developer.qdataset

From autoplot.org

Jump to: navigation, search

Purpose: write new page documenting QDataSet, after discovering a lack of resources for describing Bundles. This will replace http://autoplot.org/QDataSet, and will be completed as a part of the migration of QDataSet into the das2 core.

Audience: das2 developers and interested scientists.

Contents

  1. Introduction
  2. Simple Arrays
    1. Rank 1 Example
    2. Rank 2 Example
    3. Rank 3 Example
  3. More Metadata
  4. Combining Datasets With DEPEND Property
  5. Example Datasets
    1. Rank 1 QDataSet
    2. Rank 2 QDataSet
    3. Others
  6. High Rank and Rank 0 Datasets
  7. Some example operators
  8. Ordinal Data
  9. Bundles
    1. High Rank Bundles
  10. BINS Dimensions
  11. Schemes
  12. QDataSet Operations should Handle Metadata
  13. Comparing IDL array operations to QDataSets

1. Introduction

QDataSet is the mature interface used for representing, or modelling, data in Java. When one interface is used for all data, a rich set of operators can be developed, and software is written more quickly and correctly. Often QDataSet is referred to as a "Data Model."

QDataSet was originally developed in 2005-2007 for PaPCO, an IDL application, to provide a standard way of representing data between plug-in modules for different instruments and different spacecraft. Autoplot, a Java application, was begun in 2007, and completed this model. Note that one of the design goals of QDataSet is to do most of the work in semantics instead of syntax, so that the model can be implemented in any software framework. Autoplot is more than 10 years old and QDataSet is quite mature, providing not just the interface, but also hundreds of operators and sources for creating QDataSets from data files.

Because QDataSets carry rich metadata that allow the numbers to be interpreted, rich operations can be developed that need little documentation can be developed as well. For example waveform dataset can be passed to the fftPower function, and a power spectrum returned, with all the labels and frequency channels correctly calculated.

QDataSet reuses a number of conventions found in CDF and NetCDF files and can be thought of a software interface to these formats. QDataSet was presented at AGU: 2011 Poster

2. Simple Arrays

A QDataSet is a basically an array with metadata attached to it. An array is a set of N numbers accessed by an index from 0 to N-1. Some languages support 2-D and 3-D arrays, so to access any number, you need two or three indices, respectively.

With QDataSet, we call the number of indices the "rank" of the dataset. Core are three accessor methods:

ds.rank()         # The number of indices this QDataSet uses.
ds.length()       # The number of values in the 0th (first) index of the QDataSet.
ds.value(i)       # A value within a rank 1 QDataSet.
# or
ds.value(i,j)     # A value within a rank 2 QDataSet.
# or
ds.value(i,j,k)   # A value within a rank 3 QDataSet.

A rank 1 dataset is similar to a simple 1-D array. A number of elements are available (ds.length()), and each element is accessed with ds.value(i). The index i can take on values of 0,1,2,...,ds.length()-1.

Metadata for a QDataSet is accessed with the property method, so for example, ds.property('LABEL') returns a label for the dataset, and ds.property('UNITS') returns a units representation. QDataSet.LABEL and QDataSet.UNITS provide easy access to these property names, as well as descriptions for each.

Autoplot's Jython uses QDataSets as if they were any other array. In this case ds[i] returns each value of a rank 1 dataset. ds[i,j] accesses values of a rank 2 dataset. Examples in this document can be run in Autoplot, using [menubar]→Options→Enable Feature→Script Panel, and then pressing the "execute" button to run the script. The Log Console is usually run as well, which will show the output of print commands.

2.1. Rank 1 Example

# Create a Jython array
array= [ 2,3,2,4,5,4,2,1 ]
print 'Jython Array: ', array # Jython Array:  [2, 3, 2, 4, 5, 4, 2, 1]
 
# Convert Jython array into a QDataSet
ds= dataset( array )
print ds.rank() # 1
print ds.length() # 8
 
# Add metadata to QDataSet
ds.putProperty( QDataSet.LABEL, "frequency" )
ds.putProperty( QDataSet.UNITS, Units.hertz )
 
# Access metadata 
print ds.property( QDataSet.LABEL ) # frequency
 
# Extract data
array= [ ds.value(0), ds.value(1), ds.value(2) ]
print 'Jython Array of floats: ', array # '[2.0, 3.0, 2.0]'
# or
array= [ ds[0], ds[1], ds[2] ] 
print 'Jython Array of QDataSets: ', array # '[2.0 Hz, 3.0 Hz, 2.0 Hz]'
print 'First element has rank %d and length %d' % ( array[0].rank(), array[0].length() )

Autoplot's scripting has a command "getDataSet" which reads data using an Autoplot URI.

# read in rank 1 dataset from CDAWeb
ds= getDataSet('http://autoplot.org/data/somedata.cdf?Magnitude&doDep=no')
print ds  # 'Magnitude[24] (nT)'

2.2. Rank 2 Example

A rank 2 dataset is similar to a 2-D array. Like Java and unlike Fortran, this is an array of arrays, so a rank 2 QDataSet can be thought of as a set of rank 1 datasets, and each rank 1 dataset can have a different length. Therefore to get the length of the second index, length(i) is called, and value(i,j) is used to access the values. Note most rank 2 QDataSets are actually repeated measurements of the same thing, so length(0) will equal length(1) and all other lengths. This special case is called "Qubes," and these are like Fortran or IDL arrays. The "size" function returns the length of each dimension, and makes sure that the data really is a qube.

seed=5334
x = randomn(seed,10,5)
print '%5.2f' % x[0,0]
(nrow,ncol)= size(x)     # this would raise an exception if the data is not a Qube.
for i in xrange(nrow):
   for j in xrange(ncol):
      print  '%5.2f' % x[i,j],      # trailing comma suppresses the newline
   print ''
 
scPos= getDataSet( 'http://autoplot.org/data/somedata.cdf?doDep=no&SC_pos_GSE' )
print 'scPos should have rank 2: ', scPos.rank()

2.3. Rank 3 Example

Higher rank datasets exist as well, namely rank 3 and rank 4 datasets. These have value methods value(i,j,k) and value(i,j,k,l), and might be Qubes as well.

The method "slice" accesses the dataset at the position. For example:

ds.slice(0) 

extracts a "slice" of the dataset at the index. Calling the slice operator of a rank 3 dataset returns a rank 2, and so on. In Space Physics, we have lots of time series data sets, and typically the first index corresponds to time. The slice operator allows one to look at one of the measurements.

scPos= getDataSet( 'http://autoplot.org/data/somedata.cdf?doDep=no&SC_pos_GSE' )
print 'scPos should have rank 2: ', scPos.rank()
print 'slice of data: ', scPos.slice(0)
print 'and one measurement: ', scPos.slice(0).slice(0)

3. More Metadata

We can attach properties to datasets, like:

  • UNITS The units. In Autoplot this is a Java class of enumerated values plus additional ones encountered.
  • LABEL Human-readable label for plot axes
  • TITLE Longer, human-readable title
  • NAME Java-identifier for the dataset "flux" or "energyFlux" or "spacecraft_time", suggest a name to use in scripts.
  • VALID_MAX, double, maximum valid value (inclusive)
  • VALID_MIN, double, minimum valid value (inclusive)

4. Combining Datasets With DEPEND Property

So at what time was each measurement made? There are properties connecting QDataSets which do this. For example, suppose there is an array of timetags and another that is an array of density values at each time. The density dataset has a property "DEPEND_0" which is the timetags.

ttags= density.property( QDataSet.DEPEND_0 )

Properties:

  • DEPEND_0 The independent parameter "causing" the first index, which is also a QDataSet.
  • DEPEND_1 The independent parameter "causing" the second index, also a QDataSet.
  • DEPEND_2 The independent parameter "causing" the third index...
  • DEPEND_3 ...

When we need a quick string representation of the data, we would print

density[epoch=288]

to indicate there are 288 elements in the rank 1 dataset "density", and it DEPENDS on timetags in a dataset called "epoch".

5. Example Datasets

5.1. Rank 1 QDataSet

ds= getDataSet( 'vap+cdaweb:ds=OMNI_HRO2_5MIN&filter=omni&id=proton_density&timerange=2018-10-01' )
print ds.property( QDataSet.LABEL )     # "Proton Density (n/cc)"
print ds.property( QDataSet.UNITS )     # "n/cc"
 
epoch= ds.property( QDataSet.DEPEND_0 ) 
print epoch.property( QDataSet.UNITS )  # "cdfEpoch"
 
 
 
#

5.2. Rank 2 QDataSet

ds= getDataSet( 'https://cdaweb.gsfc.nasa.gov/istp_public/data/polar/hydra/hyd_h0/2000/po_h0_hyd_20000101_v01.cdf?ELECTRON_DIFFERENTIAL_ENERGY_FLUX' )
print ds.rank()                          # 2
print ds.property( QDataSet.DEPEND_0 )   # EPOCH
print ds.property( QDataSet.DEPEND_1 )   # ENERGY_ELE
 
from org.das2.qds import DataSetUtil
print DataSetUtil.isQube(ds) # 1 (True)
 
dep0= ds.property( QDataSet.DEPEND_0 )
print dep0.property( QDataSet.UNITS )
 
dep1=  ds.property( QDataSet.DEPEND_1 )
print dep1.property( QDataSet.UNITS )
print dep1.property( QDataSet.SCALE_TYPE )  # log

5.3. Others

Autoplot's Metadata tab allows the QDataSet to be browsed. Under the "Dataset" node, the loaded QDataSet's properties are shown, and one can see how different types of data are represented.

6. High Rank and Rank 0 Datasets

Besides rank 1 and rank 2 datasets, there are rank 3 and rank 4 datasets. You could really have any rank you would want, and should a need for a 5-index dataset arise, it would be added to Autoplot. There are also rank 0 datasets, which model scalar quantities. Note that this has just the accessor "value()" and one can still call properties to get metadata about the one value.

Note that in Jython, ds[0] doesn't actually return a double value. It returns a rank 0 dataset. Typically Jython knows how to convert this to what it needs, but sometimes you must explicitly convert the value, like so:

for i in xrange( int(ds[0]) ): print i

7. Some example operators

Autoplot's implementation of QDataSet offers lots of functionality. For example:

ds= randomn(5334,10000)   # 5334 is just a seed to make the pseudo-random sequence repeatable, which is useful for testing.
plot( histogram(ds,-5,5,0.1) )

There are hundreds more operators. Type in "ds=" and hit tab in the editor to see completions.

ds= randomn(5334,10000)
tt= timegen('2011-01-01T00:00', '1 days', 10000 )
plot( 0, tt, ds )
plot( 1, tt, smooth(ds,101) )
reset()
ds= getDataSet( 'http://autoplot.org/data/wav/chickadeesong.wav' )   # rank 1 dataset
print ds.property( QDataSet.DEPEND_0 ).property( QDataSet.UNITS )    # seconds
ps= fftPower( ds, 1024, 2, 'Hanning', monitor )          # returns rank 2 power spectrum
print ps.property(QDataSet.DEPEND_1)                     # 'dataset[511] (Hz)'
plot( ps )

8. Ordinal Data

The units representation returned by the UNITS property may be a special unit that allows encoding of ordinal data. For example, instead of "density" measured in 1/cc, we might have "instrument mode" which has the values "on", "off", and "testing."

9. Bundles

The simplest bundle is a rank 2 dataset with DEPEND_1 equal to a QDataSet containing ordinal data, labeling each channel. The following code shows this.

ll= labels( ['data-x','data-y','data-z'] )
ds= zeros(201,3)
ds[:,0]= sin( linspace(0,6*PI,201) )
ds[:,1]= sin( PI/3+linspace(0,6*PI,201) )
ds[:,2]= sin( 2*PI/3+linspace(0,6*PI,201) )
ds.putProperty( QDataSet.DEPEND_0, linspace( '2019-10-10T00:30', '2019-10-10T19:30', 201 ) )
ds.putProperty( QDataSet.DEPEND_1, ll )
plot(ds)

In this simple case, the three bundled channels must all have the same units.

We can also model more complex tuples of data, where many measurements are made at the same time. Suppose at a given time we collect density, velocity, and flux. We then have an array with elements time, density, vx, vy, vz, f1, f2, f3, f4. You can get an example bundle with org.qds.examples.Schemes:

ds= getDataSet( 'vap+hapi:https://jfaden.net/HapiServerDemo/hapi?id=Iowa+City+Conditions&timerange=2019-09-01/2019-09-11' )
plot( unbundle(ds,'Humidity') )

These values can be represented in a rank 2 QDataSet with each record having nine elements, but how do you store the different labels and units? For this, QDataSet uses the property BUNDLE_1, which is itself a QDataSet, but a odd one called a BundleDataSetDescriptor. It is always rank 2, and carries the properties for each of the bundled elements. For example:

bds= ds.property( QDataSet.BUNDLE_1 )
print bds.property( QDataSet.TITLE, 1 )   #  'Wind Speed'  
print bds.property( QDataSet.LABEL, 2 )   #  'DewPoint'
print bds.property( QDataSet.UNITS, 2 )   #  'deg F'
print bds.property( QDataSet.NAME, 3 )    #  'Humidity'
print bds.property( QDataSet.UNITS, 3 )   #  'percent'

Having the BundleDataSetDescriptor allows us to model commonly-seen rank 2 table of numbers:

2019-08-31T23:52:00.000Z 67.0 9.0 57.0 71.0 -1.0E31 -1.0E31 30.17 1021.6 0.01 
2019-09-01T05:52:00.000Z 64.0 5.0 59.0 84.0 -1.0E31 -1.0E31 30.19 1022.3 0.0 
2019-09-01T06:52:00.000Z 65.0 3.0 58.0 78.0 -1.0E31 -1.0E31 30.19 1021.9 0.0 

which can be modeled and printed like so:

print ds.rank()                                 #   '2'
bds= ds.property( QDataSet.BUNDLE_1 )  
print bds.property( QDataSet.FILL_VALUE, 4 )    #  -1e31
print bds.property( QDataSet.LABEL, 4 )         #  'WindChill'
 
dep0= ds.property( QDataSet.DEPEND_0 )
for i in range(3):
    print dep0[i].svalue(),
    for j in range(ds.length(0)):
        print ds[i,j].value(),
    print ''

and a slice results in the nine-element tuple:

ds.slice(0)                               #   results in the tuple of individual measurements

To access individual datasets from the bundle, the unbundle function can be used:

print unbundle( ds, 2 )           # results in the dew point column
print unbundle( ds, 'Humidity' )  # results in the dataset with this name.

9.1. High Rank Bundles

A special "high-rank" mode is also supported to unbundle rank 2 (or greater) datasets:

from org.das2.qds.examples.Schemes import complexBundleDataSet2
ds= complexBundleDataSet2()
v= unbundle( ds, 'V' )
plot(v)

It can do this because there is metadata within the bundle properties describing this dataset:

bds= ds.property( QDataSet.BUNDLE_1 )
print bds.property( QDataSet.NAME, 4 )
print bds.property( QDataSet.START_INDEX, 4 )   # 2, because Vx is at 2.
istart= bds.property( QDataSet.START_INDEX, 4 )
print bds.property( QDataSet.ELEMENT_NAME, istart )        # "V"
print bds.property( QDataSet.ELEMENT_DIMENSIONS, istart )  # "[3]"

Note that a rank 2 bundle could contain a rank 3 dataset. It can also contain a simple bundle, but it cannot contain another bundle. One more, slightly embarrassing, note, sometimes DEPEND_0 will be a part of the bundle itself. It's still not clear what the better form is, but codes generally support both forms.

10. BINS Dimensions

So you've seen how DEPEND_1 can describe the columns of a rank 2 spectrogram, and how BUNDLE_1 can describe the columns of an ASCII table. There are a couple of other structural properties which are useful.

Sometimes you don't have just one number describing the position of a measurement, but a range. The BINS_1 property is useful here. It is just a string identifying each of the numbers in the dimension. For example, "min,max" indicates that the two numbers are the bounds of the bounds containing the data.

11. Schemes

QDataSet is a simple interface for describing complex things. Clearly there are things which will fit into QDataSet but are not semantically correct. Suppose you have a Rank 1, 10 element dataset with a DEPEND_0 Rank 1 11 element dataset. Likewise there might be things you can imagine which do make sense but codes don't support. One example of this is rank 2 data with rank 2 DEPEND_1. There were CDF files with data like this, and when we saw this was useful, we added support for it. Every routine of the Java QDataSet library had to check the rank of DEPEND_1 to make sure it was properly handled. If the data was sliced, then its DEPEND_1 would also have to be sliced. This design has allowed QDataSet to grow organically to meet community needs.

To enumerate all the different forms of QDataSets supported, a testable Java code was introduced. This code is org.das2.qds.examples.Schemes, found at

https://sourceforge.net/p/autoplot/code/HEAD/tree/autoplot/trunk/QDataSet/src/org/das2/qds/examples/Schemes.java

To access this in Autoplot Jython, try:

from org.das2.qds.examples import Schemes
plot( 0, Schemes.simpleSpectrogram())
plot( 1, Schemes.rank2Waveform())

12. QDataSet Operations should Handle Metadata

Many operations are included in the Java/Jython implementation of QDataSet. It is expected that operators should return results with as much metadata preserved as possible. For example, the trim operator returns a subset of a dataset. This result ought to have the same UNITS and LABEL as the original data. Also, slice should take DEPEND_1 and make its result have DEPEND_0. Likewise for BINS_1 and BUNDLE_1.

Note too that Jython supports operator overloading, so "+" can be defined as the addition operator which considers units. This is all implemented as well, so adding "5cm" to "5m" will result in "505 cm."

Jython also has a dataset operator which will parse the string (or array or scalar) to turn it into a QDataSet.

print dataset("5cm")+dataset("5m")  # "505.0cm"

13. Comparing IDL array operations to QDataSets

There are a number of differences between IDLs array operations and QDataSet operations.

Note in Jython ds[2,3] is the same as ds.slice(2).slice(3) and similar to ds.value(2,3). (Similar because value(2,3) returns a double, while ds.slice(2).slice(3) returns a rank 0 dataset.)

In the following table, "ds" may change from row to row, but is the same for the two columns. TODO: these should probably be equivalent codes.

IDLQDataSets in Autoplot's Jython
ds[1,*] returns 1 by n array, reform used to remove the 1ds[1,:] returns rank 1 n-element dataset.
ds[1] is a scalards[1] is a rank 0 dataset
arrays can have 1 to 8 indicesdatasets can have 0 to 4 indices
single index aliases to multi-index ds[19] is the same as ds[3,4] of 4 by 5 arraysingle index slices to dataset of rank N-1, DataSetIterator should be used to access all elements
where command returns aliased indices for 2-D arrayswhere returns r[n,2] of indices for each dimension.
where command returns -1 to indicate no matcheswhere returns zero-length dataset for no matches.
zero-length arrays supported in IDL 8.1 and upzero-length arrays have always been supported
size command shows number of dimensions and size of each dimension (its geometry)ds.rank() and ds.length() are used to discover dataset geometry.
arrays are qubes, having the same number of columns in each roweach record could have a different length, QDataSet.QUBE property asserts that the dataset is a qube.
print, ds prints all the elementsfor d in ds: print d prints all the elements
help, ds shows information about the arrayprint ds shows information about the array
!dtor converts to radianstoRadians(ds) converts to radians
Personal tools