Configuring DNN meta data
Introduction
In order to define the meta data that describes the bindings to the DNN SVD uses a single configuration file which is set by the dnn.metadata setting in the main configuration file.
Each binding is described by a number of settings, which follow this naming scheme:
input.<tensorname>.<key> = <value>
Common settings
enabled
Boolean value, if true the binding is active, if false the binding is not active ### dim The number of dimensions of the input tensor. Possible values are 0, 1, and 2. 0 dimensions can be used for constants (e.g. training_phase = false) ### sizeX number of elements along the first axis of the tensor. For 1-dimensional tensors this is the total number of elements, for 2-dimensional tensors it is the number of rows ### sizeY number of elements along the second axis of the tensor. For 1-dimensional tensors this is 0, for 2-dimensional tensors it is the number of columns ### dtype The data type of the tensor. Currently, the following data types are supported:
| dtype | description |
|---|---|
| float | 32-bit floating point number; this is the preferred floating point type used for DNN (since calculations are faster compared 64-bit floating points (double)) |
| int16 | 16-bit signed integer (value range: -32,768 to +32,767) |
| int64 | 64-bit signed integer (value range: -9,223,372,036,854,775,808 to +9,223,372,036,854,775,807) |
| uint16 | 16-bit unsigned integer (value range: 0 - 65,535) |
| float16 | 16-bit floating point number (half precision) |
| bool | 1 bit boolean value |
Note, that not all bindings support all possible data types.
type
The type tells SVD what kind of data is expected by the binding. The following types are possible, detailed description are below.
| type | description |
|---|---|
| State | The current state of the cell (single integer value of stateId - 1). Data type is uint16. |
| ResidenceTime | the residence time of the cell in years divided by 10. Data type is float, single value. |
| Climate | climate time series for the next 10 years for the given cell. The tensor is 2-dimensional and the data type is float. |
| Neighbors | a 1-dimensional vector (data type float) with species share for the local and intermediate neighborhood. |
| Scalar | a constant (with a fixed value set during the setup process) |
| Var | a 1-dimensional vector (data type float) with variable size that is populated with user-defined values. |
| Function | a single float value that is calculated on the fly by built in functions. |
Deprecated types (used in earlier versions of SVD):
| type | description |
|---|---|
| DistanceOutside | single float value populated with the environment variable distanceOutside |
| SiteNPKA | two float values, populated with “(availableNitrogen-58.5)/41.436” and “(soilDepth-58.5)/41.436” |
Available bindings
Var
Defines a 1d vector (with size sizeX) and populates the vector with user-defined “expressions”. The expressions are defined by the transformations setting (e.g., to specify the transformations for a tensor with the name site, use input.site.transformations).
The syntax is as follows:
input.<name>.transformations = { expression1 }, { expressione2 }, ..., { expressionN }expression[1-N] corresponds to the 1st - Nth value in the vector, i.e. N equals sizeX.
The expression can make use of all available context variables for a cell (e.g. state variables and environment variables).
Example:
# environment variables for available nitrogen and percentage of sand, divided by 100
input.site.transformations={availableNitrogen/100}, {pctSand/100}Function
The Function binding is a gateway to data items that are provided by (hard-coded) functions in SVD. The function key defines which function to execute for a cell.
| function | description |
|---|---|
| DistToSeedSource | calculates the minimum distance to a cell that acts as seed source for the current cell |
Example:
# specifies to use the "DistToSeedSource" function for the "distance" tensor
input.distance.function=DistToSeedSourceNeighbors
SVD calculates the shares for the local and intermediate neighborhood for all available species (i.e. potential seed input). The species shares are calculated based on state information (mainly the compositional type). A list of available species is defined with the model.species setting in the SVD config file. For each species in model.species the function produces two values (local, and intermediate neighborhood). The size of the tensor has to be number of species * 2.
TODO: more details
Climate
TODO: more details