MLNanoShaper

accumulator(processing,logger)

A processing logger that transform logger on multiple batches Ca be used to liss numerical data, for logging to TensorBoardLogger.

AdamDelta(ρ = 0.9, ϵ = 1e-8)
AdaDelta(; [rho, epsilon])

Parameters

• Rho (ρ == rho): Factor by which the gradient is decayed at each time step.
• Machine epsilon (ϵ == epsilon): Constant to prevent division by zero (no need to change default)

AnnotedKDTree(data::StructVector,property::StaticSymbol)

Fields

• data::StructVector
• tree::KDTree

AuxiliaryParameters

The variables that do not influence the outome of the training run. This include the nb_epoch.

abstract type LossType end

LossType is an interface for defining loss functions.

Implementation

• getlossfn(::LossType)::Function : the associated loss function
• metrictype(::Type{<:LossType)::Type : the type of metrics returned by the loss function
• getlosstype(::StaticSymbol)::LossType : the function generating the losstype

Training information used in model training.

Fields

• atoms: the set of atoms used as model input
• skin : the Surface generated by Nanoshaper

TrainingParameters

The training parameters used in the model training. Default values are in the param file. The training is deterministric. Theses values are hased to determine a training run

_train(training_parameters::TrainingParameters, directories::AuxiliaryParameters)

train the model given TrainingParameters and AuxiliaryParameters.

categorical_loss(model, ps, st, (; point, atoms, d_real))

The loss function used by in training. Return the KL divergence between true probability and empirical probability Return the error with the espected distance as a metric.

continus_loss(model, ps, st, (; point, atoms, d_real))

The loss function used by in training. compare the predicted (square) distance with $\frac{1 + anh(d)}{2}$ Return the error with the espected distance as a metric.

distance(x::GeometryBasics.Mesh, y::KDTree)

Return the Hausdorff distance betwen the mesh coordinates

generate_data()

generate data from the parameters files in param/ by downloading the pdb files and running Nanoshaper.

generate_data_points(
    preprocessing::Lux.AbstractExplicitLayer, points::AbstractVector{<:Point3},
    (; atoms, skin)::TreeTrainingData{Float32}, (; ref_distance)::TrainingParameters)

generate the data_points for a set of positions points on one protein.

implicit_surface(atoms::RegularGrid{T},
    model::Lux.StatefulLuxLayer, (;
        cutoff_radius, step)) where {T}

Create a mesh form the isosurface of function `pos -> model(atoms,pos)` using marching cubes algorithm and using step size `step`.

load_data_pqr(T, name::String)

Load a TrainingData{T} from current directory. You should have a pdb and an off file with name name in current directory.

signed_distance(p::Point3, mesh::RegionMesh)::Number

returns the signed distance between point p and the mesh a positive distance means that we are inside the mesh

state

The global state manipulated by the c interface. To use, you must first load the weights using load_weights and the input atoms using load_atoms. Then you can call eval_model to get the field on a certain point.

Batch A wrapper arround a vector to design a variable dimension

ConcatenatedBatch

Represent a vector of arrays of sizes (a..., bn) where bn is the variable dimension of the batch. You can access view of individual arrays with get_slice.

ModelInput

input of the model

Fields

• point::Point3, the position of the input
• atoms::StructVector{Sphere}, the atoms in the neighboord

struct RegularGrid{T<:Number,G,F<:Function}

A structure representing the input of the surface prediction model

RegularGrid(spheres::AbstractVector,radius::Number,[center]) A constructor that take a vector of spheres. If you are not using Geometry basic spheres, you will need to give a function center that return the center of each point.

batched_sum(b::AbstractMatrix,nb_elements::AbstractVector)

compute the sum of a Concatenated batch with ndim = 2. The first dim is the feature dimension. The second dim is the the batch dim.

Given b of size (n,m) and nb_elements of size (k,), the output has size (n,k).

eval_model(x::Float32,y::Float32,z::Float32)::Float32

evaluate the model at coordinates x y z.

evaluate_model

get_cutoff_radius(x::Lux.AbstractLuxLayer)

extract the cutoff radius of a Lux Model.

light_angular_dense(; categorical=false, van_der_waals_channel=false, kargs...)

`light_angular_dense` is a function that generate a lux model.

light_soft_max_angular_dense(; categorical=false, van_der_waals_channel=false, kargs...)

`light_soft_max_angular_dense` is a function that generate a lux model.

load_atoms(start::Ptr{CSphere},length::Cint)::Cint

Load the atoms into the julia model. Start is a pointer to the start of the array of CSphere and length is the length of the array

Return an error status:

• 0: OK
• 1: data could not be read
• 2: unknow error

load_model(path::String)::Cint

Load the model from a MLNanoShaperRunner.SerializedModel serialized state at absolute path path.

Return an error status:

• 0: OK
• 1: file not found
• 2: file could not be deserialized properly
• 3: unknow error

medium_angular_dense(; categorical=false, van_der_waals_channel=false, kargs...)

`medium_angular_dense` is a function that generate a lux model.

production_instantiate(model::SerializedModel;[on_gpu])::Lux.StatefullLuxLayer{true}

Turn a SerializedModel into a Stateful Lux Layer.

tiny_angular_dense(; categorical=false, van_der_waals_channel=false, kargs...)

`tiny_angular_dense` is a function that generate a lux model.

tiny_angular_dense(; categorical=false, van_der_waals_channel=false, kargs...)

`tiny_angular_dense` is a function that generate a lux model.