ERT inversion Chi² stuck from the start in Google Colab

[Elhoule]

I’m trying to run the official 2D ERT example using PyGIMLi inside Google Colab with Miniconda. Forward modelling works but inversion is stuck: dPhi=0% and jacobian() returns RMatrix: 0 x 0. I tried backend='cg' and installed pyamg.

1. I used these lines to install PyGimli on Google Colab:
   !pip uninstall -y numba tensorflow pytensor thinc
   !pip install pygimli tetgen

2. I Restart the session

3. I follow step-by-step the 2D ERT modelling and inversion example (https://www.pygimli.org/_examples_auto/3_ert/plot_01_ert_2d_mod_inv.html#sphx-glr-examples-auto-3-ert-plot-01-ert-2d-mod-inv-py)

4. inv = mgr.invert(lam=20, verbose=True)

#################

Returns

#################

INFO:pyGIMLi:Found 2 regions.
01/07/25 - 04:43:50 - pyGIMLi - INFO - Found 2 regions.
INFO:pyGIMLi:(ERTModelling) Region with smallest marker (1) set to background.
01/07/25 - 04:43:50 - pyGIMLi - INFO - (ERTModelling) Region with smallest marker (1) set to background.
INFO:pyGIMLi:Found 2 regions.
01/07/25 - 04:43:50 - pyGIMLi - INFO - Found 2 regions.
INFO:pyGIMLi:(ERTModelling) Region with smallest marker (1) set to background.
01/07/25 - 04:43:50 - pyGIMLi - INFO - (ERTModelling) Region with smallest marker (1) set to background.
INFO:pyGIMLi:Creating forward mesh from region infos.
01/07/25 - 04:43:50 - pyGIMLi - INFO - Creating forward mesh from region infos.
INFO:pyGIMLi:Creating refined mesh (H2) to solve forward task.
01/07/25 - 04:43:50 - pyGIMLi - INFO - Creating refined mesh (H2) to solve forward task.
INFO:pyGIMLi:Mesh for forward task: Mesh: Nodes: 987 Cells: 1824 Boundaries: 1442
01/07/25 - 04:43:50 - pyGIMLi - INFO - Mesh for forward task: Mesh: Nodes: 987 Cells: 1824 Boundaries: 1442
INFO:pyGIMLi:Use median(data values)=67.9439561191829
01/07/25 - 04:43:50 - pyGIMLi - INFO - Use median(data values)=67.9439561191829
INFO:pyGIMLi:Created startmodel from forward operator:287, min/max=67.943956/67.943956
01/07/25 - 04:43:50 - pyGIMLi - INFO - Created startmodel from forward operator:287, min/max=67.943956/67.943956
INFO:pyGIMLi:Starting inversion.
01/07/25 - 04:43:50 - pyGIMLi - INFO - Starting inversion.
fop: <pygimli.physics.ert.ertModelling.ERTModelling object at 0x78837c07d120>
Data transformation: Logarithmic LU transform, lower bound 0.0, upper bound 0.0
Model transformation: Logarithmic transform
min/max (data): 42.79/104
min/max (error): 1%/1.06%
min/max (start model): 67.94/67.94
--------------------------------------------------------------------------------
inv.iter 0 ... chi² =  506.26
--------------------------------------------------------------------------------
inv.iter 1 ... chi² =  506.26 (dPhi = 0.00%) lam: 20.0
--------------------------------------------------------------------------------
inv.iter 2 ... chi² =  506.26 (dPhi = 0.00%) lam: 20.0
--------------------------------------------------------------------------------
inv.iter 3 ... chi² =  506.26 (dPhi = 0.00%) lam: 20.0
################################################################################
#                 Abort criterion reached: dPhi = 0.0 (< 2.0%)                 #
################################################################################
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
[/tmp/ipython-input-2-676997242.py](https://localhost:8080/#) in <cell line: 0>()
      2 
      3 inv = mgr.invert(lam=20, verbose=True)
----> 4 np.testing.assert_approx_equal(mgr.inv.chi2(), 0.7, significant=1)

[/usr/local/lib/python3.11/dist-packages/numpy/testing/_private/utils.py](https://localhost:8080/#) in assert_approx_equal(actual, desired, significant, err_msg, verbose)
    732         pass
    733     if np.abs(sc_desired - sc_actual) >= np.power(10., -(significant - 1)):
--> 734         raise AssertionError(msg)
    735 
    736 

AssertionError: 
Items are not equal to 1 significant digits:
 ACTUAL: 506.2573131612945
 DESIRED: 0.7

If anyone have the same problem or a solution for it I would be happy to hear it ^^

Thanks!

[halbmy]

Indeed, I can confirm this bug. It runs perfectly locally in VScode or on the command line, but it does not in Google colab. We need to make sure it is not a problem of the pip installation.

[Boorhin]

Is there an update on this? Thanks

[halbmy]

As there was no new version, probably not. I guess it is about the mesh generation. The starting model size looks too small, so I am afraid the region marker is lost and the outer region is used for inversion. Have a look at the mesh and check whether such an error also occurs for other examples.

[Boorhin]

I have tested with my own data and defining a mesh myself. It fails on Google Collab. It works on my ipython interface

[Elhoule]

It seems like you can run it locally using Google Collab. To do so, I think you can use Docker and follow the steps presented here: [https://research.google.com/colaboratory/local-runtimes.html]

[Boorhin]

Sorry I ran it on ipython on Ubuntu 24.04 not local google collab