Skip to content

Transitioning IGM from v2 to v3: Guideline

In short

This is a major structural update, and we are starting a new versioning series with IGM 3.0.0. Below are the main changes and improvements you can expect:

  • Hydra integration (game changer!): The most important update is the integration of the Hydra library for handling parameters. This allows you to run large multi-ensemble simulations (poss on multiple GPUs) from a single command line, with full traceability and reproducibility. No more manually managing complex parameter files—Hydra does it for you, and does it extremely well. This alone is a compelling reason to switch to IGM3!

  • YAML-based parameters with hierarchical structure: All parameters are now defined in YAML (instead of JSON), offering a more powerful and flexible configuration format. Especially, parameters are now organized hierarchically, making it much easier to manage large configurations. This comes with renamed parameters, but don’t worry—we provide a tool to convert old JSON files to the new YAML format, along with a conversion table.

  • New documentation website: A freshly built documentation website now automatically collects all parameter definitions, along with their default values, descriptions, and units.

  • Clean folder structure: The working directory is now properly organized into subfolders for parameters, data, user modules, and output.

  • Improved code readability: Key modules like iceflow and data_assimilation (formerly optimize) have been split into sub-files to improve readability, maintainability, and customization.

Here are the important new links:

  • The pre-release is on the usual https://github.com/instructed-glacier-model/igm , but you need to checkout the feature/hydra branch.

  • The new documentation website is here : https://instructed-glacier-model.github.io/igm-doc/

  • The separate repo containing examples : https://github.com/instructed-glacier-model/igm-examples

  • The repo with the current technical paper re-shaped for IGM 3.0.0 : https://github.com/instructed-glacier-model/igm-paper

In more details (for users)

Parameter handling

hydra library replaces parser, and YAML file

Parameters previously managed with the parser library are now handled using the hydra library. The parameter file, which was formerly a JSON file (params.json), has been transitioned to a YAML file (experiment/params.yaml) adhering to the YAML standard. This shift to hydra has significantly simplified the core IGM code. Running IGM now involves executing commands like:

 igm_run +experiment=params

where parameters can be changed within the params.yaml file, or overridden directly in the command line. For example:

igm_run +experiment=params processes.time.start=1900 processes.time.end=2100

From a user perspective, migrating to IGM 3 essentially involves converting the former parameter file params.json into the new experiment/params.yaml format. To facilitate this transition, we provide a utility script named json_to_yaml.py (in the root of IGM repo) that automates the conversion process (you may have to adjust manually, check it afterwards!).

Parameter Naming Changes, and hierarchical construction

Parameter names have been slightly modified. Previously, all parameters included a prefix like iflo to indicate they were associated with the iceflow module. Now, parameters are organized hierarchically by attributes, making the prefix redundant and therefore removed. For example: time_start is now accessible as processes.time.start.

In the iceflow module, parameters have been significantly reorganized to follow a hierarchical structure. To assist with the transition, we provide a script named json_to_yaml.py, and correspondence table to help convert old JSON files into the new YAML format. Also, IGM now raises an error if a parameter contains a typo, and suggests parameters to pick.

Example of new parameter file

Here is what the new parameter file looks like:

# @package _global_

core:
  url_data: "https://www.dropbox.com/scl/fo/8ixpy27i67s04bp7uixoq/h?rlkey=0ye7rd4zkcqfhvzx7suunw3bk&dl=0"

defaults:
  - override /inputs: [local]
  - override /processes: [smb_simple, iceflow, time, thk]
  - override /outputs: [local, plot2d]

inputs:
  local:
    input_file: input.nc

processes:
  iceflow:
    physics:
      init_slidingco: 0.0595
  time:
    start: 1880.0
    end: 2020.0
    save: 5.0 

outputs:
  plot2d:
    live: true

The above file will be read by Hydra in IGM and has the following structure:

  • core: Includes all parameters specific to the core IGM run, such as logging, downloading data prior to the run, GPU-related settings, etc.
  • defaults: Lists the input, process, and output modules to be used.
  • inputs: Contains parameters to override the defaults for input modules.
  • processes: Contains parameters to override the defaults for process modules.
  • outputs: Contains parameters to override the defaults for output modules.

Note: The former module types preprocess, process, and postprocess have been renamed to input, modules, and output, respectively.

Running multiple runs

A great advantage of Hydra is the possibility it provides for running ensemble simulations. For example, the following command sequentially runs two simulations with two different sets of parameters:

igm_run -m +experiment=params processes.time.end=2080,2110

Alternatively, if you supply two parameter files (params1 and params2), you can execute:

igm_run -m +experiment=params1,params2

You can also perform a grid search over multiple parameters. For instance, the following command runs simulations for a 3x2 parameter grid:

igm_run -m +experiment=params processes.time.start=1900,1910,1920 processes.time.end=2080,2110

In such cases, the output folder will be named multirun instead of output.

New structure of working folder

The working folder is structured as follows:

  • experiment: Contains the parameter files.
  • data: Stores input data, if any.
  • user: Contains user-defined/custom Python functions or modules.
  • output or multirun: Contains the results of the IGM runs.

The folder structure looks like this:

├── experiment
│   └── params.yaml
├── data
│   ├── ...
├── user
│   ├── code
│   │   └── processes
│   │       └── mymodule.py
│   └── conf
│       └── processes
│           └── mymodule.yaml
└── output
  ├── 2025-03-06
  │   └── 15-43-37
  │   └── 15-44-07
  │       └── output.nc
  │       └── ...

Integration and Exclusion of Former Modules

  • Splitting of iceflow modules into data_assimilation and pretraining: The iceflow module has been split into two dedicated modules: data_assimilation (formerly optimize) and pretraining. This change was made to improve code organization and maintainability, as the iceflow module had grown too large. Previously, data_assimilation and pretraining were separate but had dependency issues with iceflow. With the new structure, both modules remain dependent on iceflow, but these dependencies are now handled seamlessly. When using data_assimilation or pretraining, ensure that the iceflow module is also included in your configuration. The order of inclusion does not matter.

  • The modules anim_mayavi, anim_plotly, and anim_video have been externalized from IGM and moved to utils. These modules were purely for postprocessing and were executed at the very end. To simplify the core structure, they were externalized.

  • Modules such as print_comp and print_info have been integrated into the core functionality of IGM.

  • The write_particles functionality is now integrated into the particule module.

New local module merging netcdf and tif

A new I/O module, local, has been introduced to replace the load_XXX and write_XXX modules. The local module leverages the xarray library, which is more powerful and supports loading both NetCDF (.nc) and GeoTIFF (.tif) files.

oggm_shop requires coupling with load_ncdf or local

The oggm_shop module now exclusively handles downloading data (e.g., RGIXXXX folders) using OGGM and converting it into a NetCDF file (input.nc) that adheres to IGM's naming conventions. However, it no longer performs the task of loading this data into IGM. To process the downloaded data, you must pair oggm_shop with either the load_ncdf or local modules.

For example, if you use oggm_shop, you must include load_ncdf or local as additional inputs modules in your configuration.

oggm_shop

The oggm_shop module has been substantially re-organized.

Parameters RGI_version and RGI_product were removed since there are directly deduced from the RGI ID

utils re-structuring

The former utils.py file has been re-structured forlder-wise for clarity, this means that some import like

from ..utils import getmag

must be changed to

from igm.utils.math.getmag import getmag

Default Parameter Value Changes (Upcoming Release)

In the upcoming release, several default parameter values will be updated to improve performance and usability. These changes are not applied yet but will be included in the next version:

  • retrain_freq (current default: 10): The current default retrains the ice flow CNN every 10 time steps. This frequency has been found insufficient in many cases. The next release will increase the retraining frequency to every 3 time steps.

  • smooth_anisotropy_factor (current default: 0.2): The current value of 0.2 for anisotropic smoothing has been observed to cause chessboard effects in ice thickness. To address this, the next release will safely deactivate anisotropic smoothing by setting this parameter to 1.0.

  • convexity_weight (current default: unknown): The current convexity weight, which is particularly useful in the absence of data, has been found to be highly empirical and non-intuitive. As a result, it will be deactivated in the next release by setting its value to 0.

  • fix_opti_normalization_issue (current default: false): A normalization issue has been identified in the current cost function for data assimilation, where some terms use sums while others use means. This inconsistency has been compensated for by adjusting regularization parameters. The next release will address this issue by setting this parameter to true, ensuring consistent use of means throughout. Consequently, users will need to significantly increase the regularization parameters (approximately \(10^3\) for thickness and \(10^{10}\) for sliding coefficients).

  • log_slidingco (current default: false): In the next release, the optimization will handle the square root of slidingco (scaled) instead of slidingco directly. This approach ensures that slidingco remains positive without explicitly enforcing positivity. Note that this change will affect the scaling of slidingco (approximately \(10^{-6}\)) when this option is set to true.

  • `lr_init`` (current default: 0.0001): In the next release, it will be reduced to 0.001 to make the training more agressive.

In more details (for developpers)

Hierachical and separated parameters and code

In IGM source code, the parameters and the code are now hierarchically organized. Default parameters are stored in the igm/igm/conf folder, which is separate from the code located in the igm/igm folder. This separation ensures a clean organization of configuration files and source code. Parameters are grouped into logical categories and subcategories within the YAML files. This structure mirrors the organization of the modules and processes in IGM. For example, the folder structure for default parameters looks like this:

igm/igm/conf
├── inputs
│   └── local.yaml
│   └── ...
├── processes
│   └── iceflow.yaml
│   └── ...
└── outputs
  └── plot2d.yaml
  └── ...

In the code, all parameters are accessible through the object cfg. For example, cfg.processes.enthalpy.ref_temp refers to a parameter associated with the enthalpy processes module.

On the other hand the structure for the code looks very similar:

igm/igm
├── inputs
│   └── local
│   └── ...
├── processes
│   └── iceflow
│   └── avalanche
└── outputs
  └── plot2d
  └── ...

Custom modules (now called "user")

User modules are very useful when customizing applications. They can be used to tailor input, process, or output methods. To create such a user module, you need to create (or update) the user folder located at the root of your working directory, ensuring the following hierarchy is respected:

└── user
    ├── code
    │   └── inputs 
    │   │   └── my_inputs_module
    │   │       └── my_inputs_module.py 
    │   └── processes 
    │   │   └── my_processes_module
    │   │       └── my_processes_module.py 
    │   └── outputs 
    │       └── my_outputs_module
    │           └── my_outputs_module.py 
    └── conf
        └── inputs
            └── my_inputs_module.yaml 
        └── processes
            └── my_processes_module.yaml 
        └── outputs
            └── my_outputs_module.yaml

where my_processes_module.py has the following structure (and requires the definitions of functions initialize, update, and finalize):

def initialize(cfg,state):
  ... 

def update(cfg,state):
  cfg.processes.clim_aletsch.time_resolution
  ... 

def finalize(cfg,state):
  pass

Note that my_inputs_module.py and my_outputs_module.py only require a single function called run (with the addition of an initialize function for output), defined in the same way as the three functions above.

Parameter files located in conf/inputs, conf/processes, and conf/outputs look like 

update_freq: 1
time_resolution: 365
or in case there is no parameter (the file must exist as follows even if no parameter are defined): 
null

It is important to note that user modules take precedence over official modules. If a user module shares the same name as an official module, the user module will override the official one, and the official module will be ignored.

How to transition your own user modules?

  • Parameter changes: Update your user module code to replace each parameter reference (e.g., params.iflo_exp_glen) with the new naming convention (e.g., cfg.processes.iceflow.physics.exp_glen) as outlined in the parameter change tables below.

  • Function updates: Modify the function argument for initialize, update, and finalize to accept cfg as an argument instead of params. The state argument remains unchanged.

  • Own parameter: Remove the params(parser) function that previously defined your parameter list. Instead, transfer your parameter definitions into the appropriate YAML files within the conf folder, following the structure described in the "Custom modules (now called 'user')" section.

Complex modules made more readables, and more customizable

Key modules like iceflow and data_assimilation (formerly optimize) have been split into sub-files to improve readability, maintainability, and customization. Sometimes, you may need to modify an existing built-in module, e.g. to test a new feature in the iceflow emulator/solver, or new cost in the data assimilation. This can be achieved by creating a user module that overrides the built-in functionality. Check at the page on user modules in the documentation.

Parameter changes

Parameter name change table for iceflow module

Formerly New name
iflo_type iceflow.method
iflo_force_max_velbar iceflow.force_max_velbar
iflo_gravity_cst iceflow.physics.gravity_cst
iflo_ice_density iceflow.physics.ice_density
iflo_init_slidingco iceflow.physics.init_slidingco
iflo_init_arrhenius iceflow.physics.init_arrhenius
iflo_enhancement_factor iceflow.physics.enhancement_factor
iflo_exp_glen iceflow.physics.exp_glen
iflo_exp_weertman iceflow.physics.exp_weertman
iflo_regu_glen iceflow.physics.regu_glen
iflo_regu_weertman iceflow.physics.regu_weertman
iflo_new_friction_param REMOVED
iflo_dim_arrhenius iceflow.physics.dim_arrhenius
iflo_regu REMOVED
iflo_thr_ice_thk iceflow.physics.thr_ice_thk
iflo_min_sr iceflow.physics.min_sr
iflo_max_sr iceflow.physics.max_sr
iflo_force_negative_gravitational_energy iceflow.physics.force_negative_gravitational_energy
iflo_cf_eswn iceflow.physics.cf_eswn
iflo_cf_cond iceflow.physics.cf_cond
iflo_Nz iceflow.numerics.Nz
iflo_vert_spacing iceflow.numerics.vert_spacing
iflo_solve_step_size iceflow.solver.step_size
iflo_solve_nbitmax iceflow.solver.nbitmax
iflo_solve_stop_if_no_decrease iceflow.solver.stop_if_no_decrease
iflo_optimizer_solver iceflow.solver.optimizer
iflo_optimizer_lbfgs iceflow.solver.lbfgs
iflo_save_cost_solver iceflow.solver.save_cost
iflo_fieldin iceflow.emulator.fieldin
iflo_retrain_emulator_freq iceflow.emulator.retrain_freq
iflo_retrain_emulator_lr iceflow.emulator.lr
iflo_retrain_emulator_lr_init iceflow.emulator.lr_init
iflo_retrain_warm_up_it iceflow.emulator.warm_up_it
iflo_retrain_emulator_nbit_init iceflow.emulator.nbit_init
iflo_retrain_emulator_nbit iceflow.emulator.nbit
iflo_retrain_emulator_framesizemax iceflow.emulator.framesizemax
iflo_pretrained_emulator iceflow.emulator.pretrained
iflo_emulator iceflow.emulator.name
iflo_save_model iceflow.emulator.save_model
iflo_exclude_borders iceflow.emulator.exclude_borders
iflo_optimizer_emulator iceflow.emulator.optimizer
iflo_optimizer_emulator_clipnorm iceflow.emulator.optimizer_clipnorm
iflo_optimizer_emulator_epsilon iceflow.emulator.optimizer_epsilon
iflo_save_cost_emulator iceflow.emulator.save_cost
iflo_output_directory iceflow.emulator.output_directory
iflo_network iceflow.emulator.network.architecture
iflo_multiple_window_size iceflow.emulator.network.multiple_window_size
iflo_activation iceflow.emulator.network.activation
iflo_nb_layers iceflow.emulator.network.nb_layers
iflo_nb_blocks iceflow.emulator.network.nb_blocks
iflo_nb_out_filter iceflow.emulator.network.nb_out_filter
iflo_conv_ker_size iceflow.emulator.network.conv_ker_size
iflo_dropout_rate iceflow.emulator.network.dropout_rate
iflo_weight_initialization iceflow.emulator.network.weight_initialization

Parameter name change table for optimize (data_assimilation) module

Formerly New name
opti_control data_assimilation.control_list
opti_cost data_assimilation.cost_list
opti_nbitmin data_assimilation.optimization.nbitmin
opti_nbitmax data_assimilation.optimization.nbitmax
opti_step_size data_assimilation.optimization.step_size
opti_step_size_decay data_assimilation.optimization.step_size_decay
opti_init_zero_thk data_assimilation.optimization.init_zero_thk
opti_sole_mask data_assimilation.optimization.sole_mask
opti_retrain_iceflow_model data_assimilation.optimization.retrain_iceflow_model
opti_fix_opti_normalization_issue data_assimilation.optimization.fix_opti_normalization_issue
opti_regu_param_thk data_assimilation.regularization.thk
opti_regu_param_slidingco data_assimilation.regularization.slidingco
opti_regu_param_arrhenius data_assimilation.regularization.arrhenius
opti_regu_param_div data_assimilation.regularization.divflux
opti_smooth_anisotropy_factor data_assimilation.regularization.smooth_anisotropy_factor
opti_smooth_anisotropy_factor_sl data_assimilation.regularization.smooth_anisotropy_factor_sl
opti_convexity_weight data_assimilation.regularization.convexity_weight
opti_convexity_power data_assimilation.regularization.convexity_power
opti_to_regularize data_assimilation.regularization.to_regularize
opti_usurfobs_std data_assimilation.fitting.usurfobs_std
opti_velsurfobs_std data_assimilation.fitting.velsurfobs_std
opti_thkobs_std data_assimilation.fitting.thkobs_std
opti_divfluxobs_std data_assimilation.fitting.divfluxobs_std
opti_uniformize_thkobs data_assimilation.fitting.uniformize_thkobs
opti_include_low_speed_term data_assimilation.fitting.include_low_speed_term
opti_velsurfobs_thr data_assimilation.fitting.velsurfobs_thr
opti_log_slidingco data_assimilation.fitting.log_slidingco
opti_divflux_method data_assimilation.divflux.method
opti_force_zero_sum_divflux data_assimilation.divflux.force_zero_sum
opti_scaling_thk data_assimilation.scaling.thk
opti_scaling_usurf data_assimilation.scaling.usurf
opti_scaling_slidingco data_assimilation.scaling.slidingco
opti_scaling_arrhenius data_assimilation.scaling.arrhenius
opti_output_freq data_assimilation.output.freq
opti_plot2d_live data_assimilation.output.plot2d_live
opti_plot2d data_assimilation.output.plot2d
opti_save_result_in_ncdf data_assimilation.output.save_result_in_ncdf
opti_save_iterat_in_ncdf data_assimilation.output.save_iterat_in_ncdf
opti_editor_plot2d data_assimilation.output.editor_plot2d
opti_vars_to_save data_assimilation.output.vars_to_save
opti_infer_params data_assimilation.cook.infer_params
opti_tidewater_glacier data_assimilation.cook.tidewater_glacier
opti_vol_std data_assimilation.cook.vol_std