deep-learning-coursera

Coding assignments from Coursera Deep Learning Specialization. 🤖🦾💻

My custom projects can be found at my other repo here: https://github.com/cloudui/pilotML

Course 1: Neural Networks and Deep Learning

An introduction to Perceptrons, Multi-Layer Perceptrons, and Deep Learning with batch gradient descent without optimization methods

Projects:

• Perceptron + Logistic Regression from Scratch using numpy
• Deep Neural Nets from Scratch
• Planar Data classification
• Deep NNs Application

Course 2: Improving Neural Networks

An introduction to optimization methods: He initialization, batch norm, RMSProp, Momentum, Adam, mini-batch GD, SGD, learning decay, dropout. Exploration of hyperparameter selection and tuning.

Projects:

• Introduction to Tensorflow
• Optimization Method Implementation
  • RMSProp
  • Batch Norm
  • Momentum
  • ADAM optimizer
• Regularization Implementation
• Xavier + He Initialization Implmentation
• Gradient Checking, Gradient Decay

Course 3: Structuring Machine Learning Projects

Exploration of how to structure a long-term projects-- organizing metrics, train + validation + test sets, transfer learning, multi-task learning, data mismatch, error analysis, etc.. No programming assignments.

Course 4: Convolutional Neural Networks

Working on CNN architectures: CNNs with convolutions and pooling, inception networks, residual networks, MobileNets.

Projects:

• CNN from "scratch"
• Applying CNNs
• ResNets
• MobileNet Transfer Learning
• YOLO on self-driving data
• Image Segmentation on self-driving data
• Facial Recognition with Siamese Networks + CNNs
• Art Generation with Neural Transfer Learning

Samples

YOLO Bounding Box Prediction for Self-Driving

Sample prediction for Self-Driving Image

Image Segmentation with U-Net

Sample prediction for Image Segmentation. Note the training epochs were limited, so accuracy is limited by training time rather than model

Neural Style Transfer

Nerual Style Transfer example. The image is regenerated by capturing the style of the style photo. The first epoch is the result after 2500 epochs. The one after is after around 200000 epochs.