ADR suggestion Analysis

[henrikjacobsenfys]

General

The Analysis class is where fitting takes place. This ADR suggestions is a work in progress.

Current implementation

Only a skeleton has been implemented

Suggested implementation

There will be three Analysis classes: AnalysisBase, Analysis1d and Analysis. The base class handles all shared logic, of course.
Analysis1d will have a Q_index property, and will work with data (intensity with variances as function of energy) for that single Q_index. The Analysis class will contain a list of Analysis1d objects; one for each Q_index in the underlying experimental data.

AnalysisBase

AnalysisBase inherits from ModelBase.

It takes the following input:
display_name and unique_name (probably changing to name)
experiment: An EasyDynamics Experiment object, which contains a scipp DataArray with coordinates energy and Q. The Q from the Experiment will be used as the true Q (e.g. the one that the Analysis getters will refer to) and copied into all models.
sample_model: An EasyDynamics SampleModel object, which contains a ComponentCollection for each Q_index instrument_model: An EasyDynamics InstrumentModelobject, which contains an optionalResolutionModelandBackgroundModel(similar to theSampleModel), and a list of energy_offset Parameters, one for each Q`.

It contains setters and getters for all these models. In addition, it has a _create_convolver method, which creates a EasyDynamics Convolution object for a given Q_index, which handles the convolution of the sample ComponentCollection with the instrument resolution ComponentCollection

Analysis1d

It also takes the following two inputs:
a Q_index, which selects a particular Q of the data and all the attached models.
(slightly controversial, but I really want this): extra_parameters: A Parameter or a list of Parameter, which the user can define as they please. I do not know exactly how users will use this, but they may, for example, decide that the widths of the Lorentzians in their model should be dependent Parameters with some custom dependency on another Parameter. Most users will not touch it, but it's there to allow flexibility for advanced users.

Other methods:
get_all_variables(): gets all the variables of the relevant models at self._Q_index, all shared Parameters like the temperature of the SampleModel and any Parameters from DiffusionModels. And of course the extra Parameters that users can pass to it
calculate (or evaluate or something else): Calculate the scattering at self._Q_index. Includes checks and may regenerate the Convolution object if needed
_calculate: doethe actual calculation with minimal checks for speed for the
fit: uses the EasyScience Fitter to fit the data at the given Q_index
calculate_individual_component: calculates individual components including resolution abd background contributions for
plot_data_and_model

Analysis

Contains a list of Analysis1d objects and the following primary methods:
fit: fits the data with various settings, calling various sub-functions, such as
_fit_single_Q: calls the relevant Analysis1d.
_fit_all_Q_independently: calls all the relevant Analysis1d objects one after the other
_fit_all_Q_simultaneously: This is the fun one! It passes all the Analysis1d objects and the correct fit functions to do what the name implies, allowing fits of full diffusion models
_calculate_all_Q: for plotting
_calculate_individual_component_all_Q (name pending): for plotting
plot_data_and_model: uses a Scipp DataGroup and plopp to make slicers and other types of plots easily
plot_fit_parameter: plot e.g. the width of one (or more) of the Lorentzians. Will at first take indices, later names to choose which parameters to plot.
fit_diffusion_model or some thing similar: fit the width (and optional area) of e.g. a Lorentzian to a diffusion model. It should of course also be possible to plot the fit of this parameter together with the values.

This is more or less it for the minimal version, but

Extensions:

For the fitting I also want utility such as:
fit a single Q_index, then copy the values of the fit to the other ones to have better start guesses
Fit all Q independently, but use the previous fit as the start guess for the next one
easy access to fix Parameters across Q, and to make Parameters dependent on each other.

It will also allow extra Parameters, just like the Analysis1d class.

I also of course need to be able to pass settings to the Convolution classes, to the fitter and to the plotting. That will not be there for the minimal model, since it's effort.

A note on Energy

the energy axis is handled somewhat inconsistently throughout the library. I want users to be able to pass a list of numpy array, but it's probably easiest and best to use a scipp variable internally. The work is already growing quite big with some necessary changes to unrelated code, so I'm making an ugly and simple solution that works, and pinky promising to get back to the energy later.

[damskii9992]

A few comments here. The set_resolution_model and set_background_model should probably just be a single set_instrument_model (see my comment on #2).
The set_data should be set_measurement or set_experiment which is where you attach the Measurement (or Experiment) class.
You can still have a get_data method, but it should then just be a shorthand for Analysis.Measurement.get_data()

[henrikjacobsenfys]

Good points! I'll make those changes when I make a proper Measurement class

[rozyczko]

Based on what was discussed in e.g. easyscience/EasyDiffractionLib#97
We shouldn't use set_ for properties.
Instead, set_sample_model should be sample_models.append() and sample_models.remove(), which also means that maybe your sample model should be actually a collection. Do you not envision having more than a single model?

Similar comment is to other setters, especially with data - set_data should probably be a bit more nuanced, as @damskii9992 mentioned.
Don't we want the Experiment (or Measurement) class to own the data, anyway?
I would suggest analysis.experiment.data.append() where data is also a collection (maybe a scipp data array?)