Previously Thrifty mode (where line search results from iteration I are used to replace the forward simulations of iteration I + 1) was not properly implemented into the NoiseInversion workflow. This PR improves that functionality and also adds some QoL features and fixes:
ChangeLog
Major
New parameter: optimize.step_len_min allows the User to set a minimum step length (alpha) as a percentage of the model values. This prevents line searches from trying to minimize misfit with very small model changes, which is a waste of computational resources.
Changed Behavior: Saved Forward arrays are now overwritten by new forward simulations, for the case when Thrifty line searches need access to forward arrays in subsequent adjoint simulations
QoL Improvement: Initial and True model file export (tied to parameter export_model) now checks whether files exist and will not overwrite (previously they always overwrote), meaning Solver.setup() is now a lot faster
Bugfix: Preprocess.Pyaflowa was missing a mkdir for adjoint source directory which was leading to FileNotFoundError
Examples: Changed starting step length in Ex1 to achieve successful line search. Ex2 does NOT finish iteration, will need to fix this
Minor
Workflow Thrifty checks are now inside a property that can be checked, rather than as a function call
Preprocess reads STATION metadata (locations) from it's Event's own 'DATA' directory (e.g., scratch/solver/001/DATA/STATIONS), rather than from the master station file (e.g., path_specfem_data/STATIONS), allowing for different STATION definitions for different events. Untested, might need a little more tooling to make that work.
New internal Workflow.Inversion attribute _was_thrifty keeps track of whether a previous iteration's finalization was done in a Thrifty manner, allowing the current iteration to skip forward simulations. WARNING This parameter is NOT saved to disk, so if the workflow breaks between finalization of iteration I and forward simulations of I+1, then forward simulations will be run like a normal inversion.
Cleaned up the aesthetic look of Submit and Debug commands
Removed step_len_max from Line Search instantiation because it was not used by the module
Refactored some optimization checks for step_len_min and step_len_max with better log messages and less redundant code
Removed path_specfem_data from Preprocess init because it is not used in the module
Removed some redundant checks from Preprocess (Default and Pyaflowa)
Removed STATION file reading from default preprocessing as the Inventory was not used
Previously
Thriftymode (where line search results from iteration I are used to replace the forward simulations of iteration I + 1) was not properly implemented into the NoiseInversion workflow. This PR improves that functionality and also adds some QoL features and fixes:ChangeLog
Major
optimize.step_len_minallows the User to set a minimum step length (alpha) as a percentage of the model values. This prevents line searches from trying to minimize misfit with very small model changes, which is a waste of computational resources.export_model) now checks whether files exist and will not overwrite (previously they always overwrote), meaning Solver.setup() is now a lot fastermkdirfor adjoint source directory which was leading to FileNotFoundErrorMinor
_was_thriftykeeps track of whether a previous iteration's finalization was done in a Thrifty manner, allowing the current iteration to skip forward simulations. WARNING This parameter is NOT saved to disk, so if the workflow breaks between finalization of iteration I and forward simulations of I+1, then forward simulations will be run like a normal inversion.step_len_maxfrom Line Search instantiation because it was not used by the modulestep_len_minandstep_len_maxwith better log messages and less redundant codepath_specfem_datafrom Preprocess init because it is not used in the module