reinforcex 0.0.3

Deep Reinforcement Learning Framework
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

use super::base_distribution::BaseDistribution;
use tch::{Kind, Tensor};

pub struct GaussianDistribution {
    mean: Tensor,
    var: Tensor,
    ln_var: Tensor,
}

impl GaussianDistribution {
    pub fn new(mean: Tensor, var: Tensor) -> Self {
        let ln_var = var.log();
        GaussianDistribution { mean, var, ln_var }
    }
}

impl BaseDistribution for GaussianDistribution {
    fn params(&self) -> (&Tensor, &Tensor) {
        (&self.mean, &self.var)
    }

    fn kl(&self, q: Box<dyn BaseDistribution>) -> Tensor {
        let (q_mean, q_var) = q.params();
        let mean_diff = (&self.mean - q_mean).pow_tensor_scalar(2.0);
        let term1 = q_var.log() - &self.ln_var;
        let term2 = (&self.var + mean_diff) / q_var;
        0.5 * (term1 + term2 - Tensor::from(1.0)).sum(Kind::Float)
    }

    fn entropy(&self) -> Tensor {
        let dim: i64 = self.mean.size()[1];
        let log_term: f64 = 0.5 * ((2.0 * std::f64::consts::PI).ln() + 1.0);
        log_term * dim as f64 + 0.5 * self.ln_var.sum(Kind::Float)
    }

    fn sample(&self) -> Tensor {
        let std = self.var.sqrt();
        let noise = Tensor::randn_like(&self.mean);
        &self.mean + &std * noise
    }

    fn prob(&self, x: &Tensor) -> Tensor {
        self.log_prob(x).exp()
    }

    fn log_prob(&self, x: &Tensor) -> Tensor {
        let diff = (x - &self.mean).pow_tensor_scalar(2.0);
        let eltwise_log_prob: Tensor =
            -0.5 * ((2.0 * std::f64::consts::PI).ln() + &self.ln_var + diff / &self.var);
        eltwise_log_prob.sum_dim_intlist(&[1i64][..], false, Kind::Float)
    }

    fn copy(&self) -> Box<dyn BaseDistribution> {
        Box::new(GaussianDistribution {
            mean: self.mean.shallow_clone(),
            var: self.var.shallow_clone(),
            ln_var: self.ln_var.shallow_clone(),
        })
    }

    fn most_probable(&self) -> Tensor {
        self.mean.shallow_clone()
    }
}

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

    #[test]
    fn test_new_and_params() {
        let mean = Tensor::from_slice(&[0.0, 0.0, 0.0]).view([1, 3]);
        let var = Tensor::from_slice(&[1.0, 1.0, 1.0]).view([1, 3]);
        let gaussian: GaussianDistribution =
            GaussianDistribution::new(mean.shallow_clone(), var.shallow_clone());

        let (mean_out, var_out) = gaussian.params();
        assert_eq!(mean_out, &mean);
        assert_eq!(var_out, &var);
    }

    #[test]
    fn test_most_probable() {
        let mean = Tensor::from_slice(&[0.0, 1.0]).view([1, 2]);
        let var = Tensor::from_slice(&[1.0, 4.0]).view([1, 2]);
        let gaussian: GaussianDistribution = GaussianDistribution::new(mean.shallow_clone(), var);

        let most_probable = gaussian.most_probable();
        assert_eq!(most_probable, mean);
    }

    #[test]
    fn test_sample() {
        let mean = Tensor::from_slice(&[0.0, 1.0]).view([1, 2]);
        let var = Tensor::from_slice(&[1.0, 4.0]).view([1, 2]);
        let gaussian: GaussianDistribution = GaussianDistribution::new(mean, var);

        let sample = gaussian.sample();
        assert_eq!(sample.size(), vec![1, 2]);
    }

    #[test]
    fn test_log_prob() {
        let mean = Tensor::from_slice(&[0.0]).view([1, 1]);
        let var = Tensor::from_slice(&[1.0]).view([1, 1]);
        let gaussian = GaussianDistribution::new(mean, var);

        let x = Tensor::from_slice(&[1.0, 2.0, 3.0]).view([1, 3]);
        let log_prob = gaussian.log_prob(&x);
        let expected_log_prob: f64 = -9.7568156;
        assert!((log_prob.double_value(&[]) - expected_log_prob).abs() < 1e-6);
    }

    #[test]
    fn test_kl_divergence() {
        let mean_p = Tensor::from_slice(&[0.0]).view([1, 1]);
        let var_p = Tensor::from_slice(&[1.0]).view([1, 1]);
        let gaussian_p: GaussianDistribution = GaussianDistribution::new(mean_p, var_p);

        let mean_q = Tensor::from_slice(&[0.8]).view([1, 1]);
        let var_q = Tensor::from_slice(&[1.5]).view([1, 1]);
        let gaussian_q: Box<dyn BaseDistribution> =
            Box::new(GaussianDistribution::new(mean_q, var_q));

        let kl_div = gaussian_p.kl(gaussian_q);
        let expected_kl: f64 = 0.249399221;
        assert!((kl_div.double_value(&[]) - expected_kl).abs() < 1e-6);
    }

    #[test]
    fn test_entropy() {
        let mean = Tensor::from_slice(&[0.0]).view([1, 1]);
        let var = Tensor::from_slice(&[1.0]).view([1, 1]);
        let gaussian: GaussianDistribution = GaussianDistribution::new(mean, var);

        let entropy = gaussian.entropy();
        let expected_entropy: f64 = 1.418938533;
        assert!((entropy.double_value(&[]) - expected_entropy).abs() < 1e-6);

        let mean = Tensor::from_slice(&[0.0, 0.0]).view([1, 2]);
        let var = Tensor::from_slice(&[1.0, 1.0]).view([1, 2]);
        let gaussian: GaussianDistribution = GaussianDistribution::new(mean, var);

        let entropy = gaussian.entropy();
        let expected_entropy: f64 = 1.418938533 * 2.0;
        assert!((entropy.double_value(&[]) - expected_entropy).abs() < 1e-6);
    }
}