Incorporating External Atomic Features#

[Input: recipes/basics/extfeatures/]

The possible incorporation of user-specified, external atomic features is arguably one of the most important capabilities of Fortnet, since this kind of software can never claim that all conceivable feature generators are implemented natively. In order to give the user the greatest possible freedom in generating additional input features, Fortformat and Fortnet support external atomic features. After reading this short section, you will be able to make use of it.

Providing the Input#

As an input, we will use conventional HSD blocks, that we are already familiar with from the previous sections:

Network = BPNN {
  Hidden = 2 2
  Activation = tanh
}

Features {
  External = FromDataset {
    Indices = 1 2 3
  }
}

Training = LBFGS {
  Threshold = 1e-08
  NIterations = 1000
  NPrintout = 100
  NSaveNet = 100
  MinDisplacement = 1e-10
  MaxDisplacement = 5e-02
  LineMin = Yes
  Memory = 1000
  Loss = mse
}

Data {
  Dataset = fnetdata.hdf5
  NetstatFile = fortnet.hdf5
}

Options {
  Mode = train
  ReadNetStats = No
  RandomSeed = 123456
}

The new element is located in the External block. By specifying the Indices entry we ask Fortnet to incorporate the corresponding entries of the dataset, provided they are available. So let’s take a closer look at the specification options in the input. In this case, a dataset with three external features per atom is used. A dedicated section introduces the generation of such a dataset, using the Fortformat Python class.

There are essentially two input formats for specifying the external features to be used:

  1. Explicit specification of the feature indices as they appear in the dataset:

    Features {
     .
     .
  External = FromDataset {
    Indices = 1 2 3
  }
}

The advantage is that certain features can be left out or even used multiple times like this:

Features {
     .
     .
  External = FromDataset {
    Indices = 1 1 2 2 3
  }
}

  1. Specification of a contiguous range of indices (bounds included):

    Features {
     .
     .
  External = FromDataset {
    Indices = 1:3
  }
}

This format can therefore be used to specify all the entries in the dataset in a convenient way:

Features {
     .
     .
  External = FromDataset {
    Indices = 1:-1
  }
}

Or for example up to the penultimate entry: Features = 1:-2

Standard Output#

As far as the standard output is concerned, an additional entry will appear on the screen, which provides information about the number and indexing of the features used:

--------------------------------------------------------------------------------

External Features

nr. of external features: 3
dataset indices: 1 2 3

--------------------------------------------------------------------------------