ResNet-18 - Trained from Scratch on CIFAR-10
Assignment: DLOps Assignment 5 - Q2
Author: Vidhan Savaliya (M25CSA031, IIT Jodhpur)
Architecture: ResNet-18 (no pre-training)
Dataset: CIFAR-10
Training Details
| Parameter |
Value |
| Epochs |
50 |
| Optimizer |
SGD + Momentum |
| LR Schedule |
Cosine Annealing |
| Batch Size |
128 |
Performance
| Metric |
Value |
| Final Train Accuracy |
99.78% |
| Final Val Accuracy |
92.82% |
| Final Test Accuracy |
93.79% |
| Assignment Target |
>= 72% PASSED |
Adversarial Robustness (FGSM)
| Epsilon |
Test Accuracy |
| 0.00 |
93.79% |
| 0.01 |
75.28% |
| 0.03 |
49.34% |
| 0.10 |
32.10% |
| 0.30 |
21.13% |
WandB
https://wandb.ai/vidhan-savaliya-indian-institute-of-technology-jodhpur/dlops-ass5-q2?nw=nwuservidhansavaliya
Usage
import torch
import torchvision.models as models
model = models.resnet18(pretrained=False, num_classes=10)
state_dict = torch.load('Q2_resnet18_clean.pth', map_location='cpu')
model.load_state_dict(state_dict)
model.eval()
