Paper Replications

A collection of deep learning paper implementations in PyTorch, trained and evaluated on standard benchmarks.

---

Implementations

#	Paper	Dataset	Directory
1	AlexNet — ImageNet Classification with Deep Convolutional Neural Networks (Krizhevsky et al., 2012)	CIFAR-10	AlexNet/
2	Deep Residual Learning for Image Recognition (He et al., 2015)	Imagenette	ResNet/

---

AlexNet

PyTorch implementation of AlexNet adapted for CIFAR-10.

Prerequisites

• Python 3.8+
• PyTorch
• TorchVision

cd AlexNet
pip install -r requirements.txt

Usage

python train.py

Command Line Arguments

Argument	Description	Default
--data-dir	Directory to store CIFAR-10 data	./data
--epochs	Number of training epochs	5
--batch-size	Batch size for training	64
--lr	Learning rate	0.01
--momentum	Momentum for SGD	0.9
--weight-decay	Weight decay for SGD	0.0005
--save-path	Path to save trained model weights	alexnet.pth

Example:

python train.py --epochs 10 --batch-size 128 --lr 0.005

Files

File	Description
train.py	AlexNet model definition and training loop
alexnet.ipynb	Jupyter notebook for interactive exploration
requirements.txt	Python dependencies
NIPS-2012-imagenet-classification-with-deep-convolutional-neural-networks-Paper.pdf	Original paper

---

ResNet

A beginner-friendly ResNet-34 implementation (BasicBlock only) trained on Imagenette — a 10-class subset of real ImageNet images at 224×224. Same preprocessing pipeline as ImageNet, but auto-downloads and runs in reasonable time on a free GPU.

Prerequisites

• Python 3.8+
• PyTorch
• TorchVision

pip install torch torchvision

Recommended: Google Colab

Training ResNet-34 at 224×224 needs a GPU. Colab is the easiest way to run this — free GPU, no local setup.

1. Open ResNet/run_on_colab.ipynb in Google Colab
2. Upload resnet.py and train_resnet.py to the notebook session (or clone this repo in Colab)
3. Set runtime to GPU (Runtime → Change runtime type)
4. Run all cells — Imagenette downloads automatically

Local (if you have a GPU)

cd ResNet
python train_resnet.py

Edit the config at the top of train_resnet.py to change epochs, batch size, learning rate, etc. Lower BATCH_SIZE (e.g. 16) if you run out of VRAM on a small GPU. CPU-only training works but will be very slow.

Config defaults

Setting	Value
Architecture	ResNet-34 ([3, 4, 6, 3] layers)
Dataset	Imagenette (10 classes, 224×224)
Epochs	15
Batch size	32
Optimizer	Adam, lr 1e-3

Files

File	Description
resnet.py	ResNet-34 model (BasicBlock + ResNet class)
train_resnet.py	Training and validation loop
run_on_colab.ipynb	Colab notebook — recommended way to run

Full ImageNet?

The model uses an ImageNet-style stem (7×7 conv, 224×224), so it can train on full ImageNet in principle — but that means ~150GB of data, manual download from image-net.org, and days of GPU time. Imagenette is the practical stand-in for learning and testing this code.

---

Contributing

Each paper lives in its own subdirectory with a self-contained train.py, requirements.txt, and the original PDF.