miguelgfierro/ai_projects

AI projects

AI projects

This repo contains AI projects in multiple areas of machine learning. Many of these projects have associated articles on the blog sciblog.

You can find a list of most the post I made in this file.

Featured projects

• Introduction to Convolutional Neural Networks: In this project we explain what is a convolution and how to compute a CNN using MXNet deep learning library with the MNIST character recognition dataset. Here the blog entry.

• Introduction to Transfer Learning: In this project we use PyTorch to explain the basic methodologies of transfer learning (finetuning and freezing) and analyze in which case is better to use each of them. Here the blog entry.

• Cloud-Scale Text Classification With Convolutional Neural Networks: In these notebooks we show how to perform character level convolutions for sentiment analysis using Char-CNN and VDCNN models. Here the blog entry.

• Introduction to Data Generation: In this notebook we show a number of simple techniques to generate new data in images, text and time series. Here the blog entry.

• Introduction to Dimensionality Reduction with t-SNE: In this project we use sklearn and CUDA to show an example of t-SNE algorithm. We use a CNN to generate high-dimensional features from images and then show how they can be projected and visualized into a 2-dimensional space. Here the blog entry.

• Introduction to Distributed Training with DeepSpeed: In this project we show how to use DeepSpeed to perform distributed training with PyTorch. Here the blog entry.

• Introduction to Fraud Detection: In this notebook we design a real-time fraud detection model using LightGBM on GPU (also available on CPU). The model is then operationalized through an API using Flask and websockets. Here the blog entry.

• Introduction to Machine Learning API: In this notebook we show how to create an image classification API. The system works with a pretrained CNN using CNTK deep learning library. The API is setup with Flask for managing the end point services and CherryPy as the backend server. Here the blog entry.

• Introduction to Recommendation Systems with Deep Autoencoders: In this notebook we make an overview to recommendation systems and implement a recommendation API using a deep autoencoder with PyTorch and the Netflix dataset. Here the blog entry.

• Introduction to Natural Language Processing with fastText: In this project we show how to implement text classification, sentiment analysis and word embedding using the library fastText. We also show a way to represent the word embeddings in a reduced space using t-SNE algorithm. Here the blog entry.

• Time Series Forecasting of Stock Price: In this tutorial we show how to implement a simple stock forecasting model using different variants of LSTMs and Keras. Here the blog entry.

• Visualization of Football Matches with Datashader: In this notebook we explain how to visualize all matches in the UEFA Champions League since its beginning using the python library datashader. To create the project we use the Lean Startup method. Here the blog entry.

ai_projects's Issues

[POST] Flask API for object detection in Azure

https://docs.microsoft.com/en-us/azure/app-service-web/web-sites-python-create-deploy-flask-app

https://github.com/Azure/ObjectDetectionUsingCntk

create the image evaluation function with CNTK

Understanding time difference in finetunning vs freeze and train

Posted in pytorch forum: https://discuss.pytorch.org/t/understanding-time-difference-between-finetuning-and-training-the-last-layer-with-frozen-weights/10796

[POST] Transfer learning with pytorch

• Get TL paper from Bengio
• Select datasets
• Code for loading model (ResNets)
• Code for removing the last layer
• Code for funetunning
• Metric for measuring the difference between datasets

Manage grayscale images

we can create fake RGB images from 1 channel images by replicating the channels:

data_transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Lambda(lambda x: torch.cat([x, x, x], 0))
])

source

cherrypy log

http://stackoverflow.com/questions/21191233/how-to-get-custom-logging-working-in-cherrypy

Test map background in football visualization

http://www.naturalearthdata.com/downloads/

Make the map tiles work in datashader

Map tiles are not working with datashader. Fix the issue.

run multigpu pytorch

Create a markdown with the complete process: end point+image evaluation+response

Load model and synset at the same time as api

We need to load the model and the synset when the API starts.

Clean localhost gunicorn method and add comments

Create map of matches using datashader

Take all matches and all locations and create a dataframe and then plot it in the map

See if it is needed an image transformation

def get_preprocessed_image(my_image, mean_image):
''' Reshape and flip RGB order '''
my_image = my_image.astype(np.float32)
bgr_image = my_image[:, :, ::-1] # RGB -> BGR
image_data = np.ascontiguousarray(np.transpose(bgr_image, (2, 0, 1)))
image_data -= mean_image
return(image_data)

Use Stamen toner maps to create a black and white map of Europe

Use the Stamen toner maps http://tile.stamen.com/toner-background/{Z}/{X}/{Y}.png, and use the alpha parameter to change the brightness level to make it gray and black.

Reading characters from card

I would like to read characters from scratch cards ... I still don't really understand how/where you connected the trained dataset result to CNN.

change function create_image to accept parameters

Should accept parameters like the color of the map

Gunicorn with Nginx

Set up gunicorn initially with Nginx:
https://www.digitalocean.com/community/tutorials/how-to-serve-flask-applications-with-gunicorn-and-nginx-on-ubuntu-14-04

Set up gunicorn with the API

https://www.digitalocean.com/community/tutorials/how-to-serve-flask-applications-with-gunicorn-and-nginx-on-ubuntu-14-04

Explore kubernetes to schedule GPUs

https://en.m.wikipedia.org/wiki/Kubernetes

simpsons dataset

Change comments in football notebook to reflect the lean startup method with TDSP

Comments the complete story and link tghe post

Improve route('/') with a picture of Bender

https://github.com/miguelgfierro/sciblog_support/blob/api_detection/Intro_to_Machine_Learning_API/api.py#L61

CNTK API notebook fixes

fix typos
new bombardment: https://github.com/ilkarman/Azure-WebApp-w-CNTK/blob/master/readme_support/REST_Speed_Test.ipynb

improve plot_metrics function

the legend is two points but it would be nicer if it were two points connected with a line

Test the CNTK API with nginx and gunicorn

NLP project: wrap up content and merge

• Explain fastText algorithm
• Why is important in comparison with deep learning
• Execute all from scratch
• Review
• Merge with master

Install a Kubernetes cluster & drivers

1. Install Kubernetes cluster
   https://github.com/ritazh/acs-engine/blob/enable-k8v1.6-multiplegpu/docs/kubernetes.md -> shows how to create a Kubernetes cluster

NOTE: Make sure to configure the agent nodes with vm size Standard_NC12 or above to utilize the GPUs

2. Install drivers:

• SSH into each node and run the following scripts :
  install-nvidia-driver.sh

curl -L -sf https://raw.githubusercontent.com/ritazh/acs-k8s-gpu/master/install-nvidia-driver.sh | sudo sh

To verify, when you run kubectl describe node <node-name>, you should get something like the following:

Capacity:
alpha.kubernetes.io/nvidia-gpu:    2
cpu:                               12
memory:                            115505744Ki
pods:                              110

Plot lines with datashader independently

I would like to plot lines with datashader so I can control the initial and final points.

end point to classify images from url

Test cherrypy

basic config: http://flask.pocoo.org/snippets/24/

advanced config: http://fgimian.github.io/blog/2012/12/08/setting-up-a-rock-solid-python-development-web-server/

performance comparison cherrypy vs gunicorn vs bjoern: https://blog.appdynamics.com/engineering/a-performance-analysis-of-python-wsgi-servers-part-2/

Visualization of word vector with t-sne

t-sne algorithm in sklearn: http://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html
t-sne in JS: http://karpathy.github.io/2014/07/02/visualizing-top-tweeps-with-t-sne-in-Javascript/

Test if apache is easier to manage

https://www.vioan.eu/blog/2016/10/10/deploy-your-flask-python-app-on-ubuntu-with-apache-gunicorn-and-systemd/

repo: https://github.com/vioan/flask-apache-gunicorn-systemd

other option: https://gist.github.com/vehrka/7799b263a32124e7b10e

[POST] Post about fastText in NLP

[POST] Post about NIPS

New post about the highlights at NIPS

Analyze Twitter data

Get coordinates from European stadiums

I found this one that contains teams from Spain, Germany, England, Scotland and France. I'm looking for the rest of the teams in Europe and Russia. Best case scenario would be from teams all over the world.

Set up a question in SO.

Setup a WiFi repeater

http://lifehacker.com/5563196/turn-your-old-router-into-a-range-boosting-wi-fi-repeater

Word representation

Entry about word representation https://github.com/salestock/fastText.py#word-representation-learning

Text classification with DBpedia

Create new stadium dataset with other teams in Europe

Right now stadium.csv has teams from England, Scotland, Spain, France and Germany. I need to create a new one with teams from the rest countries that participated in the Champions League.

Do you have demo url for this project :) ?

Update final image in blog lean startup

Update final image and put it the black background

Set up fast text for python

Install fast text https://github.com/facebookresearch/fastText and python wrapper https://github.com/salestock/fastText.py

Sentiment analysis with Amazon polarity

Find a tool to benchmark the API performance

with python: https://github.com/ilkarman/CV_end_to_end/blob/master/00-Resnet/Server%2BTests.ipynb

with nodejs: http://www.abhishekshukla.com/javascript/benchmark-your-rest-api-the-easy-way/

Quantify difference between datasets

2 dimensions as stated in this transfer learning tutorial: size and similarity to the original dataset.
A) The size can be in number of examples and number of examples per class, here maybe we can do a weighted quadratic difference.
B) similarity, there is literature on image similarity (maybe KL for colour and texture)

Basic API with flask

Post and get service

execute api from notebook

%%bash --bg
/home/hoaphumanoid/anaconda3/envs/cntk-py35/bin/python /home/hoaphumanoid/installer/sciblog_support/Intro_to_Machine_Learning_API/cntk_api.py

Reconsider datashader project: from visualization to data science process

Focus on how to develop a DS process following the lean startup approach