ANN

	This is part of the annfunc module
	It is only available if you configure PLUMED with ./configure –enable-modules=annfunc . Furthermore, this feature is still being developed so take care when using it and report any problems on the mailing list.

Calculates the ANN-function.

This module implements ANN class, which is a subclass of Function class. ANN class takes multi-dimensional arrays as inputs for a fully-connected feedforward neural network with specified neural network weights and generates corresponding outputs. The ANN outputs can be used as collective variables, inputs for other collective variables, or inputs for data analysis tools.

Examples

Assume we have an ANN with numbers of nodes being [2, 3, 1], and weights connecting layer 0 and 1 are

[[1,2], [3,4], [5,6]]

weights connecting layer 1 and 2 are

[[7,8,9]]

Bias for layer 1 and 2 are [10, 11, 12] and [13], respectively.

All activation functions are Tanh.

Then if input variables are l_0_out_0, l_0_out_1, the corresponding ANN function object can be defined using following plumed script:

Click on the labels of the actions for more information on what each action computes

ann: ANN ...
   
   
ARG
the input for this action is the scalar output from one or more other actions. 
=l_0_out_0,l_0_out_1 
   
NUM_LAYERS
compulsory keyword 
number of layers of the neural network 
=3 
   
NUM_NODES
compulsory keyword 
numbers of nodes in each layer of the neural network 
=2,3,1 
   
ACTIVATIONS
compulsory keyword 
activation functions for the neural network 
=Tanh,Tanh 
   
WEIGHTS0
flattened weight arrays connecting adjacent layers, WEIGHTS0 represents flattened
weight array connecting layer 0 and layer 1, WEIGHTS1 represents flattened weight
array connecting layer 1 and layer 2, ... 
=1,2,3,4,5,6 
   
WEIGHTS1
flattened weight arrays connecting adjacent layers, WEIGHTS0 represents flattened
weight array connecting layer 0 and layer 1, WEIGHTS1 represents flattened weight
array connecting layer 1 and layer 2, ... 
=7,8,9 
   
BIASES0
bias array for each layer of the neural network, BIASES0 represents bias array for
layer 1, BIASES1 represents bias array for layer 2, ... 
=10,11,12 
   
BIASES1
bias array for each layer of the neural network, BIASES0 represents bias array for
layer 1, BIASES1 represents bias array for layer 2, ... 
=13 
... You cannot view the components that are calculated by each action for this input file. Sorry 

To access its components, we use "ann.node-0", "ann.node-1", ..., which represents the components of neural network outputs.

Glossary of keywords and components

Description of components

By default the value of the calculated quantity can be referenced elsewhere in the input file by using the label of the action. Alternatively this Action can be used to calculate the following quantities by employing the keywords listed below. These quantities can be referenced elsewhere in the input by using this Action's label followed by a dot and the name of the quantity required from the list below.

Quantity	Description
node	components of ANN outputs

Compulsory keywords

PERIODIC	if the output of your function is periodic then you should specify the periodicity of the function. If the output is not periodic you must state this using PERIODIC=NO
NUM_LAYERS	number of layers of the neural network
NUM_NODES	numbers of nodes in each layer of the neural network
ACTIVATIONS	activation functions for the neural network

Options

NUMERICAL_DERIVATIVES	( default=off ) calculate the derivatives for these quantities numerically
ARG	the input for this action is the scalar output from one or more other actions. The particular scalars that you will use are referenced using the label of the action. If the label appears on its own then it is assumed that the Action calculates a single scalar value. The value of this scalar is thus used as the input to this new action. If * or *.* appears the scalars calculated by all the proceeding actions in the input file are taken. Some actions have multi-component outputs and each component of the output has a specific label. For example a DISTANCE action labelled dist may have three components x, y and z. To take just the x component you should use dist.x, if you wish to take all three components then use dist.*.More information on the referencing of Actions can be found in the section of the manual on the PLUMED Getting Started. Scalar values can also be referenced using POSIX regular expressions as detailed in the section on Regular Expressions. To use this feature you you must compile PLUMED with the appropriate flag. You can use multiple instances of this keyword i.e. ARG1, ARG2, ARG3...
WEIGHTS	flattened weight arrays connecting adjacent layers, WEIGHTS0 represents flattened weight array connecting layer 0 and layer 1, WEIGHTS1 represents flattened weight array connecting layer 1 and layer 2, ... You can use multiple instances of this keyword i.e. WEIGHTS1, WEIGHTS2, WEIGHTS3...
BIASES	bias array for each layer of the neural network, BIASES0 represents bias array for layer 1, BIASES1 represents bias array for layer 2, ... You can use multiple instances of this keyword i.e. BIASES1, BIASES2, BIASES3...