concision_neural/model/params/
impl_model_params_rand.rs

1/*
2    appellation: impl_model_params_rand <module>
3    authors: @FL03
4*/
5
6use crate::model::params::{DeepParamsBase, ShallowParamsBase};
7
8use crate::model::ModelFeatures;
9use cnc::init::{self, Initialize};
10use cnc::params::ParamsBase;
11use ndarray::{DataOwned, Ix2};
12use num_traits::{Float, FromPrimitive};
13
14use rand_distr::uniform::{SampleUniform, Uniform};
15use rand_distr::{Distribution, StandardNormal};
16
17impl<A, S> ShallowParamsBase<S, Ix2>
18where
19    S: DataOwned<Elem = A>,
20{
21    /// consumes the controller to initialize the various parameters with random values
22    pub fn init(self) -> Self
23    where
24        A: Float + num_traits::FromPrimitive + rand_distr::uniform::SampleUniform,
25        rand_distr::StandardNormal: rand_distr::Distribution<A>,
26    {
27        // initialize the hidden layer(s)
28        let hidden = ParamsBase::glorot_normal(self.hidden().dim());
29        // Controller input weights (alphabet + state -> hidden)
30        let input = ParamsBase::glorot_normal(self.input().dim());
31        // initialize the output layers
32        let output = ParamsBase::glorot_normal(self.output().dim());
33        // return a new instance with the initialized layers
34        Self {
35            hidden,
36            input,
37            output,
38        }
39    }
40    /// returns a new instance of the model initialized with the given features and random
41    /// distribution
42    pub fn init_rand<G, Ds>(features: ModelFeatures, distr: G) -> Self
43    where
44        G: Fn((usize, usize)) -> Ds,
45        Ds: Clone + Distribution<A>,
46        S: DataOwned,
47    {
48        Self {
49            input: ParamsBase::rand(features.dim_input(), distr(features.dim_input())),
50            hidden: ParamsBase::rand(features.dim_hidden(), distr(features.dim_hidden())),
51            output: ParamsBase::rand(features.dim_output(), distr(features.dim_output())),
52        }
53    }
54    /// initialize the model parameters using a glorot normal distribution
55    pub fn glorot_normal(features: ModelFeatures) -> Self
56    where
57        A: Float + FromPrimitive,
58        StandardNormal: Distribution<A>,
59    {
60        Self::init_rand(features, |(rows, cols)| {
61            cnc::init::XavierNormal::new(rows, cols)
62        })
63    }
64    /// initialize the model parameters using a glorot uniform distribution
65    pub fn glorot_uniform(features: ModelFeatures) -> Self
66    where
67        A: Float + FromPrimitive + SampleUniform,
68        <A as SampleUniform>::Sampler: Clone,
69        Uniform<A>: Distribution<A>,
70    {
71        Self::init_rand(features, |(rows, cols)| {
72            init::XavierUniform::new(rows, cols).expect("failed to create distribution")
73        })
74    }
75}
76
77impl<A, S> DeepParamsBase<S, Ix2>
78where
79    S: DataOwned<Elem = A>,
80{
81    /// returns a new instance of the model initialized with the given features and random
82    /// distribution
83    pub fn init_rand<G, Ds>(features: ModelFeatures, distr: G) -> Self
84    where
85        G: Fn((usize, usize)) -> Ds,
86        Ds: Clone + Distribution<A>,
87        S: DataOwned,
88    {
89        let input = ParamsBase::rand(features.dim_input(), distr(features.dim_input()));
90        let hidden = (0..features.layers())
91            .map(|_| ParamsBase::rand(features.dim_hidden(), distr(features.dim_hidden())))
92            .collect::<Vec<_>>();
93
94        let output = ParamsBase::rand(features.dim_output(), distr(features.dim_output()));
95
96        Self::new(input, hidden, output)
97    }
98    /// initialize the model parameters using a glorot normal distribution
99    pub fn glorot_normal(features: ModelFeatures) -> Self
100    where
101        A: Float + FromPrimitive,
102        StandardNormal: Distribution<A>,
103    {
104        Self::init_rand(features, |(rows, cols)| {
105            cnc::init::XavierNormal::new(rows, cols)
106        })
107    }
108    /// initialize the model parameters using a glorot uniform distribution
109    pub fn glorot_uniform(features: ModelFeatures) -> Self
110    where
111        A: Clone + Float + FromPrimitive + SampleUniform,
112        <S::Elem as SampleUniform>::Sampler: Clone,
113        Uniform<S::Elem>: Distribution<S::Elem>,
114    {
115        Self::init_rand(features, |(rows, cols)| {
116            init::XavierUniform::new(rows, cols).expect("failed to create distribution")
117        })
118    }
119}
concision_neural/model/params/impl_model_params_rand.rs

concision_neural/model/params/
impl_model_params_rand.rs
