think about it.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/14
• Originally Reported By:
• Originally Created At: 2013-10-25T00:26:36.651

Currently differential operators are only tested in 3D on regular meshes. Please add test cases for 2D and tensor meshes with different (random) mesh sizes.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/1
• Originally Reported By: Lars Ruthotto
• Originally Created At: 2013-07-23T00:17:34.437

This will make the project backwards compatable

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/31
• Originally Reported By:
• Originally Created At: 2013-11-08T20:22:31.698

Survey I think fits what we are talking about more than a 'Data'. It is confusing to call this part of the framework 'Data'. We do a Survey, which has information about Tx, Rx, and has dobs, we then pair a problem and survey together to simulate dpred.

This naming convention is also confusing when we start having more complicated data-structures than just vectors, for example if you have many freqs or transmitters, and a ndarray just doesn't cut it anymore.

Talked about this with @sgkang and @lheagy today. @dwfmarchant do you agree? @lacmajedrez?

Inversion framework needs a good way of switching between betas.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/10
• Originally Reported By:
• Originally Created At: 2013-10-25T00:13:28.848

Put these in sputils.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/29
• Originally Reported By:
• Originally Created At: 2013-11-06T19:42:34.226

Dealing with bound constraints in the optimize problem.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/12
• Originally Reported By:
• Originally Created At: 2013-10-25T00:16:55.209

It looks like the properties BaseMesh.n and BashMesh.nCv do the same things. We should get rid of one of them.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/34
• Originally Reported By: David Marchant
• Originally Created At: 2013-11-26T20:09:56.875

Mesh creation doesn't work with integers in the hx, hy, hz, vectors.

This code works:

#!python

hx = r_[1.,2.,3.]
hy = r_[1.,2.,3.]
hz = r_[1.,2.,3.]
m = SimPEG.mesh.TensorMesh((hx,hy,hz))
spy(m.cellGrad)

This code doesn't:

#!python

hx = r_[1,2,3]
hy = r_[1,2,3]
hz = r_[1,2,3]
m = SimPEG.mesh.TensorMesh((hx,hy,hz))
spy(m.cellGrad)

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/9
• Originally Reported By: David Marchant
• Originally Created At: 2013-10-23T03:58:07.209

Currently differential operators are only tested in 3D on regular meshes. Please add test cases for 2D and tensor meshes with different (random) mesh sizes.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/1
• Originally Reported By: Lars Ruthotto
• Originally Created At: 2013-07-23T00:17:34.437

These should be implemented in fortran or something with a light python wrapper.

put these in SimPEG.Solver

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/16
• Originally Reported By:
• Originally Created At: 2013-10-25T01:11:26.841

aveCC2F should take into account the length scale. This can be tested by using a randomTensorMesh.

The other averaging matrices should be tested more rigorously as well.

area calculation (i.e. 'volume' in 2D) is not correct if the cell is concave.

Currently averaging the four parallelograms, and I think this works out to give the area of the convex hull of the cell.

Need to write test cases and explore this issue.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/3
• Originally Reported By:
• Originally Created At: 2013-08-01T17:33:02.126

Any thoughts as to why this might be happening? I can't get any other information about the crash from my terminal.

This should test derivatives of all parts of your problem. Makes writing and testing code easier for research applications.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/13
• Originally Reported By:
• Originally Created At: 2013-10-25T00:20:37.324

Currently differential operators are only tested in 3D on regular meshes. Please add test cases for 2D and tensor meshes with different (random) mesh sizes.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/1
• Originally Reported By: Lars Ruthotto
• Originally Created At: 2013-07-23T00:17:34.437

These should really be returning scalars.

Maybe notation like nFx and nFxv similar to nC and nCv.

This reads, number of faces in the X direction as a vector.

Thoughts?

I know there was a menu around here somewhere. If someone has one, can we scan it and attach it to the Wiki? It would make life easier to change it up when ordering sushi (unless we like the same stuff over and over again).

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/2
• Originally Reported By: Kris Davis
• Originally Created At: 2013-07-23T19:36:53.868

This should live in the mesh class. Or be inherited by each mesh class.

should provide an interpolation matrix that can take data (on cc, n, f, or e) and return the interpolated value.

This is needed for generating data (dpred) from the fields etc.

Implementation should have a few test cases as well.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/21
• Originally Reported By:
• Originally Created At: 2013-11-01T22:56:16.751

make
from SimPEG import *
work!

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/32
• Originally Reported By:
• Originally Created At: 2013-11-08T23:38:29.642

Make a base class for a time domain problem.

Some things to keep in mind:

timeStep = [1e-4, (1e-5, 50), 10]
assert getDt(30)==1e-5

use the 1D mesh code, superclass, and get rid of some methods?

checkpointing?

Grids should be X,Y,Z again

Interp locations should be Fx, Ey etc.

Interp code should also call interputils.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/24
• Originally Reported By:
• Originally Created At: 2013-11-05T21:08:00.878

Is there a difference?

This is a conceptual problem that needs to be addressed.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/7
• Originally Reported By:
• Originally Created At: 2013-10-19T02:55:43.320

Start with ave and then define where they go from and to:

For example:

#!python

M.aveF2CC
M.aveN2CC
M.aveCC2F

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/6
• Originally Reported By:
• Originally Created At: 2013-10-18T23:31:24.317

add it

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/27
• Originally Reported By:
• Originally Created At: 2013-11-06T19:27:56.081

add :

approxJ
approxJt

to the problem class, and update the inversion class so that these are called if available in the inversion.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/15
• Originally Reported By:
• Originally Created At: 2013-10-25T01:08:46.943

This should take all the nodes and average to the centers.

I think the current one averages to edges.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/28
• Originally Reported By:
• Originally Created At: 2013-11-06T19:41:07.765

This will make handling errors more manageable and scalable..

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/5
• Originally Reported By:
• Originally Created At: 2013-10-18T20:13:24.164

I ran into an issue in the scipy.linalg.cg algorithm.

The code does not allow you to run two CG loops at once. For example, if you want to use an iterative preconditioner (perhaps solving regularization WtW x = -g) inside inexact gauss newton. The code breaks, because scipy doesn't let you use two CG algorithms at the same time.

https://github.com/scipy/scipy/blob/master/scipy/sparse/linalg/isolve/iterative.py#L82

This might be a big issue if we are going to large-scale simulations where we cannot directly solve anything inside our inexact gauss newton, and have to use CG in a lot of other places while inside a big CG iteration.

Also, it might be bad if we need to parallelize a lot of CG algorithms at once?

Has any one else run into these problems? Should we be using another package for our iterative solvers?

Thoughts?

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/33
• Originally Reported By:
• Originally Created At: 2013-11-20T21:58:39.372

It would be good to have testing of the Cyl1D mesh

should follow similar structure as test_TensorMesh.py

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/8
• Originally Reported By:
• Originally Created At: 2013-10-22T23:36:29.442

This example is bit annoying for me.

Current example
h1 = ( (numPad, sizeStart [, increaseFactor]), (numCore, sizeCore), (numPad, sizeStart [, increaseFactor]) )

We may just say
in 1D
hx = Utils.meshTensor( (numPad, sizeStart,increaseFactor), (numCore, sizeCore), (numPad, sizeStart,increaseFactor) )

2D and 3D .. and so on

Write the communication using pubsub, this will allow you to hook into the communication from any other point in your code.

Should be used in Inversion and Optimize

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/11
• Originally Reported By:
• Originally Created At: 2013-10-25T00:16:04.496

All the visualization should live in a different folder.

e.g.

SimPEG.vis.mpl
SimPEG.vis.vtk

Could help the maintenance of the visualization code.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/22
• Originally Reported By:
• Originally Created At: 2013-11-01T22:58:17.179

Weak form, and test order of convergence.

Inhomogeneous Dirichlet, Neumann.

Get octree mesh implemented.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/19
• Originally Reported By:
• Originally Created At: 2013-11-01T22:50:55.900

Be able to create a figure of a 3D field sliced in various diections.

Z-slices, equally spaced, edges, faces, cell centers

overlay mesh

quiver plots

real, imag, amplitude

downsampling

change axes, get rid of padding

(make them into videos!!)

Rename nF to nFv and make nF return sum(nFv)

nE needs to be updated as well.

Tests and internal code need to be updated.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/25
• Originally Reported By:
• Originally Created At: 2013-11-05T21:19:13.068

h = [4,np.r_[1,2], np.random.rand(2)]
for dim in [2,3]:
    M = Mesh.TensorMesh(h[:dim])
    Mf = M.getFaceInnerProduct()
    Af = M.getFaceMass()
    print np.linalg.norm(Mf.diagonal()/dim - Af.diagonal())
    Me = M.getEdgeInnerProduct()
    Ae = M.getEdgeMass()
    print np.linalg.norm(Me.diagonal()/dim - Ae.diagonal())

1.30185189291e-16
1.30185189291e-16
1.95397176322e-17
1.9664450708e-17

It looks like the faceMass is 2 or 3 times the getFaceMass, but otherwise these methods are identical.

It would be good to remove one of the methods if they are indeed duplicates.
I would remove the Innerproduct code that is user facing, and make Mass code point to that instead, as it works for anisotropic properties.

Get the projected gradient code working for bound constraint problems.

perhaps add hooks for projection etc. and active constraints into the main Minimize class.

Reference the algorithm? And provide tests and documentation.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/20
• Originally Reported By:
• Originally Created At: 2013-11-01T22:53:17.867

Have the option to write to an hdf5 file

Be able to save a full optimization routine to disk.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/30
• Originally Reported By:
• Originally Created At: 2013-11-08T19:55:00.507

Dave and I were talking about moving this project to Github.

There are better source management tools, and a more involved community.

Thoughts?

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/35
• Originally Reported By:
• Originally Created At: 2013-11-26T20:10:12.688

I am having some problems with the file directory and importing things.

This is especially the case when importing something of the same name as the file:

#!python

from SimPEG.forward.LinearProblem import example  #  Works!
import SimPEG.forward.LinearProblem as LinearProblem
LinearProblem.example   #  does not exist

This is because LinearProblem is a Class inside the file called LinearProblem.

Things to think about:

• I think we should avoid having file names and the classes inside named the same thing
• This may mean that we have to shuffle some things around or combine files
• We need to figure out how to put all of the different forward problems in the directory.

Thoughts?
Does anyone have experience setting up the structure of python directories? and knows what actually goes in the init.py file?

I am kinda lost.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/26
• Originally Reported By:
• Originally Created At: 2013-11-06T03:15:52.140

Need a similar interpolation scheme for the logically orthogonal mesh.

Inverse distance weighting? Delaunay Triangulation?

This will be expensive as the mesh gets big, and in 3D.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/23
• Originally Reported By:
• Originally Created At: 2013-11-05T19:41:11.645

Wasted way to much of @sgkang time today..sorry.

There is a serious issue multiplying matrices time vectors in numpy and the action of numpy matrices times numpy vectors.

Consider the following

#!python

from SimPEG import utils as ut
import numpy as np
import matplotlib.pyplot as plt
from SimPEG import TensorMesh
from scipy import sparse as sp
from SimPEG import utils as ut
import scipy.sparse.linalg.dsolve as dsl

generate a sparse matrix A and a dense matrix C

#!python

h = [np.array([1,1,1,1]),np.array([1,1,1,1]),np.array([1,1,1,1])]
M = TensorMesh(h)
G = M.nodalGrad
I = sp.eye(125)
A = G.T*G
A = A+I
C = np.matrix(np.random.randn(4,125))
b = np.random.randn(125)

can do that

#!python

z = A*b

cannot do that

#!python

y = C*b

If on the other hand we set

#!python

b = mkvc(b,2)

then

#!python

b = ut.mkvc(b,2)

can do that

#!python

z = A*b

cannot do that

#!python

y = C*b

Then it all works

Now, we still need to be careful as

#!python

print np.shape(b.T*b), np.shape(np.inner(b,b)), np.shape(np.outer(b,b))

all gives 125 times 125 arrays

sigh ...

We need to come up with a solution to this!

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/4
• Originally Reported By: Eldad Haber
• Originally Created At: 2013-08-03T01:07:28.018

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/17
• Originally Reported By:
• Originally Created At: 2013-10-25T01:37:49.813

There is a bug in interpolation in TensorMesh if x0 is set to <0
and if the interpolated point is on the boundary of the domain.

For Example:

mesh = Mesh.TensorMesh([4, 4],x0=[0, 0])
spy(mesh.getInterpolationMat([[0,1]],'CC'))

These need to be documented with test cases.

These should be a call to another functions that create the RHS based on source locations and mesh.

• Bitbucket: https://bitbucket.org/rcockett/simpeg/issue/18
• Originally Reported By:
• Originally Created At: 2013-10-31T19:53:15.899

Want starting and stopping h values

clean up the headers

take function for first value