Finished Bayesian tutorial #383

benRenard · 2025-02-10T16:25:44Z

No description provided.

Updated broken link

nghi-truyen

Merci Ben,
Ma première remarque est que la doc prend du temps pour compiler comme il y a pas mal de fonction d'optimisation dans ce tuto. Serait-il possible de rajouter le maxiter dans termination_crit de optimize_options pour limiter le temps ça prenne pour l'optim (je pense qu'on a pas forcément besoin d'une convergence dans un tuto); et aussi si possible utiliser le Lez au lieu de la Cance pour être encore plus rapide ?

asjeb

model0 -> model_0 (Truyen m'a corrigé sur des tuto à moi)

asjeb · 2025-02-11T11:50:35Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+3. it naturally handles multi-gauge calibration, with a built-in weighting of the information brought by each gauge. 
+
+This tutorial builds on the :ref:`tutorial on model calibration<user_guide.classical_uses.split_sample_temporal_validation>`: we recommend that you follow it first.


Suggested change

This tutorial builds on the :ref:`tutorial on model calibration<user_guide.classical_uses.split_sample_temporal_validation>`: we recommend that you follow it first.

This tutorial builds on the :ref:`<user_guide.classical_uses.fully_distributed_calibration>`: we recommend that you follow it first.

asjeb · 2025-02-11T11:51:51Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+	model0 = smash.Model(setup, mesh)
+
+We finally optimize this model using the standard, non-Bayesian approach described in the :ref:`model calibration tutorial<user_guide.classical_uses.split_sample_temporal_validation>`. Note that in addition to the four model parameters ``cp``, ``ct``, ``kexc`` and ``llr``, we also calibrate the initial state values (``hp`` and ``ht``) so that we work with 'good' initial state values in the remainder of this tutorial. In addition, a single gauge is used for calibration (the most downstream one, with code "V3524010").


Suggested change

We finally optimize this model using the standard, non-Bayesian approach described in the :ref:`model calibration tutorial<user_guide.classical_uses.split_sample_temporal_validation>`. Note that in addition to the four model parameters ``cp``, ``ct``, ``kexc`` and ``llr``, we also calibrate the initial state values (``hp`` and ``ht``) so that we work with 'good' initial state values in the remainder of this tutorial. In addition, a single gauge is used for calibration (the most downstream one, with code "V3524010").

We optimize the model using the standard approach described in the :ref:`<user_guide.classical_uses.fully_distributed_calibration>`. In addition to the four model primary parameters ``cp``, ``ct``, ``kexc`` and ``llr``, we also calibrate the initial state values (``hp`` and ``ht``) so that we work with 'good' initial state values in the remainder of this tutorial. A single gauge is considered for calibration (the most downstream one, with code "V3524010").

asjeb · 2025-02-11T11:53:06Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

-Bayesian Estimation Approach
-============================
+==============================
+Bayesian Estimation in `smash`


Suggested change

Bayesian Estimation in `smash`

Bayesian Estimation Approach

asjeb · 2025-02-11T12:04:25Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+.. ipython:: python
+
+	calibrated_parameters=["cp","ct","kexc","llr","hp","ht"] 


Suggested change

calibrated_parameters=["cp","ct","kexc","llr","hp","ht"]

calibrated_parameters = ["cp", "ct", "kexc", "llr"," hp", "ht"]

asjeb · 2025-02-11T12:07:24Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+	optimize_options={"parameters": calibrated_parameters}
+	cost_options={"gauge": calibration_gauges}
+	model0.optimize(optimize_options=optimize_options,cost_options=cost_options)


Suggested change

calibration_gauges=["V3524010"]

optimize_options={"parameters": calibrated_parameters}

cost_options={"gauge": calibration_gauges}

model0.optimize(optimize_options=optimize_options,cost_options=cost_options)

optimize_options={

"parameters": calibrated_parameters

}

calibration_gauges = ["V3524010"]

cost_options = {

"gauge": calibration_gaugesy

}

model_0.optimize(

optimize_options = optimize_options,

cost_options = cost_options

)

asjeb · 2025-02-11T15:13:22Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+	optimize_options={"parameters": calibrated_parameters}
+	cost_options={"gauge": calibration_gauges}
+	model0.optimize(optimize_options=optimize_options,cost_options=cost_options)


Suggested change

model0.optimize(optimize_options=optimize_options,cost_options=cost_options)

model_0.optimize(optimize_options=optimize_options,cost_options=cost_options)

asjeb · 2025-02-11T15:13:46Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+
+	optimize_options={"parameters": calibrated_parameters,"control_tfm":'keep'}
+	info=smash.optimize_control_info(model=model0,optimize_options=optimize_options)


Suggested change

info=smash.optimize_control_info(model=model0,optimize_options=optimize_options)

info=smash.optimize_control_info(model=model_0,optimize_options=optimize_options)

asjeb · 2025-02-11T15:14:24Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+	optimize_options={"parameters": calibrated_parameters}
+	cost_options={"gauge": calibration_gauges}
+	model=smash.bayesian_optimize(model=model0,optimize_options=optimize_options,cost_options=cost_options)


Suggested change

model=smash.bayesian_optimize(model=model0,optimize_options=optimize_options,cost_options=cost_options)

model=smash.bayesian_optimize(model=model_0,optimize_options=optimize_options,cost_options=cost_options)

asjeb · 2025-02-11T15:14:52Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+	optimize_options={"parameters": calibrated_parameters}
+	cost_options={"gauge": calibration_gauges,"control_prior":priors}
+	model=smash.bayesian_optimize(model=model0,optimize_options=optimize_options,cost_options=cost_options)


Suggested change

model=smash.bayesian_optimize(model=model0,optimize_options=optimize_options,cost_options=cost_options)

model=smash.bayesian_optimize(model=model_0,optimize_options=optimize_options,cost_options=cost_options)

asjeb · 2025-02-11T15:15:20Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+	optimize_options={"parameters": calibrated_parameters}
+	cost_options={"gauge": calibration_gauges}
+	model=smash.bayesian_optimize(model=model0,optimize_options=optimize_options,cost_options=cost_options)


Suggested change

model=smash.bayesian_optimize(model=model0,optimize_options=optimize_options,cost_options=cost_options)

model=smash.bayesian_optimize(model=model_0,optimize_options=optimize_options,cost_options=cost_options)

asjeb · 2025-02-11T15:29:45Z

Je t'ai proposé des modif directement à commiter depuis le git. Mais je pense que pour la prochaine passe, je ferai du pull/push.
La compilation est longue j'aurais tendance à proposer que tu passes les cellules python en block, comme utiliser dans le contributor_guide/development_process_details.rst, ensuite tu peux référencer les images comme dans math_num/forward_structure.rst que tu mets dans le _static ; ou faire un max_iter comme suggérer par Truyen.
Convention anglo-sax pour moi avec des espaces autour des =, pas obligatoire mais ça rend le code plus lisible.
Je t'ai espacé certain de tes appels de fonctions et tes dico, dis-moi ce que tu en penses.

benRenard · 2025-02-11T17:46:00Z

Je ne suis pas fan de rajouter des "max_iter" pour aller plus vite, ca pollue inutilement le tuto je trouve. Je prefere soit utiliser le Lez (effectivement ca va plus vite, mais bon j'ai ecrit tout le tuto sur la base de la Cance donc c'est pas trivial), soit utiliser les cellules "block". C'est quoi un temps de compil "acceptable" ?

Pour les conventions, je vous fais confiance, il faut privilegier celles de smash.

nghi-truyen · 2025-02-13T12:16:27Z

@benRenard , @asjeb , j'ai crée un PR #385 pour simplifier la génération des outputs lorsqu’on utilise code-block au lieu de ipython dans les tutoriels. Vous pouvez y jeter un œil, en particulier sur le fichier contributor_guide (pas besoin d’entrer dans les détails du code).

pag13

Très intéressant ! J'ai mis des suggestions mineures de forme.

pag13 · 2025-02-14T12:18:53Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+1. it allows using prior information on parameters when available;
+2. it explicitly recognizes the uncertainties affecting both the model-simulated discharge (structural uncertainty) and the discharge data used to calibrate the model (observation uncertainty);
+3. it naturally handles multi-gauge calibration, with a built-in weighting of the information brought by each gauge. 


expliciter encore "built-in", intrinsec/automatic weighting property?

pag13 · 2025-02-14T12:20:01Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+	model0 = smash.Model(setup, mesh)
+
+We finally optimize this model using the standard, non-Bayesian approach described in the :ref:`model calibration tutorial<user_guide.classical_uses.split_sample_temporal_validation>`. Note that in addition to the four model parameters ``cp``, ``ct``, ``kexc`` and ``llr``, we also calibrate the initial state values (``hp`` and ``ht``) so that we work with 'good' initial state values in the remainder of this tutorial. In addition, a single gauge is used for calibration (the most downstream one, with code "V3524010").


bonne idée de mq il est aussi possible d'optimiser les états

pag13 · 2025-02-14T12:43:31Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+		plt.legend();
+		plt.title(title);
+		return(plt)


Modifier verbose out pour les appels faits ensuite de cette fonction doplot qui produit dans la doc html générée des messages:
OUTXX: <module 'matplotlib.pyplot' from '/home/pagarambois/anaconda3/envs/smash-dev/lib/python3.12/site-packages/matplotlib/pyplot.py'>

pag13 · 2025-02-14T12:49:05Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+
+The figure below compares the observed and the simulated discharge time series at gauge "V3517010" and indeed shows a quite poor fit, leading to a rather high uncertainty.
+


And unrealisticly high parameters values?
Sur la doc que j'ai compilé les incertitudes sont rather high en effet (10e10), c'est cela ?
Les paramètres obtenus sont très grandds aussi :
cp-0 estimated at : 1123919500.0000
ct-0 estimated at : 1127735353.0000
kexc-0 estimated at : 995713886.0000
llr-0 estimated at : 1110648003.0000
sg0-V3524010 estimated at : 1044549471.0000
sg0-V3515010 estimated at : 897988541.0000
sg0-V3517010 estimated at : 897988541.0000
sg1-V3524010 estimated at : 1045220557.0000
sg1-V3515010 estimated at : 1050589267.0000
sg1-V3517010 estimated at : 1052602533.0000

Ajouter un commentaire sur cela?

pag13 · 2025-02-14T12:53:58Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

+		print(f"{info['name'][i]} estimated at : {info['x'][i]:.4f}")
+
+The figure below compares the observed and the simulated discharge time series at gauge "V3517010" and indeed shows a quite poor fit, leading to a rather high uncertainty.


Le nom de jauge pourrait être ajouté dans le titre du graphe mais très bien en l'état déjà.

nghi-truyen · 2025-02-17T14:57:49Z

doc/source/user_guide/in_depth/bayesian_estimation_approach.rst

@@ -1,8 +1,204 @@
 .. _user_guide.in_depth.bayesian_estimation_approach:


Reminder/note: we can use smash.default_optimize_options to show the default optimization options..;
mention that Bayes optimize is currently not available for ANN mapping..

Finished Bayesian tutorial

76951a2

benRenard requested review from pag13, nghi-truyen and asjeb February 10, 2025 16:26

Update bayesian_estimation_approach.rst

f74d59d

Updated broken link

nghi-truyen reviewed Feb 11, 2025

View reviewed changes

asjeb reviewed Feb 11, 2025

View reviewed changes

pag13 reviewed Feb 14, 2025

View reviewed changes

nghi-truyen reviewed Feb 17, 2025

View reviewed changes

asjeb linked an issue Feb 18, 2025 that may be closed by this pull request

Add a 'Bayesian estimation' case study #128

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Finished Bayesian tutorial #383

Finished Bayesian tutorial #383

benRenard commented Feb 10, 2025

nghi-truyen left a comment •

edited

Loading

asjeb left a comment

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb Feb 11, 2025

asjeb commented Feb 11, 2025 •

edited

Loading

benRenard commented Feb 11, 2025

nghi-truyen commented Feb 13, 2025

pag13 left a comment

pag13 Feb 14, 2025

pag13 Feb 14, 2025

pag13 Feb 14, 2025

pag13 Feb 14, 2025

pag13 Feb 14, 2025

nghi-truyen Feb 17, 2025 •

edited

Loading

		3. it naturally handles multi-gauge calibration, with a built-in weighting of the information brought by each gauge.

		This tutorial builds on the :ref:`tutorial on model calibration<user_guide.classical_uses.split_sample_temporal_validation>`: we recommend that you follow it first.

	This tutorial builds on the :ref:`tutorial on model calibration<user_guide.classical_uses.split_sample_temporal_validation>`: we recommend that you follow it first.
	This tutorial builds on the :ref:`<user_guide.classical_uses.fully_distributed_calibration>`: we recommend that you follow it first.

		model0 = smash.Model(setup, mesh)

		We finally optimize this model using the standard, non-Bayesian approach described in the :ref:`model calibration tutorial<user_guide.classical_uses.split_sample_temporal_validation>`. Note that in addition to the four model parameters ``cp``, ``ct``, ``kexc`` and ``llr``, we also calibrate the initial state values (``hp`` and ``ht``) so that we work with 'good' initial state values in the remainder of this tutorial. In addition, a single gauge is used for calibration (the most downstream one, with code "V3524010").

		.. ipython:: python

		calibrated_parameters=["cp","ct","kexc","llr","hp","ht"]

	calibrated_parameters=["cp","ct","kexc","llr","hp","ht"]
	calibrated_parameters = ["cp", "ct", "kexc", "llr"," hp", "ht"]

-	calibration_gauges=["V3524010"]
-	optimize_options={"parameters": calibrated_parameters}
-	cost_options={"gauge": calibration_gauges}
-	model0.optimize(optimize_options=optimize_options,cost_options=cost_options)
+	optimize_options={
+	    "parameters": calibrated_parameters
+	}
+	calibration_gauges = ["V3524010"]
+	cost_options = {
+	    "gauge": calibration_gaugesy
+	}
+	model_0.optimize(
+	    optimize_options = optimize_options,
+	    cost_options = cost_options
+	)

	model0.optimize(optimize_options=optimize_options,cost_options=cost_options)
	model_0.optimize(optimize_options=optimize_options,cost_options=cost_options)


		optimize_options={"parameters": calibrated_parameters,"control_tfm":'keep'}
		info=smash.optimize_control_info(model=model0,optimize_options=optimize_options)

	model=smash.bayesian_optimize(model=model0,optimize_options=optimize_options,cost_options=cost_options)
	model=smash.bayesian_optimize(model=model_0,optimize_options=optimize_options,cost_options=cost_options)


		The figure below compares the observed and the simulated discharge time series at gauge "V3517010" and indeed shows a quite poor fit, leading to a rather high uncertainty.

		print(f"{info['name'][i]} estimated at : {info['x'][i]:.4f}")

		The figure below compares the observed and the simulated discharge time series at gauge "V3517010" and indeed shows a quite poor fit, leading to a rather high uncertainty.

		@@ -1,8 +1,204 @@
		.. _user_guide.in_depth.bayesian_estimation_approach:

Finished Bayesian tutorial #383

Are you sure you want to change the base?

Finished Bayesian tutorial #383

Conversation

benRenard commented Feb 10, 2025

nghi-truyen left a comment • edited Loading

Choose a reason for hiding this comment

asjeb left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

asjeb commented Feb 11, 2025 • edited Loading

benRenard commented Feb 11, 2025

nghi-truyen commented Feb 13, 2025

pag13 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

nghi-truyen Feb 17, 2025 • edited Loading

Choose a reason for hiding this comment

nghi-truyen left a comment •

edited

Loading

asjeb commented Feb 11, 2025 •

edited

Loading

nghi-truyen Feb 17, 2025 •

edited

Loading