unimib-thesis-mad
/
reCTBN
1
0
Fork 0
Code Issues 9 Pull Requests 1 Releases Activity

Bug fix of sampling + implementation of Bayesian Approach

Browse Source
pull/19/head
AlessandroBregoli 3 years ago
parent 5c816ebba7
commit f87900fdbd
6 changed files with 171 additions and 63 deletions
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Show Stats Download Patch File Download Diff File
  1. 2
      Cargo.toml
  2. 41
      src/parameter_learning.rs
  3. 10
      src/params.rs
  4. 122
      src/tools.rs
  5. 55
      tests/parameter_learning.rs
  6. 2
      tests/tools.rs

2
Cargo.toml
Unescape View File

@ -7,7 +7,7 @@ edition = "2021"
[dependencies]
ndarray = "*"
ndarray = {version="*", features=["approx"]}
thiserror = "*"
rand = "*"
bimap = "*"

41
src/parameter_learning.rs
Unescape View File

@ -112,3 +112,44 @@ impl ParameterLearning for MLE {
return (CIM, M, T);
}
}
pub struct BayesianApproach {
default_alpha: usize,
default_tau: f64
}
impl ParameterLearning for BayesianApproach {
fn fit(
&self,
net: Box<&dyn network::Network>,
dataset: &tools::Dataset,
node: usize,
parent_set: Option<BTreeSet<usize>>,
) -> (Array3<f64>, Array3<usize>, Array2<f64>) {
//TODO: make this function general. Now it works only on ContinousTimeDiscreteState nodes
//Use parent_set from parameter if present. Otherwise use parent_set from network.
let parent_set = match parent_set {
Some(p) => p,
None => net.get_parent_set(node),
};
let (mut M, mut T) = sufficient_statistics(net, dataset, node.clone(), &parent_set);
M.mapv_inplace(|x|{x + self.default_alpha});
T.mapv_inplace(|x|{x + self.default_tau});
//Compute the CIM as M[i,x,y]/T[i,x]
let mut CIM: Array3<f64> = Array::zeros((M.shape()[0], M.shape()[1], M.shape()[2]));
CIM.axis_iter_mut(Axis(2))
.zip(M.mapv(|x| x as f64).axis_iter(Axis(2)))
.for_each(|(mut C, m)| C.assign(&(&m/&T)));
//Set the diagonal of the inner matrices to the the row sum multiplied by -1
let tmp_diag_sum: Array2<f64> = CIM.sum_axis(Axis(2)).mapv(|x| x * -1.0);
CIM.outer_iter_mut()
.zip(tmp_diag_sum.outer_iter())
.for_each(|(mut C, diag)| {
C.diag_mut().assign(&diag);
});
return (CIM, M, T);
}
}

10
src/params.rs
Unescape View File

@ -102,7 +102,7 @@ impl ParamsTrait for DiscreteStatesContinousTimeParams {
Option::Some(cim) => {
let mut rng = rand::thread_rng();
let lambda = cim[[u, state, state]] * -1.0;
let x: f64 = rng.gen_range(0.0..1.0);
let x: f64 = rng.gen_range(0.0..=1.0);
Ok(-x.ln() / lambda)
}
Option::None => Err(ParamsError::ParametersNotInitialized(String::from(
@ -119,15 +119,17 @@ impl ParamsTrait for DiscreteStatesContinousTimeParams {
Option::Some(cim) => {
let mut rng = rand::thread_rng();
let lambda = cim[[u, state, state]] * -1.0;
let x: f64 = rng.gen_range(0.0..1.0);
let urand: f64 = rng.gen_range(0.0..=1.0);
let next_state = cim.slice(s![u, state, ..]).map(|x| x / lambda).iter().fold(
(0, 0.0),
|mut acc, ele| {
if &acc.1 + ele < x && ele > &0.0 {
acc.1 += x;
if &acc.1 + ele < urand && ele > &0.0 {
acc.0 += 1;
}
if ele > &0.0 {
acc.1 += ele;
}
acc
},
);

122
src/tools.rs
Unescape View File

@ -1,22 +1,25 @@
use ndarray::prelude::*;
use crate::network;
use crate::node;
use crate::params;
use crate::params::ParamsTrait;
use ndarray::prelude::*;
pub struct Trajectory {
pub time: Array1<f64>,
pub events: Array2<usize>
pub events: Array2<usize>,
}
pub struct Dataset {
pub trajectories: Vec<Trajectory>
pub trajectories: Vec<Trajectory>,
}
pub fn trajectory_generator(net: Box<&dyn network::Network>, n_trajectories: u64, t_end: f64) -> Dataset {
let mut dataset = Dataset{
trajectories: Vec::new()
pub fn trajectory_generator(
net: Box<&dyn network::Network>,
n_trajectories: u64,
t_end: f64,
) -> Dataset {
let mut dataset = Dataset {
trajectories: Vec::new(),
};
let node_idx: Vec<_> = net.get_node_indices().collect();
@ -24,72 +27,93 @@ pub fn trajectory_generator(net: Box<&dyn network::Network>, n_trajectories: u64
let mut t = 0.0;
let mut time: Vec<f64> = Vec::new();
let mut events: Vec<Array1<usize>> = Vec::new();
let mut current_state: Vec<params::StateType> = node_idx.iter().map(|x| {
net.get_node(*x).params.get_random_state_uniform()
}).collect();
let mut next_transitions: Vec<Option<f64>> = (0..node_idx.len()).map(|_| Option::None).collect();
events.push(current_state.iter().map(|x| match x {
params::StateType::Discrete(state) => state.clone()
}).collect());
let mut current_state: Vec<params::StateType> = node_idx
.iter()
.map(|x| net.get_node(*x).params.get_random_state_uniform())
.collect();
let mut next_transitions: Vec<Option<f64>> =
(0..node_idx.len()).map(|_| Option::None).collect();
events.push(
current_state
.iter()
.map(|x| match x {
params::StateType::Discrete(state) => state.clone(),
})
.collect(),
);
time.push(t.clone());
while t < t_end {
for (idx, val) in next_transitions.iter_mut().enumerate(){
for (idx, val) in next_transitions.iter_mut().enumerate() {
if let None = val {
*val = Some(net.get_node(idx).params
.get_random_residence_time(net.get_node(idx).params.state_to_index(&current_state[idx]),
net.get_param_index_network(idx, &current_state)).unwrap() + t);
*val = Some(
net.get_node(idx)
.params
.get_random_residence_time(
net.get_node(idx).params.state_to_index(&current_state[idx]),
net.get_param_index_network(idx, &current_state),
)
.unwrap()
+ t,
);
}
};
}
let next_node_transition = next_transitions
.iter()
.enumerate()
.min_by(|x, y|
x.1.unwrap().partial_cmp(&y.1.unwrap()).unwrap())
.unwrap().0;
.min_by(|x, y| x.1.unwrap().partial_cmp(&y.1.unwrap()).unwrap())
.unwrap()
.0;
if next_transitions[next_node_transition].unwrap() > t_end {
break
break;
}
t = next_transitions[next_node_transition].unwrap().clone();
time.push(t.clone());
current_state[next_node_transition] = net.get_node(next_node_transition).params
.get_random_state(
net.get_node(next_node_transition).params.
state_to_index(
&current_state[next_node_transition]),
net.get_param_index_network(next_node_transition, &current_state))
.unwrap();
events.push(Array::from_vec(current_state.iter().map(|x| match x {
params::StateType::Discrete(state) => state.clone()
}).collect()));
current_state[next_node_transition] = net
.get_node(next_node_transition)
.params
.get_random_state(
net.get_node(next_node_transition)
.params
.state_to_index(&current_state[next_node_transition]),
net.get_param_index_network(next_node_transition, &current_state),
)
.unwrap();
events.push(Array::from_vec(
current_state
.iter()
.map(|x| match x {
params::StateType::Discrete(state) => state.clone(),
})
.collect(),
));
next_transitions[next_node_transition] = None;
for child in net.get_children_set(next_node_transition){
for child in net.get_children_set(next_node_transition) {
next_transitions[child] = None
}
}
events.push(current_state.iter().map(|x| match x {
params::StateType::Discrete(state) => state.clone()
}).collect());
events.push(
current_state
.iter()
.map(|x| match x {
params::StateType::Discrete(state) => state.clone(),
})
.collect(),
);
time.push(t_end.clone());
dataset.trajectories.push(Trajectory {
time: Array::from_vec(time),
events: Array2::from_shape_vec((events.len(), current_state.len()), events.iter().flatten().cloned().collect()).unwrap()
events: Array2::from_shape_vec(
(events.len(), current_state.len()),
events.iter().flatten().cloned().collect(),
)
.unwrap(),
});
}
dataset
}

55
tests/parameter_learning.rs
Unescape View File

@ -46,10 +46,10 @@ fn learn_binary_cim_MLE() {
let (CIM, M, T) = mle.fit(Box::new(&net), &data, 1, None);
print!("CIM: {:?}\nM: {:?}\nT: {:?}\n", CIM, M, T);
assert_eq!(CIM.shape(), [2, 2, 2]);
assert_relative_eq!(-1.0, CIM[[0, 0, 0]], epsilon=0.2);
assert_relative_eq!(-4.0, CIM[[0, 1, 1]], epsilon=0.2);
assert_relative_eq!(-6.0, CIM[[1, 0, 0]], epsilon=0.2);
assert_relative_eq!(-2.0, CIM[[1, 1, 1]], epsilon=0.2);
assert!(CIM.abs_diff_eq(&arr3(&[
[[-1.0, 1.0], [4.0, -4.0]],
[[-6.0, 6.0], [2.0, -2.0]],
]), 0.2));
}
@ -87,9 +87,48 @@ fn learn_ternary_cim_MLE() {
let (CIM, M, T) = mle.fit(Box::new(&net), &data, 1, None);
print!("CIM: {:?}\nM: {:?}\nT: {:?}\n", CIM, M, T);
assert_eq!(CIM.shape(), [3, 3, 3]);
assert_relative_eq!(-1.0, CIM[[0, 0, 0]], epsilon=0.2);
assert_relative_eq!(-4.0, CIM[[0, 1, 1]], epsilon=0.2);
assert_relative_eq!(-1.0, CIM[[0, 2, 2]], epsilon=0.2);
assert_relative_eq!(0.5, CIM[[0, 0, 1]], epsilon=0.2);
assert!(CIM.abs_diff_eq(&arr3(&[
[[-1.0, 0.5, 0.5], [3.0, -4.0, 1.0], [0.9, 0.1, -1.0]],
[[-6.0, 2.0, 4.0], [1.5, -2.0, 0.5], [3.0, 1.0, -4.0]],
[[-1.0, 0.1, 0.9], [2.0, -2.5, 0.5], [0.9, 0.1, -1.0]],
]), 0.2));
}
#[test]
fn learn_ternary_cim_MLE_no_parents() {
let mut net = CtbnNetwork::init();
let n1 = net
.add_node(generate_discrete_time_continous_node(String::from("n1"),3))
.unwrap();
let n2 = net
.add_node(generate_discrete_time_continous_node(String::from("n2"),3))
.unwrap();
net.add_edge(n1, n2);
match &mut net.get_node_mut(n1).params {
params::Params::DiscreteStatesContinousTime(param) => {
param.cim = Some(arr3(&[[[-3.0, 2.0, 1.0],
[1.5, -2.0, 0.5],
[0.4, 0.6, -1.0]]]));
}
}
match &mut net.get_node_mut(n2).params {
params::Params::DiscreteStatesContinousTime(param) => {
param.cim = Some(arr3(&[
[[-1.0, 0.5, 0.5], [3.0, -4.0, 1.0], [0.9, 0.1, -1.0]],
[[-6.0, 2.0, 4.0], [1.5, -2.0, 0.5], [3.0, 1.0, -4.0]],
[[-1.0, 0.1, 0.9], [2.0, -2.5, 0.5], [0.9, 0.1, -1.0]],
]));
}
}
let data = trajectory_generator(Box::new(&net), 100, 200.0);
let mle = MLE{};
let (CIM, M, T) = mle.fit(Box::new(&net), &data, 0, None);
print!("CIM: {:?}\nM: {:?}\nT: {:?}\n", CIM, M, T);
assert_eq!(CIM.shape(), [1, 3, 3]);
assert!(CIM.abs_diff_eq(&arr3(&[[[-3.0, 2.0, 1.0],
[1.5, -2.0, 0.5],
[0.4, 0.6, -1.0]]]), 0.2));
}

2
tests/tools.rs
Unescape View File

@ -41,3 +41,5 @@ fn run_sampling() {
assert_eq!(4, data.trajectories.len());
assert_relative_eq!(1.0, data.trajectories[0].time[data.trajectories[0].time.len()-1]);
}

Write Preview
Loading…
Cancel
Save