reinforcex 0.0.4

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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236

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

pub struct SoftmaxDistribution {
    logits: Tensor,
    beta: f64,
    min_prob: f64,
    n_classes: i64,
}

unsafe impl Sync for SoftmaxDistribution {}
unsafe impl Send for SoftmaxDistribution {}

impl SoftmaxDistribution {
    pub fn new(logits: Tensor, beta: f64, min_prob: f64) -> Self {
        assert!(
            logits.dim() == 2,
            "Logits must be 2D: (batch_size, n_classes)"
        );
        let n_classes = logits.size()[1];
        assert!(
            min_prob * (n_classes as f64) <= 1.0,
            "Invalid min_prob value"
        );
        Self {
            logits,
            beta,
            min_prob,
            n_classes,
        }
    }
}

impl BaseDistribution for SoftmaxDistribution {
    fn params(&self) -> (&Tensor, &Tensor) {
        (&self.logits, &self.logits)
    }

    fn all_prob(&self) -> Tensor {
        let scaled_logits = &self.logits * self.beta;
        let softmax = scaled_logits.softmax(-1, Kind::Float);
        softmax * (1.0 - self.min_prob * (self.n_classes as f64)) + self.min_prob
    }

    fn all_log_prob(&self) -> Tensor {
        self.all_prob().log()
    }

    fn detach(&mut self) {
        self.logits = self.logits.detach();
    }

    fn kl(&self, q: &Box<dyn BaseDistribution>) -> Tensor {
        (self.all_prob() * (self.all_log_prob() - q.all_log_prob())).sum_dim_intlist(
            [-1].as_ref(),
            false,
            Kind::Float,
        )
    }

    fn entropy(&self) -> Tensor {
        -(self.all_prob() * self.all_log_prob()).sum_dim_intlist([-1].as_ref(), false, Kind::Float)
    }

    fn sample(&self) -> Tensor {
        let probs = self.all_prob();
        let noise = Tensor::rand(&probs.size(), (Kind::Float, probs.device())).log() * -1.0;
        (probs.log() + noise).argmax(-1, false)
    }

    fn prob(&self, x: &Tensor) -> Tensor {
        self.all_prob().gather(-1, &x, false).squeeze_dim(-1)
    }

    fn log_prob(&self, x: &Tensor) -> Tensor {
        self.all_log_prob().gather(-1, &x, false).squeeze_dim(-1)
    }

    fn copy(&self) -> Box<dyn BaseDistribution> {
        Box::new(Self::new(
            self.logits.shallow_clone().detach(),
            self.beta,
            self.min_prob,
        ))
    }

    fn most_probable(&self) -> Tensor {
        self.all_prob().argmax(-1, false)
    }

    fn concat(&mut self, others: Vec<Box<dyn BaseDistribution>>) {
        // self の logits を最初に入れる
        let mut logits = vec![self.logits.shallow_clone()];

        // others の logits を追加
        logits.extend(others.iter().map(|d| d.params().0.shallow_clone()));

        // まとめて cat
        self.logits = Tensor::cat(&logits, 0);
    }
}

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

    #[test]
    fn test_all_prob() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);

        // Test probabilities sum to 1
        let all_prob = dist.all_prob();
        assert_eq!(all_prob.dim(), 2);
        assert_eq!(all_prob.size()[0], 1);
        assert!(all_prob.sum(Kind::Float).double_value(&[]) - 1.0 < 1e-6);

        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0]).reshape(&[2, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);

        // Test probabilities sum to 1
        let all_prob = dist.all_prob();
        assert_eq!(all_prob.dim(), 2);
        assert_eq!(all_prob.size()[0], 2);
        // assert!(all_prob.sum(Kind::Float).allclose(&Tensor::from_slice(&[1.0, 1.0]).to_kind(Kind::Float), 1e-6, 1e-6, true));
    }

    #[test]
    fn test_all_prob_with_min_prob() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let beta = 1.0;
        let min_prob = 0.1;
        let dist = SoftmaxDistribution::new(logits, beta, min_prob);

        // Ensure minimum probability constraint is applied
        let all_prob = dist.all_prob();
        let min_val = all_prob.min().double_value(&[]);
        assert!(min_val >= min_prob);
    }

    #[test]
    fn test_all_log_prob() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);

        let all_log_prob = dist.all_log_prob();
        let log_sum = all_log_prob.sum(Kind::Float).double_value(&[]);
        assert!(log_sum.is_finite());
    }

    #[test]
    fn test_sample() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);
        let sample = dist.sample();
        assert_eq!(sample.size(), [1]);
        assert!(0 <= sample.double_value(&[]) as i64);
        assert!(sample.double_value(&[]) as i64 <= 3);

        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0]).reshape(&[2, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);
        let sample = dist.sample();
        assert_eq!(sample.size(), [2]);
    }

    #[test]
    fn test_prob() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);
        let prob = dist.prob(&Tensor::from_slice(&[0 as i64]).reshape(&[1, 1]));
        assert!((prob.double_value(&[]) - 0.032058604).abs() < 1e-6);
        let prob = dist.prob(&Tensor::from_slice(&[1 as i64]).reshape(&[1, 1]));
        assert!((prob.double_value(&[]) - 0.087144318).abs() < 1e-6);
        let prob = dist.prob(&Tensor::from_slice(&[2 as i64]).reshape(&[1, 1]));
        assert!((prob.double_value(&[]) - 0.236882818).abs() < 1e-6);
        let prob = dist.prob(&Tensor::from_slice(&[3 as i64]).reshape(&[1, 1]));
        assert!((prob.double_value(&[]) - 0.643914260).abs() < 1e-6);
    }

    #[test]
    fn test_log_prob() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);
        let log_prob = dist.log_prob(&Tensor::from_slice(&[0 as i64]).reshape(&[1, 1]));
        assert!((log_prob.double_value(&[]) - (-3.440189702)).abs() < 1e-6);
        let log_prob = dist.log_prob(&Tensor::from_slice(&[1 as i64]).reshape(&[1, 1]));
        assert!((log_prob.double_value(&[]) - (-2.440189702)).abs() < 1e-6);
        let log_prob = dist.log_prob(&Tensor::from_slice(&[2 as i64]).reshape(&[1, 1]));
        assert!((log_prob.double_value(&[]) - (-1.440189702)).abs() < 1e-6);
        let log_prob = dist.log_prob(&Tensor::from_slice(&[3 as i64]).reshape(&[1, 1]));
        assert!((log_prob.double_value(&[]) - (-0.440189702)).abs() < 1e-6);
    }

    #[test]
    fn test_entropy() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);
        let entropy = dist.entropy();
        assert!(entropy.double_value(&[]) >= 0.0);
        assert!((entropy.double_value(&[]) - 0.947536964).abs() < 1e-6)
    }

    #[test]
    fn test_most_probable() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.0);
        let most_probable = dist.most_probable();
        assert_eq!(most_probable.int64_value(&[]), 3);

        let logits = Tensor::from_slice(&[1.0, 3.5, 1.0, 2.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.0, 0.1);
        let most_probable = dist.most_probable();
        assert_eq!(most_probable.int64_value(&[]), 1);

        let logits = Tensor::from_slice(&[1.0, 3.5, 5.0, 2.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 2.0, 0.1);
        let most_probable = dist.most_probable();
        assert_eq!(most_probable.int64_value(&[]), 2);

        let logits = Tensor::from_slice(&[5.1, 3.5, 5.0, 4.0]).reshape(&[1, 4]);
        let dist = SoftmaxDistribution::new(logits, 1.5, 0.0);
        let most_probable = dist.most_probable();
        assert_eq!(most_probable.int64_value(&[]), 0);
    }

    #[test]
    fn test_invalid_min_prob() {
        let logits = Tensor::from_slice(&[1.0, 2.0, 3.0, 4.0]).reshape(&[1, 4]);
        let min_prob = 0.3;
        let result = std::panic::catch_unwind(|| {
            SoftmaxDistribution::new(logits, 1.0, min_prob);
        });
        assert!(result.is_err());
    }
}