Saturday, March 25, 2017

Difference between Batch Gradient Descent and Stochastic Gradient Descent - Udacity Machine Learning Nanodegree Coursera

Recommend this great 13 minutes crystal clear video by Andrew Ng on Coursera explaining the differences between batch gradient descent aka gradient descent aka normal gradient descent versus Stochastic Gradient Descent. It's clear simple and easy to understand without prerequisite. Andrew Ng shows you how the formula differs, how the step by step train strategy differs and a visualization of the trajectory to find global minimum (the center of all the ellipse in his graph).

  • Summary
    • Gradient Descent may have issues when the scale of the data is large
      • If the number of training samples is large
      • Gradient Descent algorithm requires summing over all of m
        • e.g. US population census population of 300MM
    • Stochastic Gradient Descent is a modification of gradient descent
      • In other words, the cost functions are different
    • Stochastic every iteration is faster
    • Steps: randomly shuffle dataset, optimize one training data at a time, improve parameters early one at a time, instead of looking at the examples together as a batch
    • Weakness: generally moves towards global minimum, but doesn't always go there, can reach the general vicinity of the global minimum. Does not converge as nicely as gradient descent. 
    • In reality, practical data science, once it gets close to the global minimum its parameters are good enough. In real life, it works out.

K Means Clustering Unsupervised Learning - Udacity Machine Learning Nanodegree Flash Card

  • Draw a line connecting two centroids and use the half way line as a division line for two hyperplanes (if two clusters). Results vary greatly.

  • Initial positions of centroid can strongly influence result. Different initial positions give completely different results.

  • Analogy "Rubber Band"

  • Center of the cluster is called a centroid

  • Number of centroids at initiation can heavily influence the result. 

  • Great for ... PROS:

  • Bad for ... CONS ... limitations:
    • Hill climbing algorithm.
    • Result depends on initiation
    • If initiation is close to local optima, may be sticky. Never move away. Ignore global optima. Bad initial centroids exist
    • If there are more potential clusters, there are more local optima. Run iterate the algorithm many times to avoid being stuck. 

Thursday, March 23, 2017

Follow my new website - Zero Budget Growth Hacking for Small Businesses

Dear entrepreneurs, small business owners and startup techies, how do you go from zero to one with no marketing budget? I will show you how in my new blog. Here's my background highlighted in the first post

What makes me a special growth hacker? I don't just advertise, I code, hacked and actually took multiple stores, youtube channel, and contents from zero to one.

My Biography
TL;DR Dilys is a social media growth hacker. Dilys' background is the intercept of business, technology, and startup. She has experience working with giant corporations and top YCombinator startups. She contributed to USATODAY, Fast Company, VentureBeat, Crunchies by TechCrunch and was invited to Google social media studies, tech conferences. She ran campaigns to kickstart e-Commerce stores: Chinese Alibaba Taobao 0 to Level 6, eBay 0 to PowerSeller, Shopify 0 to Shopify & Uber partner. She recently took an experimental Youtube partner channel from 0 to 400,000 minutes watched, 0 to 300,000 views, 0 to 900 subscribers in just one month (February 2017 the shortest month too!).

Can't wait to share all my unique experiences as a seller, growth hacker, startup growth person with you. FREE. Just content and some Google ads. That's it. No subscription needed. Follow my blog now.

Wednesday, March 22, 2017

Udacity Digital Marketing Nanodegree Reviews (updating in progress)

This review is updated continuously throughout the program. Yay I just joined the Udacity Nanodegree for Digital Marketing! I am such an Udacity and learning junkie LOL. What grabbed my attention was the line-up of partners, the real world projects and also Avinash Kaushik's presence. I wonder what's the oracle of Google Analytics doing promoting this course.

  • First impression, clean beautiful videos, unlike some of the programming Georgia Tech videos Udacity has
  • The partners really do show up early in the syllabus and seems like they will participate
  • Though jobs are not guaranteed, there are mentions of hiring partners
  • Classmates are young and energetic marketing veterans. Already very active on slack
  • Meet the students use hashtag #ImInDMND on instagram
  • Realworld like non-trivial business cases and owner / user statements
  • What are the projects like? Udacity allows you to use Udacity as a real-world marketing project.
  • Amazing speakers, famous authors and speakers including author of crossing the chasm, avinash kaushik Google Analytics evangelist
  • Mentorship - mentorship is available. My mentor has been unresponsive and unhelpful so far. I do not recommend.
  • Mini interviews with industry giants
  • The Facebook Ad project is extremely useful. The real world project experience can have tangible results. It is resume worthy. I got a sizable view and conversion of which I am comfortable to talk about in future interviews.
  • Does it directly translate into full time jobs? Probably not, Udacity took away the guarantee a long time ago. It does give exposure to Moz, Google Ads, which are not as accessible experience as Facebook ads and mailchimp, which any one can try on one's own. 

Subscribe to our mailing list

Tuesday, March 21, 2017

Udacity Machine Learning Nanodegree - Projects Step by Step Walkthrough High Level Cheat Sheet

High level steps to solve Udacity Machine Learning Nanodegree projects:

  • Import dependencies: numpy, pandas, sklearn, matplotlib
  • Data cleaning:
    • Replace all data with numeric value such as binaries 0 and 1 or scale down to between -1 to 1, or 0 to 1 (normalization). 
    • Replace yes/no binary answers with 1,0
    • Replace categorical data A, B, C with dummy columns |A|B|C| use 1 if true, 0 if false
  • Split data into features and target aka label
  • Perform initial exploration, turns data CSV into Pandas.DataFrame
    • Computer summary stats: mean, counts etc.
  • from sklearn import model
  • clf = sklearnmodel.model() #specify the classifier
  • ... ) #fit the model wither parameters
  • clf.predict() #make predictions
  • Metrics:
    • R^2 R squared - great for linear regression 0 to 1, 1 being the best
  • Errors:
  • This list is under construction

Sklearn machine learning model cheat sheet
What are the best algorithms to use for each machine learning problem?
Classification versus regression
Supervised versus unsupervised

Saturday, March 18, 2017

Commonly seen python error messages - Learn to code Python for Beginners

  • Python KeyError if dict[key]: cannot do this have to change to if key in dict: 

Pandas Sample Code - Udacity Machine Learning

  • .groupby()
  • .count()
  • pandas.DataFrame.count
  • .sum()
  • df[df["class"]==1].count()["value"]
  • countOfColumn = myDataFrame[conditionColumn["myCondition"]=="myCondValue"].count()["conditionColumn"] get row count by column condition and value
  • pandas.DataFrame.count
  • df[(df['A']>0) & (df['B']>0) & (df['C']>0)]
  • pandas.DataFrame.sum
  • df.groupby('a').count()
  • df.first()

Sunday, March 12, 2017

Udacity Machine Learning Nanodegree Bayes Rule Bayesian Analysis Walkthrough

<xi, di>
di = f(xi) + err
x, d, h(x) = x mod 9, h(x) = x/3, h(x) = 2,
1, 1, 1%9 = 1, 1/3, 2,
3, 0, 3%9 = 3,  1, 2,
6, 5, 6%9 = 6, 2, 2,
10, 2, 10%9= 1, 10/3, 2,
11, 1, 11%9= 2, 11/3, 2,
13, 4, 13%9 = 4, 13/3, 2,

sum of squared errors for each (excel calc)

h(x) = x mod 9
sum of squared errors = 12

h(x) = x/3
sum of squared errors = 19.44

h(x) = 2
sum of squared errors = 19

Use the smallest
Or better way: write a python script

Saturday, March 11, 2017

R Squared Coefficient of Determination - Machine Learning Concept

*coefficient of determination*](


coefficient of determination
useful statistics for regression analysis
measures how good the model makes prediction.

R^2 range {0, 1}
can be negative, arbitrarily worse
percentage of square correlection between predicted and actual values of target variable

indicates what percentage of the target variable, using this model, can be explained by the **features**.

r2_score from sklearn.metrics

Wednesday, March 8, 2017

Pandas Numpy Data Analysis Tool Kit - Udacity Machine Learning Nanodegree 01

Numpy perfect for statistical analysis, matrix manipulation. Learn to Code Notes.
Numpy Documentation

Code pattern 01 numpy use array().T to get matrix transpose
X = [1,2,3]
XT = array(X).T, series2)

Pandas Numpy Data Analysis Tool Kit - Udacity Machine Learning Nanodegree 00


Basic units data structures of Pandas, data analysis using Python

Allows users to store a large amount of information and perform data analysis

Dataframe documentation:

A dictionary
  • Dict of 1D ndarrays, lists, dicts, or Series
  • 2-D numpy.ndarray
  • Structured or record ndarray
  • Series
  • Another DataFrame

Sample Code: 

d = {'key_name':Series([1,2,3], index=['a','b','c'])}

Analogy : Excel Spreadsheet
Will also return number of rows and columns



More sample code:
   my_data = pd.DataFrame(data)
    print my_data.dtypes
    print ""
    print my_data.describe()
    print ""
    print my_data.head()
    print ""

    print my_data.tail()

# Retrieve columns
# Retrieve rows

df[df['col_name'] >= 30]

get row column counts of Pandas Dataframe
.count() count each column of the entire table

React UI, UI UX, Reactstrap React Bootstrap

React UI MATERIAL  Install yarn add @material-ui/icons Reactstrap FORMS. Controlled Forms. Uncontrolled Forms.  Columns, grid