machine unlearning methods and their effects on models.

Uses

Direct Use

These models can be directly used for evaluating the effect of excluding specific classes from the CIFAR-10 dataset during training.

Out-of-Scope Use

The models are not suitable for tasks requiring general-purpose image classification beyond the CIFAR-10 dataset.

Model Definition Example

import torch
import torch.nn as nn
from torchvision.models import resnet18

def ResNet18_FashionMNIST(num_classes=10, in_channels=1):
    model = resnet18(weights=None)
    # Modify first conv for grayscale input:
    model.conv1 = nn.Conv2d(in_channels, 64, kernel_size=3, stride=1, padding=1, bias=False)
    model.maxpool = nn.Identity()
    model.fc = nn.Linear(model.fc.in_features, num_classes)
    return model