Tag: Visualization

Configuring KNIME to work with Python 2.7.x on Windows

Apparently it is tricky to get Python integration working in the KNIME Analytics Platform. If you read the official guide too quickly you can miss some critical information at the bottom of the page. I was getting an error complaining that the google.protobuf library was missing even though I thought that I had everything installed correctly:

Library google.protobuf is missing, required minimum version is 2.5.0

Here is the correct sequence of actions to fix this problem and get Python integration working in KNIME. It worked for me, so I hope it works for you too!

  1. If you don’t already have it on your system, download and install Anaconda Python. I use the Python 2.7 version of Anaconda. I recommend installing Anaconda in the folder “C:\Anaconda2”.
  2. Create a new conda environment by using the following code at the windows command prompt (or power shell):
    conda create -y -n py27_knime python=2.7 pandas jedi protobuf
  3. Create a new Windows batch file with the following contents:
    @REM Adapt the directory in the PATH to your system
    @SET PATH=C:\Anaconda2\Scripts;%PATH%
    @CALL activate py27_knime || ECHO Activating py27_knime failed
    @python %*

    Of course, you might have to change the path to your Anaconda2 installation if you installed it to a different path. Save the batch file somewhere, like your user directory or the KNIME workspace folder. I named my file “py27.bat” and placed it in the knime-workspace folder.
    Just in case anyone reading this is confused…. a windows batch file is just a text file that is saved with the file extension “.bat”. You can create it in any text editor by creating a new empty text file, pasting the above four lines of text into it and saving the file as “py27.bat”.

  4. If you haven’t already, download and install KNIME Analytics Platform. At the time of writing, the latest version is 3.4.0, so that’s the one I used. You might as well get the installer with all the free extensions included, presuming you have enough disk space for it and your internet connection is decent enough. This will save having to install further packages later, although to be fair KNIME seems to do a fairly good job at installing packages on the fly (after asking you first) whenever you load a workspace that needs them.

    Downloading KNIME

  5. Start KNIME. Go to File>Preferences, then KNIME > Python scripting. In the “path to the local python executable” paste the full path to your batch file, e.g. “C:\Users\yourname\knime-workspace\py27.bat”. To be future-proofed, also do this for KNIME > Python (Labs).
    KNIME now calls our batch file instead of calling the system-wide Python executable directly. The batch file activates the “py27_knime” environment before calling Python. This makes sure that the additional libraries required by KNIME are available.
    I guess you could also get your usual Python environment working with KNIME by installing additional packages, but let’s just do what the knowledgeable guys at KNIME have suggested this time. 🙂

    Python scripting preferences for KNIME Analytics Platform in Windows

  6. Now restart KNIME. Try opening or creating a workflow with a node that uses some Python code. It should work now!



Plotting multivariate data with Matplotlib/Pylab: Edgar Anderson’s Iris flower data set

The problem of how to visualize multivariate data sets is something I often face in my work. When using numerical optimization we might have a single objective function and multiple design variables that can be represented by columnar data in the form {x1, x2, x3, … xn, y} a.k.a. NXY. With design spaces of more than a few dimensions it is difficult to visualize them in order to estimate the relationship between each independent variable and the objective, or perform a sensitivity study.

JensG / Pixabay

While perusing recent work in and tools for visualizing such data I stumbled across some nice examples of multivariate data plotting using a famous data set known as the “Iris data set”, also known as Fisher’s Iris data set or Edgar Anderson’s Iris flower data set. It contains data from 50 flowers each of three different flower species, collected in the GaspĂ© Peninsula. This set is not in the NXY form typical of optimization routines, but instead each flower has a number of parameters measured and tabulated; namely sepal length, sepal width, petal length and petal width. In other words there is no resultant Y data that is a function of the design space vector. Instead, it is interesting to plot relationships between the measured parameters to determine if they correlate with each other.

A quick internet search brings up a number of examples where the set has been plotted as a gridded set of subplots, using various software tools. For example, Mike Bostock’s blog post demonstrating his D3.js package, and the version on the wikipedia page.

I decided to try and code a Matplotlib script to generate a similar gridded multiplot from the data set. I did so within a Jupyter Notebook (formerly known as iPython Notebook) running Python 2.7. The data was imported using Pandas and made use of Matplotlib’s Pyplot module. Pandas was used to import the data but it could have been done in a number of different ways; it is just that Pandas is designed to work with csv files containing a mix of types.

The resulting image can be seen below.

Iris flower data set visualization using Matplotlib/pyplot.

Fisher’s Iris data set sometimes known as Anderson’s Iris data set, visualization by Simon Bance using Matplotlib/Pyplot. A multivariate data set introduced by Ronald Fisher in 1936 from data collected by Edgar Anderson on Iris flowers in the GaspĂ© Peninsula.

Here is the script:


"""
https://en.wikipedia.org/wiki/Iris_flower_data_set
A script for plotting multivariate tabular data as gridded scatter plots.
"""
import os
import pandas as pd
import matplotlib.pyplot as plt

inFile = r'iris.dat'

# Check if data file exists:
if not os.path.exists(inFile): sys.exit("File %s does not exist" % inFile)

rootFolder = os.path.dirname(os.path.abspath(inFile))

# Read in the data file
df = pd.read_csv(inFile, delimiter="\t")
headers = list(df.columns.values)
df.head(5) # Prints first n lines to check if we loaded the data file as expected.

# We also have n=4 distinct species in the Species column and I will
# list the species names so we can distinguish them later for plotting:
species = list(df.Species.unique()) # normal python list, thank you very much!
print type(species)

# Here we specify how many columns prepend and append the columns that we want to use.
# For Dakota this would include the objective function(s) column(s) appended to the end.
num_precols = 0
num_obj_fn = 1

# Work out the number of dimensions in each design vector:
num_dims = df.shape[1] - num_obj_fn # We know that there are 3 additional columns (and hope that it stays consistent in future)!
print "Our design vector has %s dimensions: %s" % (num_dims, headers[num_precols:-1])
gridshape = (num_dims, num_dims)
num_plots = num_dims**2
print "Our multivariate grid will therefore be of shape", gridshape, "with a total of", num_plots, "plots"

# Plot the data in a grid of subplots.
fig = plt.figure(figsize=(12, 12))

# Iterate over the correct number of plots.
n = 1

# Create an empty 2D list to store created axes. This alows us to edit them somehow.
axes = [[False for i in range(num_dims)] for j in range(num_dims)]

for j in range(num_dims):
for i in range(num_dims):

# e.g. plt.subplot(nx, ny, plotnumber)
ax = fig.add_subplot(num_dims, num_dims, n) # Plot numbering in this case starts from 1 not zero (MATLAB style indexing)!

# Choose your list of colours
colors = ['red', 'green', 'blue']

for index, s in enumerate(species):

# x axis: For each in the species list look at all rows with that value in the Species column.
# Use the ith column of that subset as the x series.
# y axis: Likewisem, but use the jth column.

if i != j:
ax.scatter(df.where(df['Species'] == s).ix[:,i], df.where(df['Species'] == s).ix[:,j], color=colors[index], label=s)
else:
# Put the variable name on the i=j subplots:
ax.text(0.25, 0.5, headers[i])
pass

# Set axis labels:
ax.set_xlabel(headers[i])
ax.set_ylabel(headers[j])

# Hide axes for all but the plots on the edge:
if j < num_dims - 1: ax.xaxis.set_visible(False) if i > 0:
ax.yaxis.set_visible(False)

if i == 1 and j == 0:
ax.legend(bbox_to_anchor=(3.5, 1), loc=2, borderaxespad=0., title="Species name:")

# Add this axis to the list.
axes[j][i] = ax

n += 1

plt.subplots_adjust(left=0.1, right=0.85, top=0.85, bottom=0.1)

plt.savefig("%s/iris.png" % rootFolder, dpi=300)
plt.show()

Further so-called “classic data sets” are listed at https://en.wikipedia.org/wiki/Data_set#Classic_data_sets.




The new default colormap for matplotlib is called “viridis” and it’s great!

It’s probably not news to anyone in data visualization that the most-used “jet” colormap (sic) (sometimes referred to as “rainbow”) is a bad choice for many reasons.

  • Doesn’t work when printed black & white
  • Doesn’t work well for colourblind people
  • Not linear in colour space, so it’s hard to estimate numerical values from the resulting image

The Matlab team recently developed a new colormap called “parula” but amazingly because Matlab is commercially-licensed software no-one else is allowed to use it!
The guys at Matplotlib have therefore developed their own version, based on the principles of colour theory (covered in my own BSc lecture courses on Visualization 🙂 ) that is actually an improvement on parula. The new Matplotlib default colormap is named “viridis” and you can learn all about it in the following lecture from the SciPy 2015 conference (YouTube ):

Viridis will be the new default colour map from Matplotlib 2.0 onwards, but users of v1.5.1 can also choose to use it using the cmap=plt.cm.viridis command.
I don’t know about you, but I like it a lot and will start using it immediately!