OK, another night, another trial. I must say, tonight was a lot more fun than the last couple of nights, because I really felt that I learned something, which is really the whole point of this exercise. So the example I was trying to code tonight is a simple EOF of a 3D data series. This is something that I just had to code up at work today, so it was a perfect chance for me to try out Sage. For work I ended up altering an existing m-file and running the EOFs in Matlab, but that’s OK, because now I know what I expect to see after running this in Sage.
The data names have been changed to protect the innocent.

# Load in required modules
sage: from import *
sage: from pylab import *
sage: from scipy.stats.stats import nanmean
sage: import datetime

#Load data from NetCDF file
sage: ncfile = netcdf_file('','r')
sage: varnames = ncfile.variables.keys()
sage: varnames


#Now that I have the order I can load into arrays
sage: lon = ncfile.variables[varnames[0]][:]
sage: lat = ncfile.variables[varnames[2]][:]
sage: dates = ncfile.variables[varnames[1]][:]
sage: raw = ncfile.variables[varnames[3]][:,0:50,:] #I only want 50 records in Y
sage: data = raw.copy() #make a copy
sage: data.shape
(124, 50, 151)
sage: (ncycles, ny, nx) = data.shape

#deal with dates
sage: ncfile.variables[varnames[1]].attributes

{'axis': 'TIME',
'time_origin': '15-JAN-1901 00:00:00',
'units': 'HOURS since 1901-01-15 00:00:00'}

sage: off = datetime.datetime(1901,1,15,0,0,0)
sage: months = ones(ncycles)

sage: for i in range(0,ncycles):
....tdel = datetime.timedelta(days=dates[i]/24) = off + tdel
....months[i] = td.month

sage: ind = where(raw<0)
sage: data[ind] = nan

And here was the first real bottleneck, as things just slowed to a crawl as python tried to find all the instances where the data was less than zero. This is something that is instantaneous in Matlab, and took over 30 seconds to go through 124*50*151 values. There must be a faster way to do this.

#Take out monthly averages
sage: mclim = ones((50,151))
sage: for i in range(1,13):
....index = where(months==i)[0]
....mclim = nanmean(data[index,:,:])
....data2[index,:,:] = data[index,:,:] - mclim

data2.shape = (ncycles, nx*ny)
ltmean = nanmean(data2) #get mean of each time series

#take out long term mean
sage: anom = data2.copy()
sage: for i in range(0,ncycles):
....anom[i,:] = data2[i,:] - ltmean

sage: EOF = nan_to_num(anom) #push land back to zero
sage: [u,s,v] = linalg.svd(EOF)
sage: for i in range(0,ncycles):#build array so that we can project eigenvalues back onto timeseries
....s2[i,i] = s[i]
sage: amp = dot(s2.transpose(),u.transpose()) #get amplitude
sage: spatial = v[0:4,:]# pull out spatial fields
sage: ratios = pow(s,2)/sum(pow(s,2))*100 #get %variance explained for each mode
sage: temp = spatial[0,:]
sage: temp.shape = (ny,nx) #push back to original dims
sage: plot(amp)
sage: savefig('amplitude.png')
sage: imshow(flipud(temp))
sage: savefig('spatial.png')


I actually really felt positive about this whole example as I really learned a lot more. This also was probably too large of an array to test out (measure twice cut once!) but it’s what I was working with so I wanted a real world example. The more that I worked in sage the more comfortable I felt as well. The geographic projection issue is still there, as well as some indexing speed issues, but overall, I was really impressed with the Sage/SciPy/NumPy experience today. Overall I feel that more of a transition was made for me last night/today. Which was great timing as a co-worker actually called me and asked if I knew of any free replacements for Matlab…

Technorati Tags:
, , , , , ,

This was quite possibly the worst idea for title naming that I could have thought of. Anyway, I played around a bit more tonight, and I thought that I would give an update to the three people that are waiting with bated breath.

Anywho, I decided to continue trying to map the data from the netcdf file onto a projection, and here’s what I ran into.

It looks like the basemap module is installed (as basemap) but that it depends on matplotlib > 0.98 and 0.91 is installed. I tried to be tricky and move my locally installed matplotlib over to the sage/local/lib/python2.5/site-packages directory but then that version of matplotlib needed a newer version of numpy than what was installed. At this point I tried

hostname $> sage -upgrade

to see if updated packages/modules were available. This started a huge chain reaction of downloads and source compiling to get to the latest, greatest versions. This process took exactly 59m10.482s to complete (I know because it told me!).

But once again, I get this error:

sage: from basemap import basemap

ImportError: your matplotlib is too old – basemap requires version 0.98 or higher, you have version 0.91.1

At this point though, it’s not working on either the linux or OSX platforms due to outdated dependencies, so either I need to find another way to plot mapped projections or use something else.

Again, this isn’t a knock against Sage, because I really don’t think that is an ideal test for this software. But honestly, a lot of why I went for this approach was to avoid having to use separate approaches for data manipulation and visualization, and this would be a common task. Matlab’s mapping toolbox is useless to me for plotting, so I end up using m_map, which is still not as good as GMT, but it gets the job done in house.

My main thoughts at this point are that it seems easy to get into dependency hell here, as one module upgrade can force another, and so on. At this point it’s another block of time spent on setup, and no result. Time to stop for the time being.

Technorati Tags:
, , , , , , ,

Part 1 of the sage experience was just installing the software. This was incredibly easy on both OSX and linux (CentOS 5.2 and Fedora 9). For the Fedora 9 install I just downloaded the latest version of Sage which was compiled for Fedora 8, and this seemed to be just fine.

So for me, I really just wanted to be able to do a few different examples which would be close to “real world applications” for me.

Some things that I would like to be able to do in sage:

1. Load in a 2-D NetCDF satellite data file and display it as a map projection. This should be really simple. I would usually just use GMT for this (a small shell script wrapping psbasemap, grdimage, and pscoast).

2. Load in a data series with dates and locations, and match this to corresponding satellite data in time and space. Normally I would use a perl script that I wrote many moons ago to do this. I would basically sort the data, then match a block of data at a time using GMT’s grdtrack function. I know that this is inefficient, and really I would like to be able to pull extra data in x,y, or t and take the mean or median value, which would be more CPU intensive, but better than matching just one point in space and time to the nearest pixel.

3. Load in a multivariate data series and do multivariate statistics (e.g. LME, GLM/GAM, RDA). This is where the R interface would come into play. Normally I would prepare the data elsewhere, then import the flat table into R and use the R functions. This may involve installing more packages (nlme, mgcv, etc).

4. Load in a 3-D set (x,y,t) of satellite data files and perform an EOF analysis on them (akin to SVD in Matlab). Normally I would do this in Matlab or Ferret. I’m just curious how easy it would be to do this here.

There are other things that I could do, but these are a few off the top of my head, and things that I am doing now, so it would be incentive to try Sage out with. For tonight, I’ll just work on #1, which should be really fast.

The data file I’m using is just a NetCDF file (created by GMT) which I can read with pupynere in python. Here I’m going to use the module (which is actually based on pupynere I believe).

sage: from import *
sage: from pylab import *

# Read in file metadata to object
sage: ncfile = netcdf_file(‘RS2006001_2006031_sst.grd’,’r’)

# get the variables in the data file
sage: ncfile.variables

{‘x’: < object at 0xb47b08c>,
‘y’: < object at 0xb47b16c>,
‘z’: < object at 0xb47b1ec>}

# Yank out data
sage: longitude = ncfile.variables[‘x’][:]
sage: latitude = ncfile.variables[‘y’][:]
sage: sst = ncfile.variables[‘z’][:]

# just plot sst to test 2D image plotting
sage: plot(sst)
[<matplotlib.AxesImage instance at 0xc03636c>]

Nice, but it’s upside down. Let’s flip it vertically.

sage: clf
sage: plot(flipud(sst))
[<matplotlib.AxesImage instance at 0xb86a2ac>]
sage: savefig('temp.png')


Easy, but I want to put this on a projection. Normally I would use the basemap tools which are an add on to matplotlib. I don’t see these installed, and I didn’t see them in the extra sage packages on line, so I downloaded them from SourceForge and installed them.

The first step you have to do is to install the geos package, just read the README in the geos folder and hit


and then we get our first epic fail. Something in the geos chain won’t compile, and I’m just about fried enough to call it quits for this evening.

At this point I’ve been playing with this for more than 2 hours, and I still have yet to make a simple map on a projection. There has to be something I’m missing, but at this point I’m going to pause until tomorrow. So not the best testing evening, but there are some positives so far. The bundling of most packages is a plus, and the ease of loading in NetCDF files is nice. Data displays well using the Pylab interface, even though I am still forced to save to a file at this point.

So immediate goals:

1. Get a backend working for viewing plots in widgets (akin to ipython -pylab)

2. Get the basemap tools installed so that I can make a map with a projection!

Technorati Tags:
, , , , , ,