Page History
Below is a Python example that gathers some data from an experiment and saves it to a hdf5 file for future use. The gathered data will be used to select which shots to use for future analysis. It demonstrates the use of h5py to write and read compound data types to hdf5 files.
The data used is a CXI tutorial run from the Tutorial test data/reg/d/psdm/cxi/cxitut13/.
Code Block | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|
| ||||||||||
import sys import numpy as np import h5py from collections import defaultdict import psana def getFirstEpicsValue(epics,pvName): '''convenience function to get first epics value from the given pv ''' return epics.getPV(pvName).data()[0] def gatherData(eventsToProcessmaxEventsToProcess=None): '''Goes through a specific experiment and run, gathers specific values from the data, returns a two numpy arrays: compoundDatadataArray: the gathered data compoundTimetimeArray: the event times for each row of the gathered data has two fields - 'seconds' 'nanoseconds', both integers ''' dsdataSource = psana.DataSource("exp=cxitut13:run=0022") src_diodesrcDiode = psana.Source('DetInfo(CxiDg4.0:Ipimb.0)') src_gasdetsrcGasDet = psana.Source('BldInfo(FEEGasDetEnergy)') epicsepicsStore = dsdataSource.env().epicsStore() dataLists# = defaultdict(list) for num,evt in enumerate(ds.events()):define data type for data array and time array dataDtype = np.dtype([('aG',np.float), if eventsToProcess is not None and num >= eventsToProcess: ('aD',np.float), ('aH',np.float), break diode = evt.get(psana.Ipimb.DataV2,src_diode) ('aX',np.float), ('aY',np.float), ('aZ',np.float)]) timeDtype dataLists['aDiode'].append(diode.channel0Volts()+diode.channel1Volts()+= np.dtype([('seconds',np.uint32),('nanoseconds',np.uint32)]) # to minimize realloc of arrays, set the number of elements in a block size blockSize = 100000 dataArray = np.zeros(blockSize, dtype=dataDtype) timeArray = diodenp.channel2Volts()+diode.channel3Volts())zeros(blockSize, dtype=timeDtype) # motor calibration GasDet EncoderScale = evt.get(psana.Bld.BldDataFEEGasDetEnergy,src_gasdet) 0.005 Xoffset = 17050. Yoffset = 1056510. dataLists['aGasDet'].append(GasDet.f_11_ENRC())Zoffset = 1830400. eventIdnextArrayRow = evt.get(psana.EventId)0 for eventNumber, evt dataLists['aTime'].append(eventId.timein enumerate(dataSource.events()): if maxEventsToProcess is not None and dataLists['aSampleX'].append(getFirstEpicsValue(epics,'CXI:SC1:MZM:08:ENCPOSITIONGET')) nextArrayRow >= maxEventsToProcess: dataLists['aSampleY'].append(getFirstEpicsValue(epics,'CXI:SC1:MZM:09:ENCPOSITIONGET')) break dataLists['aSampleZ'].append(getFirstEpicsValue(epics,'CXI:SC1:MZM:10:ENCPOSITIONGET'))diode = evt.get(psana.Ipimb.DataV2,srcDiode) dataLists['aKB1Hpitch'].append(getFirstEpicsValue(epics,'CXI:KB1:MMS:07.RBV'))gasDet = evt.get(psana.Bld.BldDataFEEGasDetEnergy,srcGasDet) if# check numthat %all 1000data == 0: is present in event before adding to dataArray print "event number=%8dif diode.chann8el0Volts=%f GasDet.f_11_ENCR=%f" % \ is None or gasDet is None: continue (num, diode.channel0Volts(),GasDet.f_11_ENRC()) # motor calibration # grow arrays if they are to small EncoderScaleif nextArrayRow >= 0.005 timeArray.size: Xoffset = 17050.timeArray.resize(timeArray.size + blockSize) Yoffset = 1056510. Zoffset = 1830400. dataArray.resize(dataArray.size + blockSize) compoundDataTypetimeArray[nextArrayRow] = npevt.dtype([('aG',np.float), ('aD',np.float), ('aH',np.float), get(psana.EventId).time() diodeSum = diode.channel0Volts() + diode.channel1Volts() + \ ('aX',np.float), ('aY',np.float), ('aZ',np.float)]) diode.channel2Volts() + diode.channel3Volts() compoundData = np.zeros(len(dataLists dataArray[nextArrayRow]['aTimeaD']), dtype=compoundDataType) diodeSum compoundDatadataArray[nextArrayRow]['aG'] = gasDet.f_11_ENRC() sampleX = dataLists['aGasDet'] epicsStore.getPV('CXI:SC1:MZM:08:ENCPOSITIONGET').value(0) compoundData['aD'] sampleY = dataLists['aDiode'] epicsStore.getPV('CXI:SC1:MZM:09:ENCPOSITIONGET').value(0) compoundData['aH'] sampleZ = dataLists['aKB1Hpitch']epicsStore.getPV('CXI:SC1:MZM:10:ENCPOSITIONGET').value(0) compoundData dataArray[nextArrayRow]['aX'] = np.array(dataLists['aSampleX'])sampleX * EncoderScale - Xoffset compoundData dataArray[nextArrayRow]['aY'] = np.array(dataLists['aSampleY'])sampleY * EncoderScale - Yoffset compoundDatadataArray[nextArrayRow]['aZ'] = np.array(dataLists['aSampleZ'])sampleZ * EncoderScale - Zoffset timeDataType = np.dtype([('seconds',np.uint32),('nanoseconds',np.uint32)]) compoundTime = np.zeros(len(dataLists['aTime']), dtype=timeDataType) compoundTime[:] = dataLists['aTime'] dataArray[nextArrayRow]['aH'] = \ epicsStore.getPV('CXI:KB1:MMS:07.RBV').value(0) if nextArrayRow % 1000 == 0: print "event %8d diode.channel0Volts=%f gasDet.f_11_ENCR=%f" % \ (eventNumber, diode.channel0Volts(),gasDet.f_11_ENRC()) nextArrayRow += 1 # shrink arrays to number of events we stored data from timeArray.resize(nextArrayRow) dataArray.resize(nextArrayRow) return compoundDatadataArray, compoundTimetimeArray def writeCompoundDataToH5H5WriteDataAndTime(compoundDatadataArray, compoundTimetimeArray, h5filename): f = h5py.File(h5filename,'w') data_comp_typedataDtype = compoundDatadataArray.dtype data_dsetdataDset = f.create_dataset('data', compoundDatadataArray.shape, data_comp_typedataDtype) data_dsetdataDset[:] = compoundDatadataArray time_comp_typetimeDtype = compoundTimetimeArray.dtype time_dsettimeDset = f.create_dataset('time',compoundTimetimeArray.shape, time_comp_typetimeDtype) time_dsettimeDset[:] = compoundTimetimeArray f.close() def readCompoundDataFromH5H5ReadDataAndTime(h5filename): f = h5py.File(h5filename) data_dsetdataDset = f['data'] data_arraydataArray = data_dsetdataDset[:] time_dsettimeDset = f['time'] time_arraytimeArray = time_dsettimeDset[:] f.close() return data_arraydataArray, time_arraytimeArray if __name__ == '__main__': maxArraySize = None if len(sys.argv) > 1: maxArraySize = int(sys.argv[1]) datadataArray,timePairtimeArray = gatherData(maxArraySize) writeCompoundDataToH5H5WriteDataAndTime(datadataArray,timePair timeArray, "saved_output.h5") datadataArray,timePairtimeArray = readCompoundDataFromH5H5ReadDataAndTime("saved_output.h5") |
To use this script:
- place it in your release directory
- run ipython
enter the commands:
import gather_savegatherSave
datadataArray,timePairtimeArray = gather_savegatherSave.gatherData()
gather_savegatherSave.writeCompoundDataToH5H5WriteDataAndTime(datadataArray,timePair timeArray, "saved_output.h5")
datadataArray,timePairtimeArray = gather_save.readCompoundDataFromH5H5ReadDataAndTime("saved_output.h5")
The function gatherData() is one that needs to be modified for different datasets. writeCompoundData H5WriteDataAndTime and readCompoundData H5ReadDataAndTime will not.
data dataArray is a numpy array with 6 named fields that gather different values from the events, epics pv's , a value from the gas detector, and the voltage sum of a Diode. The fields have names like 'aD' (the Diode sum) and 'aG' for the gas detector value.
...
Code Block | ||
---|---|---|
| ||
logicalIndex = datadataArray['aG'] > 0.84 # a mask that is 1 when 'aG' is greater than 0.84 data['aD'][logicalIndex] # the mask is used to get diode values when 'aG' is > than 0.84 logicalIndex.nonzero()[0] # turn the mask into a list of positions, see the documentation on # the numpy function nonzero # http://docs.scipy.org/doc/numpy/reference/generated/numpy.nonzero.html # likewise one can do import numpy as np np.where(data['aG'] > 0.84)[0] # the [0] is necessary to get the indicies along the first axis |
...
Below we discuss how things are done in the example.
...
Creation of Numpy Array with Named Fields
Numpy arrays are very efficient data structures in Python. This example creates two numpy arrays to store the event data. These arrays have named fields which provides a dictionary style access to the data. Note the two numpy dtypes (data types) created that define these arrays:
Code Block | ||
---|---|---|
| ||
from collections import defaultdict ... dataLists = defaultdict(list) # define data type for data array and time array for num,evt in enumerate(ds.events()): ...dataDtype = np.dtype([('aG',np.float), ('aD',np.float), ('aH',np.float), ('aX',np.float), ('aY',np.float), ('aZ',np.float)]) timeDtype dataLists['aTime'].append(eventId.time()) |
The use of the standard Python defaultdict allows us to avoid initializing the keys in dataList. The price is that if we later make a typo when we retrieve 'aTime', from dataLists, the error message will not make this clear.
Creation of Numpy Array with Named Fields
= np.dtype([('seconds',np.uint32),('nanoseconds',np.uint32)]) |
A 32 bit unsigned int is used for the seconds and nanoseconds for the timeArray as this is how these fields are stored in the xtc, but one could use np.int or np.uint as well.
For more information on numpy dtypes visit the documentation: Numpy Dtypes
Checking that all Data is Present in the Event
One needs to check that the data one wants from an event is present. In the example there are two counters - eventNumber and nextArrayRow. eventCounter keeps track of which event we are reading and is only used for printing a status message. nextArrayRow is a zero-up counter of events that include both the diode and gasDet data. If both are not present - we go on to the next event and do not increment nextArrayRow.
Using Array Blocks to Read in Data
Manipulations using numpy arrays are most efficient when the final size of the array is known ahead of time. Although numpy arrays have an append method, using it for each event can When creating numpy arrays, it is more efficient to create with a known size. You can append to an existing numpy array, but to do this with every event may lead to a great deal of memory reallocation. In this example, we read the data into Python lists. Once we have all the data, we create the numpy array of the known size.To create a numpy array with named fields you must define a dtype. For this example where each field is a float, it is fairly straightforward's of the arrays. Therefore we start with arrays that hold 100,000 elements, and grow them by 100,000 if we need to. In the end we shrink the arrays down to the size that is used. Note the lines:
Code Block | ||
---|---|---|
| ||
import numpy as np ... compoundDataType # to minimize realloc of arrays, set the number of elements in a block size blockSize = 100000 dataArray = np.dtype([('aG',np.float), ('aD',np.float), ('aH',np.float),zeros(blockSize, dtype=dataDtype) timeArray = np.zeros(blockSize, dtype=timeDtype) ... # grow arrays if they are to small if ('aX',np.float), ('aY',np.float), ('aZ',np.float)]) compoundData = np.zeros(len(dataLists['aTime']), dtype=compoundDataType)nextArrayRow >= timeArray.size: timeArray.resize(timeArray.size + blockSize) dataArray.resize(dataArray.size + blockSize) ... # shrink arrays to number of events we stored data from timeArray.resize(nextArrayRow) dataArray.resize(nextArrayRow) |
numpy dtypes can get quite complicated, for more information visit the documentation: http://docs.scipy.org/doc/numpy/reference/arrays.dtypes.htmlWri
Writing an HDF5 File of Compound Data
...
Code Block | ||
---|---|---|
| ||
import h5py f = h5py.File('myfile.h5''h5filename,'w') data_comp_typedataDtype = compoundDatadataArray.dtype data_dsetdataDset = f.create_dataset('data', compoundDatadataArray.shape, data_comp_type) data_dsetdataDtype) dataDset[:] = dataArray timeDtype = timeArray.dtype timeDset = f.create_dataset('time',timeArray.shape, timeDtype) timeDset[:] = compoundDatatimeArray f.close() |
It is important to call the close() method of the h5py.File object when done.