relearn 0.3.1

A Reinforcement Learning library
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

//! Deterministic distribution
use crate::utils::distributions::ArrayDistribution;
use tch::{Device, Kind, Tensor};

/// A deterministic distribution over size-0 vectors.
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub struct DeterministicEmptyVec {
    /// `batch_shape + [0]`
    sampling_shape: Vec<i64>,
}

impl DeterministicEmptyVec {
    #[must_use]
    pub fn new(mut batch_shape: Vec<i64>) -> Self {
        batch_shape.push(0);
        Self {
            sampling_shape: batch_shape,
        }
    }
}

impl ArrayDistribution<Tensor, Tensor> for DeterministicEmptyVec {
    fn batch_shape(&self) -> Vec<usize> {
        self.sampling_shape
            .split_last()
            .unwrap()
            .1
            .iter()
            .map(|&s| s.try_into().unwrap())
            .collect()
    }

    fn element_shape(&self) -> Vec<usize> {
        vec![0]
    }

    fn sample(&self) -> Tensor {
        Tensor::empty(&self.sampling_shape, (Kind::Int64, Device::Cpu))
    }

    fn log_probs(&self, _elements: &Tensor) -> Tensor {
        Tensor::zeros(
            self.sampling_shape.split_last().unwrap().1,
            (Kind::Float, Device::Cpu),
        )
    }
    fn entropy(&self) -> Tensor {
        Tensor::zeros(
            self.sampling_shape.split_last().unwrap().1,
            (Kind::Float, Device::Cpu),
        )
    }
    fn kl_divergence_from(&self, other: &Self) -> Tensor {
        let self_batch_shape = self.sampling_shape.split_last().unwrap().1;
        let other_batch_shape = other.sampling_shape.split_last().unwrap().1;
        let shape =
            broadcast_shapes(self_batch_shape, other_batch_shape).expect("Mismatched shapes");
        Tensor::zeros(&shape, (Kind::Float, Device::Cpu))
    }
}

fn broadcast_shapes<'a>(mut a: &'a [i64], mut b: &'a [i64]) -> Option<Vec<i64>> {
    // Ensure that b is not longer. It will be prepended with 1s
    if b.len() > a.len() {
        std::mem::swap(&mut a, &mut b);
    }
    let mut broadcasted = Vec::new();
    for (&ai, &bi) in a
        .iter()
        .rev()
        .zip(b.iter().rev().chain(std::iter::repeat(&1)))
    {
        if ai == bi || bi == 1 {
            broadcasted.push(ai);
        } else if ai == 1 {
            broadcasted.push(bi)
        } else {
            return None;
        }
    }
    broadcasted.reverse();
    Some(broadcasted)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn batch_shape() {
        let d = DeterministicEmptyVec::new(vec![2, 3]);
        assert_eq!(d.batch_shape(), vec![2, 3]);
    }

    #[test]
    fn element_shape() {
        let d = DeterministicEmptyVec::new(vec![2, 3]);
        assert_eq!(d.element_shape(), vec![0]);
    }

    #[test]
    fn sample() {
        let d = DeterministicEmptyVec::new(vec![2, 3]);
        assert_eq!(
            d.sample(),
            Tensor::empty(&[2, 3, 0], (Kind::Int64, Device::Cpu))
        );
    }

    #[test]
    fn log_probs() {
        let d = DeterministicEmptyVec::new(vec![2, 3]);
        let elements = Tensor::empty(&[2, 3, 0], (Kind::Int64, Device::Cpu));
        assert_eq!(
            d.log_probs(&elements),
            Tensor::zeros(&[2, 3], (Kind::Float, Device::Cpu))
        );
    }

    #[test]
    fn entropies() {
        let d = DeterministicEmptyVec::new(vec![2, 3]);
        assert_eq!(
            d.entropy(),
            Tensor::zeros(&[2, 3], (Kind::Float, Device::Cpu))
        );
    }

    #[test]
    fn kl_divergence() {
        let d = DeterministicEmptyVec::new(vec![2, 3]);
        assert_eq!(
            d.kl_divergence_from(&d),
            Tensor::zeros(&[2, 3], (Kind::Float, Device::Cpu))
        );
    }

    #[test]
    fn kl_divergence_broadcast() {
        let d1 = DeterministicEmptyVec::new(vec![1, 3, 1]);
        let d2 = DeterministicEmptyVec::new(vec![4, 2, 1, 1]);
        assert_eq!(
            d1.kl_divergence_from(&d2),
            Tensor::zeros(&[4, 2, 3, 1], (Kind::Float, Device::Cpu))
        );
    }
}