rsrl 0.8.1

A fast, extensible reinforcement learning framework in Rust
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166

use super::*;
use ndarray::{Ix2, IntoDimension};
use std::ops::{Add, AddAssign, Mul, MulAssign};

type GradMap = ::std::collections::HashMap<[usize; 2], f64>;

#[derive(Clone, Debug, Default)]
#[cfg_attr(
    feature = "serde",
    derive(Serialize, Deserialize),
    serde(crate = "serde_crate")
)]
pub struct Sparse {
    dim: Ix2,
    grads: GradMap,
}

impl Sparse {
    pub fn new<D: IntoDimension<Dim = Ix2>>(dim: D, grads: GradMap) -> Result<Sparse, String> {
        if grads.len() == 0 {
            Err("No gradient information passed into Sparse::new().".to_owned())
        } else {
            let dim = dim.into_dimension();
            let (nr, nc) = dim.into_pattern();

            if grads.keys().all(|&[r, c]| r < nr && c < nc) {
                Ok(Sparse { dim, grads, })
            } else {
                Err("Inconsistent dimensions in Sparse::new().".to_owned())
            }
        }
    }

    pub fn new_unchecked<D: IntoDimension<Dim = Ix2>>(dim: D, grads: GradMap) -> Sparse {
        Sparse {
            dim: dim.into_dimension(),
            grads,
        }
    }
}

impl Buffer for Sparse {
    type Dim = Ix2;

    fn raw_dim(&self) -> Ix2 { self.dim }

    fn addto<D: DataMut<Elem = f64>>(&self, weights: &mut ArrayBase<D, Ix2>) {
        for (&idx, pd) in self.grads.iter() {
            if let Some(w) = weights.get_mut(idx) {
                w.add_assign(pd);
            }
        }
    }

    fn scaled_addto<D: DataMut<Elem = f64>>(&self, alpha: f64, weights: &mut ArrayBase<D, Ix2>) {
        for (&idx, &pd) in self.grads.iter() {
            if let Some(w) = weights.get_mut(idx) {
                *w = w.mul_add(alpha, pd);
            }
        }
    }

    fn to_dense(&self) -> Array2<f64> {
        let mut g_matrix = Array2::zeros(self.dim);

        for (&[r, c], &g) in self.grads.iter() {
            g_matrix[(r, c)] = g;
        }

        g_matrix
    }
}

impl BufferMut for Sparse {
    fn zeros<D: IntoDimension<Dim = Ix2>>(dim: D) -> Self {
        Self::new_unchecked(dim, GradMap::new())
    }

    fn map(&self, f: impl Fn(f64) -> f64) -> Self {
        Sparse {
            dim: self.dim.clone(),
            grads: self.grads.iter().map(|(&k, &pd)| (k, f(pd))).collect(),
        }
    }

    fn map_into(self, f: impl Fn(f64) -> f64) -> Self {
        Sparse {
            dim: self.dim,
            grads: self.grads.into_iter().map(|(k, pd)| (k, f(pd))).collect(),
        }
    }

    fn map_inplace(&mut self, f: impl Fn(f64) -> f64) {
        self.grads.values_mut().for_each(|pd| *pd = f(*pd));
    }

    fn merge_inplace(&mut self, other: &Self, f: impl Fn(f64, f64) -> f64) {
        if self.dim != other.dim {
            panic!("Incompatible dimensionality.")
        }

        for (k, x) in self.grads.iter_mut() {
            *x = f(*x, other.grads.get(k).copied().unwrap_or(0.0));
        }

        for (&k, y) in other.grads.iter() {
            self.grads.entry(k).or_insert_with(|| f(0.0, *y));
        }
    }
}

impl Into<Array2<f64>> for Sparse {
    fn into(self) -> Array2<f64> { self.to_dense() }
}

impl Add<Sparse> for Sparse {
    type Output = Sparse;

    fn add(mut self, other: Sparse) -> Sparse {
        self.add_assign(&other);

        self
    }
}

impl AddAssign<Sparse> for Sparse {
    fn add_assign(&mut self, other: Sparse) {
        for (idx, pd_other) in other.grads.into_iter() {
            if let Some(pd) = self.grads.get_mut(&idx) {
                pd.add_assign(&pd_other);
            } else {
                self.grads.insert(idx, pd_other);
            }
        }
    }
}

impl AddAssign<&Sparse> for Sparse {
    fn add_assign(&mut self, other: &Sparse) {
        for (&idx, pd_other) in other.grads.iter() {
            if let Some(pd) = self.grads.get_mut(&idx) {
                pd.add_assign(pd_other);
            } else {
                self.grads.insert(idx, pd_other.clone());
            }
        }
    }
}

impl Mul<f64> for Sparse {
    type Output = Sparse;

    fn mul(mut self, factor: f64) -> Sparse {
        self.mul_assign(factor);

        self
    }
}

impl MulAssign<f64> for Sparse {
    fn mul_assign(&mut self, factor: f64) {
        for pd in self.grads.values_mut() {
            pd.mul_assign(factor);
        }
    }
}