resnet18-imagenet100-from-scratch

ResNet-18 — ImageNet-100 Pretraining from Scratch

A ResNet-18 implementation built in PyTorch and trained from scratch on ImageNet-100.

This project was created as part of my effort to understand computer vision architectures from first principles rather than relying on pretrained torchvision models.

The model is intended both for ImageNet-100 classification and as a pretrained visual backbone for transfer-learning experiments.

Model Architecture

• Architecture: ResNet-18
• Parameters: 11,227,812
• Residual blocks: [2, 2, 2, 2]
• Input: 3 × 224 × 224
• Stem: 7 × 7 convolution, stride 2
• MaxPool: 3 × 3, stride 2
• Final feature representation: 512 dimensions
• Classification head: 512 → 100

The architecture was implemented manually in PyTorch.

Training

The model was trained for 100 epochs using:

• Loss: CrossEntropyLoss
• Optimizer: SGD
• Initial learning rate: 0.1
• Momentum: 0.9
• Weight decay: 1e-4
• Batch size: 128
• Scheduler: StepLR
• Step size: 30 epochs
• Gamma: 0.1

Data Augmentation

Training preprocessing:

v2.Compose([
    v2.RandomResizedCrop(224),
    v2.RandomHorizontalFlip(),
    v2.ToImage(),
    v2.ToDtype(scale=True, dtype=torch.float32),
    v2.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

Validation preprocessing:

v2.Compose([
    v2.Resize(224),
    v2.CenterCrop(224),
    v2.ToImage(),
    v2.ToDtype(dtype=torch.float32, scale=True),
    v2.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

Performance

Best ImageNet-100 validation accuracy:

78.72%

This was reached at epoch 81.

The uploaded checkpoint corresponds to the best validation model.

Loading the Model

import torch
from model import ResNet18

model = ResNet18(num_classes=100)

state_dict = torch.load(
    "model.pth",
    map_location="cpu",
    weights_only=True
)

model.load_state_dict(state_dict)
model.eval()

Feature Extraction

The classification layer can be removed to use the network as a 512-dimensional visual feature extractor:

model.fc = torch.nn.Identity()

An input batch with shape:

[B, 3, 224, 224]

will then produce:

[B, 512]

representations.

Motivation

The purpose of this project is not only image classification.

The model was pretrained to investigate transfer learning: whether visual representations learned from natural images can transfer to substantially different visual domains.

The next stage of the project explores transfer learning from ImageNet-100 to knee MRI data for the RSNA Knee Abnormality Detection task.

Code

The complete training notebook and broader computer vision learning project are available on GitHub:

https://github.com/sirohi-ml/Image_Detection_from_scratch

Limitations

This model was trained on ImageNet-100 rather than the complete ImageNet-1K dataset.

It should therefore not be treated as equivalent to standard ImageNet-1K pretrained ResNet-18 weights.

Performance on downstream domains has not yet been established and should be evaluated for each transfer-learning task.