|
Let's say, for example, that you have written a Perceptron or Neural
Network learning algorithm, and you have been asked to plot the accuracy
versus epochs. This tutorial will show you how to do that.
- First, put your data into a text file of
comma-separated values (also known as a CSV file).
- The easiest way to put your data into a CSV file is
to simply print it to stdout, and when you
run your program, pipe its output to a file.
- For example, let's say your main training loop looks
something like this:
int epochs = 0; while(true) { do_one_training_epoch(trainingSet); epochs++; double mseTraining = evaluate_accuracy(trainingSet); double mseValidation = evaluate_accuracy(validationSet); if(check_stopping_criteria(mseValidation)) break; }
Since we want to plot accuracy versus
epochs, we will print those values:
int epochs = 0; while(true) { do_one_training_epoch(trainingSet); epochs++; double mseTraining = evaluate_accuracy(trainingSet); double mseValidation = evaluate_accuracy(validationSet); System.out.write(epochs); System.out.write(", "); System.out.write(mseTraining); System.out.write(", "); System.out.write(mseValidation); System.out.write("\n"); if(check_stopping_criteria(sse)) break; }
- Now, just run your program and pipe the output to a
file. Example:
./MLSystemManagerDbg -L neuralnet -A iris.arff -E training > accuracy_vs_epochs.csv
2.
3.
Next, open your favorite spreadsheet
program. Spreadsheet programs like MS-Excel are commonly used tools to plot
data. I also recommend the spreadsheet program in LibreOffice
because it is free and open, and it is good enough. Import your CSV file.
Make sure that each value is in its own cell.
4.
Use your mouse to draw a box around
your data in the spreadsheet.
5.
Click on "insert"->"chart".
You will see several chart types. The one you want is called something like
"XY (Scatter)". This type of chart will use the first column for
the horizontal axis, and all the other columns will be plotted on the
vertical axis.
6.
If the chart is too small, grab one
of the corners and stretch it bigger.
7.
At this point, I like to take a
screen shot of the chart, and paste it into a simple painting program. (I
like to use Kolourpaint because it is simple and
fast.) I use the painting program to label the axes, to scale it to the right
size for my report, to label my lines, etc. Yes, I know that the spreadsheet
provides a mechanism to label the axes, but it isn't very flexible. The
painting program will let me do whatever I want. I can circle things that I wish
to highlight for the reader, I can label my lines right on top of the chart,
I can draw arrows to direct the reader's attention to interesting trends in
the graph, etc. Don't be constrained by the capabilities of your plotting
tool--in fact, I recommend not even wasting your time learning them, since
painting programs will always be more flexible.
Now, where should you go when you're ready to move beyond the basic
capabilities of your spreadsheet program? My advice is to choose a plotting
program that will enable you to automate your processes. If you discover a
bug in your algorithm, or if you think of a cool tweak that you would like to
try, you may not like the idea of having to regenerate your plots by hand. On
the other hand, it's not such a big deal if all you have to do is run a
script and it regenerates all of your results, and even the charts come out
ready for you to add your final touches.
Here are some plotting tools that can be (at least partially) automated.
(These have a much steeper learning curve than the spreadsheets, but they
will pay you back in the long run if you make a lot of charts.)
- GnuPlot (Warning: Not really open source,
despite the deceptive prefix.)
- Waffles
- Know of any others?
|