nelscorrea

README

Palm Beach Data Science Meetup (07/18/2019)

https://www.meetup.com/Palm-Beach-Data-Meetup/events/262988444/
Slides and Jupyter notebooks from the Palm Beach Data Science Meetup of Thursday, July 18, 2019.

---

Deep Learning Architectures for Image Classification and Object Detection

Speaker: Nelson Correa, Ph.D.
Twitter: @nelscorrea
https://linkedin.com/in/ncorrea

Description

Object detection is a task in computer vision with many practical applications that can now be achieved with super-human levels of performance on selected benchmarks using deep neural networks. In this talk we define the object detection task and present J. Redmon’s YOLO (You Only Look Once) V3 deep neural network architecture. As preliminaries to object detection and YOLOv3, we first describe image classification on the Pascal VOC and ImageNet benchmark datasets, and introduce a series of deep learning neural network architectures that include the multilayer perceptron (MLP), convolutional neural networks (CNNs), and other networks with dystopian names such as AlexNet, GoogLeNet/Inception, VGG16, ResNet, and Region-CNN (R-CNN). We conclude with note of recent developments, including capsule networks (CapNets) by G. Hinton and deep networks with visual feedback. Slides and notebooks with code will be available after the talk.

---

Slides (HTML)

• PBDS Main Slide deck
• MNIST Digit Classification (MLP and CNN)
• VGG16 Image Classification
• YOLOv3 Object Detection

---

Jupyter Notebooks

• MNIST Digit Classification (MLP and CNN)
• VGG16 Image Classification
• YOLOv3 Object Detection

---

Required files

VGG16_Image_Classification.ipynb

To train the custom output classifier of VGG16_Image_Classification.ipynb, a subset of the Kaggle dogs vs. cats data is needed, and the variable kaggle_dogcats_dir needs to be set to your local directory with the data (e.g., '/your_datasets/dogs-vs-cats/'). Your may download the data from:

• dogs vs. cats data: https://kaggle.com/c/dogs-vs-cats/data

The image data directory should have test, train and validation subdirectories, each with cats and dogs subdirectories.

YOLOv3_Object_Detection.ipynb

To run the YOLOv3_Object_Detection.ipynb notebook you must download the pre-trained yolov3.weights file from J. Redmon’s site to a local directory on your machine, and change the variable weights_path in the second code cell of the notebook to your local directory.

Steps:

• Download yolov3.weights from https://pjreddie.com/media/files/yolov3.weights.
• weights_path = '/Your_Local_Directory_Models/yolov3.weights'

---

This site is open source. Improve this page.