Stitch1D investigation

[Owen Arnold]

Mentioned in #9440 (see description). There may be a difference in scale factor between Mantid's Stitch1D, and the equivalent algorithm generated by genie. I strongly doubt that Stitch1DMany is the cause of this because it's simply a wrapper of Stitch1D.

Stitch1D is well used by the reflectometry group, and is based on their previous combineMulti script, so I doubt that there are really serious issues here. We should investigate anyway, but this is not critical.

I will need user help testing this.

[Owen Arnold]

One thought, the way that the reported scale factor is calculated is to take the ratio of only one corresponding y bin. i.e. y1[0]/y2[0] (see Stitch1D) this is a possible source of the difference.

[Samuel Jackson]

I've been trying to write a usage example for this algorithm and came across the following problem. If I create two workspaces with the same function, but with different x ranges and attempt to stitch them together I get a function that looks reasonable, but the scale appears to be doubled.

The following script generates two workspaces with a sine with over a upper and lower range and tries to stitch them together. The result is good, but the scale is double what I would expect for the result:

import numpy as np

def createSineWorkspace(name,start,end):
        dataX = np.arange(start,end,0.1)
        dataY = np.sin(dataX)
        ws = CreateWorkspace(dataX, dataY, OutputWorkspace=name)
        ws = ConvertToHistogram(ws, OutputWorkspace=name)
        return ws

#create some histograms to test the algorithm with
ws1 = createSineWorkspace('left', 0, 10)
ws2 = createSineWorkspace('right', 9, 20)

#stitch the two workspaces together
ws3, sf = Stitch1D(ws1, ws2)

#check against reference workspace
ref = createSineWorkspace('reference',0,20)

plotSpectrum(['left', 'right', 'ws3'], 0)
print sf

[Owen Arnold]

output sf comes out right!

[Samuel Jackson]

Another interesting point:

import numpy as np

def gaussian(x, mu, sigma):
        return (1/ sigma * np.sqrt(2 * np.pi)) * np.exp( - (x-mu)**2  / (2*sigma**2))

x1 = np.arange(-1, 1, 0.02)
x2 = np.arange(0.4, 1.6, 0.02)
ws1 =  CreateWorkspace(UnitX="1/q", DataX=x1, DataY=gaussian(x1[:-1], 0, 0.1))
ws2 =  CreateWorkspace(UnitX="1/q", DataX=x2, DataY=gaussian(x2[:-1], 1, 0.05))

stitched, scale = Stitch1D(LHSWorkspace=ws1, RHSWorkspace=ws2, Params=0.02)
print scale

The script above produces two peaks and attempts to stitch them together. Because the overlap is in a range where all values are 0, the second peak does not appear in the final stitched workspace. Adding +1 to every y value produces a result more like what I would expect.

[Owen Arnold]

refs #9549. Regression test demonstrates failure.

Changeset: 9aa1a3994f3e51737e2ed2faae8459e31858849d

[Owen Arnold]

refs #9549. Regression test fixed.

Changeset: 45cc409d47a5a5a65a1c37bce59c86f7e251b9f2

[Owen Arnold]

refs #9549. Regression test demonstrates failure.

Changeset: 9aa1a3994f3e51737e2ed2faae8459e31858849d

[Owen Arnold]

refs #9549. Regression test fixed.

Changeset: 45cc409d47a5a5a65a1c37bce59c86f7e251b9f2

[Owen Arnold]

refs #9549. Extend solution to scaling RHSWS

Changeset: 443ef2d14369917a905306db65832cd901f00b08

[Owen Arnold]

refs #9549. Extend tests for multithreaded zero search

Changeset: 83a9faed401bdf4df5e8a3a7e31dcbddbabef745

[Owen Arnold]

refs #9549. Protect against bad scale factors.

Changeset: a197ebc34f025f141852233e797347adc84a860d

[Owen Arnold]

refs #9549. Change warning conditions

Changeset: cfcb678c2bb4d7c73e0707ac32f8f158a43af435

[Owen Arnold]

refs #9549. Unused var removed.

Changeset: 25ad45f4e4df54e0be36c9cfabda63c94325b20e

[Martyn Gigg]

Fix input scalefactor value for case when manual value is not used.

The algorithm now requires the scalefactor to be greater than zero so even if it is not used the value must still satisfy the validator. Refs #9549

Changeset: 1a0da074ca3d64b16b234aae8066e3f822b02854

[Samuel Jackson]

Merge remote-tracking branch 'origin/feature/9549_stitch1d_defects'

Full changeset: ​0b43f3970e59d7253502392c99e07a9e38e78405

[Samuel Jackson]

Changes appear to have fixed the issue. Also tried with using a manual scale factor on scaling left and right hand workspaces. The results were also both fine. Code review looks good too.

[Stuart Campbell]

This ticket has been transferred to github issue ​10392