Saving a settings file

In order to launch a training, you need to save a setting file using the interactive tools contained in the folder notebooks (offline_settings_training.ipynb and online_settings_training.ipynb). In this page, I want to present you the different parameters that you will be asked to define. This description will also allow to you to completely understand how large is the model implemented (i.e. what you can do using directly this project and what you can’t).

The following screen shot presents an example of the interactive tool to define the settings for a training.

General parameters for training

• name_model
  String
  Name of the architecture of the network. The name is chosen among the architecture defined in the package MRF.models.
  
  

If the chosen model starts by a projection :

• normalization
  String
  Type of normalization of the data. The possible types of normalization that can be used are defined in the class Normalization of the module Training_parameters.
• initilization
  String
  Type of initilization for the weights of the first linear layer which should perform the projection. The possible types of initialization that can be used are defined in the class Initialization of the module Training_parameters.
• dimension_projection
  Int
  Dimension of the subspace on which the projection will be performed.
• namepca
  str
  Name of the file containing the basis functions obtained with an online PCA algorithm in order to initalize or define the first linear layer supposed to perform the projection. This file needs to be a .mat file with the basis functions saved with the key word basis. The basis functions need to be saved with the size (number of functions) x (length fingerprint).
  
  

• optimizer
  String
  Name of the optimizer used for the training. The possible names to use are defined in the class Optimizer of the module Training_parameters.
  
  Number of epochs.
  
  
• lr
  Float
  Learning rate.
  
  
• noise_type
  String
  Type of noise that you want to use. The different types of noise are defined in the module Training_parameters.
  
  
• noise_level
  Float
  Level of noise that you want to use.
  
  
• batch_size
  Int
  Batch size.
  
  
• nb_iterations
  Int
  Number of iterations.
  
  
• nb_epochs
  Int
  Number of epochs.
  
  
• loss
  Array of String
  Type of loss used. The array has a length 5 corresponding to the 5 parameters (m_{0s}, T_{1f}, T_{2f}, R and T_{2s}). The i-th element of this array gives the type of transformation applied to the parameter i. The different type of losses are defined in the class Loss in the module Training_parameters.
  
  
• params
  Array of Int
  Define the parameters that you want to estimate. Number 0 corresponds to m0s, number 1 to T_{1f}, etc. (The details are defined in the the module Training_parameters with the vectors nametoparam and paramtoname).
  
  
• minpd
  Float
  Minimum proton density.
  
  
• maxpd
  Float
  Maximum proton density.
  
  
• save_name
  String
  Name given to your training.
  
  
• validation
  Bool
  Boolean which is True if you want to track validation loss and errors during the training.
  
  

If validation is True :

• validation_size
  Int
  Number of fingerprints used for the validation dataset.

---

Parameters specific to the Offline setting

• nb_files
  String
  Number of files used for training and validation. I recall that your files needs to be saved in a folder with a name that you can choose and that should be located in MRF/Offline/loading_data.
  
  
• path_files
  String
  Name that you gave to the folder located in MRF/Offline/loading_data which contains the files. Two different type of files should be saved in this folder:

– the files fingerprints1.npy, fingerprints2.npy, … Each file contains an array of size n x (length fingerprint) where n is an integer that should be greater than the batch size chosen. Each fingerprint should be computed considering that the proton density is equal to 1.

– the files params1.npy, params2.npy, … Each file contains an array of size n x (length fingerprint) and gives the parameters used to compute the fingerprints saved in the files fingerprints1.npy, fingerprints2.npy, …

– If your loss requires the knowledge of the Cramer Rao Bound, you also need to save n this folder the files CRBs1.npy, CRBs2.npy, … Each file contains an array of size n x 6 and gives the Cramer Rao Bounds for the parameters m_{0s}, T_1, T_{2f}, R, T_{2s} and PD in this order.

If validation is True :

• small_validation_size
  Int
  Number of fingerprints used for the smaller validation dataset.

---

Parameters specific to the Online setting

• save_samples
  Bool
  Boolean which when set to True will allow to save the sampled points during the training.
  
  
• sampling
  String
  Name of the sampling strategy hat you want to use. The different possible sampling name are defined in the class Sampling of the module Training_parameters. Each name is associated with a particular sampling strategy defined in the method sample of the class Data_class.
  
  
• min_values
  Array of float
  Five dimensional vector containing the minimum values that you want to consider when the parameters space will be sampled. The components of the vector corresponds to the parameters m_{0s}, T_1, T_{2f}, R and T_{2s} in this order.
  
  
• max_values
  Array of float
  Five dimensional vector containing the maximum values that you want to consider when the parameters space will be sampled. The components of the vector corresponds to the parameters m_{0s}, T_1, T_{2f}, R and T_{2s} in this order.
  
  
• t2_wrt_t1
  String
  Define the way you want to sample the parameter T_{2f} with respect to T_1. The different ways to use this option are defined in the class T2wrtT1 of the module Training_parameters.