Intro to Neural Networks

Simple implementations of basic neural networks in both Keras and PyTorch

College Admissions
2-Layer fully connected neural network used to solve binary classification task. Aim is to predict if student is admitted or not based no theirs GPC & GPA scores

Code: Keras PyTorch
MNIST Handwritten Digits
2-Layer fully connected neural network used to solve popular MNIST dataset. MNIST is set of 60k images. Each one is 28x28 grayscale

Code: Keras PyTorch
IMDB Movie Reviews
2-layer fully connected neural network used to solve IMDB movie reviews dataset. Uses bag-of-words approach.

Code: Keras PyTorch

NumPy Implementations

Quick implementation of basic neural network building blocks in pure NumPy

Perceptron - Classification
Single-Layer Perceptron used to solve binary classification task. Trained on college admission dataset.

Code: NumPy
Feedforward - Regression
Feedforward Multi-Layer Perceptron used to solve regression task. Trained on bike-rental dataset.

Code: NumPy
Feedforward - Classification
Feedforward Multi-Layer Perceptron used to solve image classification task. Trained on Fashion-MNIST dataset.

Code: NumPy
RNN - Counter
Recurrent Neural Network used as a counter. This RNN has many-to-one arrangement. Dataset is synthetic. Both unfolded and recurrent implementations.
Code not tidied.

Code: NumPy
RNN - Text Generation
Character-level Recurrent Neural Network used to generate novel text. This RNN has many-to-many arrangement. Dataset is composed of 300 dinosaur names.
Code not tidied.

Code: NumPy

Vision Basics

Convolutional Network
Small convolutional neural network used for image classification on the CIFAR-10 dataset. Images are RGB 32x32 pixels.

Code: Keras
Batch Normalisation
Small convnet with batch normalisation used on CIFAR-10 dataset.

Code: Keras
Data Augmentation
Small convnet with data augmentation to reduce overfitting on CIFAR-10 dataset.

Code: Keras
Custom Generator
Small convnet with custom data generator trained on CIFAR-10 dataset.

Code: Keras
ResNet from Scratch
ResNet-50 implemented from scratch using Keras functional API. Then trained from scratch on Oxford VGG Flowers 17 dataset.

Code: Keras
Bottleneck Features
Pre-trained DenseNet-201 is further trained on Oxford VGG Flowers 102 dataset. Bottleneck features are extracted and only classifier is trained.

Code: PyTorch

Mini Projects

Image Classification
Keras: ResNet-50 trained on Oxford VGG Flower 17 dataset.
PyTorch: DenseNet-201 trained on Oxford VGG Flower 102 dataset.

Code: Keras PyTorch
Multi-Digit Detection
ConvNet trained to detect variable length house numbers on Street View House Numbers dataset.

Code: PyTorch
Keypoint Detection
Pre-trained MobileNetV2 is used for keypoint detection on CATs dataset.

Code: Keras PyTorch
Object Detection
YOLOv2 on DarkNet backbone trained on BCCD dataset.

Code: Keras
Image Segmentation
Segnet on VGG16 backbone trained on Pascal VOC2007 dataset.
Code not tidied.

Code: PyTorch

NLP Basics

Time Series Prediction
Vanilla RNN trained to perform time series prediction on sine wave.

Code: PyTorch
Text Generation
Char-level LSTM trained to generate new text trained on Shakespeare plays.

Code: PyTorch
Word2Vec Skipgram
Example how to train embedding layer using Word2Vec Skipgram. Trained on popular wiki8 dataset

Code: PyTorch

Mini Projects

Sentiment Analysis
Word-level LSTM trained to generate predict if movie review is positive or negative. Uses IMDB Movie Reviews dataset

Code: PyTorch
Language Translation
Two models are trained to perform English to French translation.
• Many-to-Many (m2m)
• Sequence-to-Sequence (s2s)

Code: Keras (m2m) Keras (s2s)

Implementaitons of simple GAN architectures

Autoencoder
Simple autoencoder trained on MNIST dataset.

Code: Keras
GAN
Fully connected Generative Adversarial Network trained on MNIST dataset.

Code: Keras
DCGAN
Deep Convolutional GAN trained on CelebA dataset.

Code: PyTorch