burn-rl 0.21.0

RL crate for the Burn framework
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

#![warn(missing_docs)]
#![cfg_attr(docsrs, feature(doc_cfg))]

//! A library for training reinforcement learning agents.

/// Module for implementing an environment.
pub mod environment;
/// Module for implementing a policy.
pub mod policy;
/// Transition buffer.
pub mod transition_buffer;

pub use environment::*;
pub use policy::*;
pub use transition_buffer::*;

#[cfg(test)]
pub(crate) type TestBackend = burn_flex::Flex;

#[cfg(test)]
pub(crate) mod tests {
    use crate::{Batchable, Policy, PolicyState, TestBackend};

    use burn_core::record::Record;
    use burn_core::{self as burn};

    /// Mock policy for testing
    ///
    /// Calling `forward()` with a [MockObservation](MockObservation) (list of f32) returns a [MockActionDistribution](MockActionDistribution)
    /// containing a list of 0s of the same length as the observation.
    ///
    /// Calling `action()` with a [MockObservation](MockObservation) (list of f32) returns a [MockAction](MockAction) with a list of actions of the same length as the observation.
    /// The actions are all 1 if the call is requested as deterministic, or else 0.
    #[derive(Clone)]
    pub(crate) struct MockPolicy {}

    impl MockPolicy {
        pub fn new() -> Self {
            Self {}
        }
    }

    impl Policy<TestBackend> for MockPolicy {
        type Observation = MockObservation;
        type ActionDistribution = MockActionDistribution;
        type Action = MockAction;
        type ActionContext = MockActionContext;
        type PolicyState = MockPolicyState;

        fn forward(&mut self, obs: Self::Observation) -> Self::ActionDistribution {
            let mut dists = vec![];

            for _ in obs.0 {
                dists.push(MockActionDistribution(vec![0.]));
            }
            MockActionDistribution::batch(dists)
        }

        fn action(
            &mut self,
            obs: Self::Observation,
            deterministic: bool,
        ) -> (Self::Action, Vec<Self::ActionContext>) {
            let mut actions = vec![];
            let mut contexts = vec![];

            for _ in obs.0 {
                if deterministic {
                    actions.push(MockAction(vec![1]));
                } else {
                    actions.push(MockAction(vec![0]));
                }
                contexts.push(MockActionContext);
            }

            (MockAction::batch(actions), contexts)
        }

        fn update(&mut self, _update: Self::PolicyState) {}

        fn state(&self) -> Self::PolicyState {
            MockPolicyState
        }

        fn load_record(
            self,
            _record: <Self::PolicyState as PolicyState<TestBackend>>::Record,
        ) -> Self {
            self
        }
    }

    /// Mock observation for testing represented as a vector of f32. Can call `batch()` and `unbatch` on it.
    #[derive(Clone)]
    pub(crate) struct MockObservation(pub Vec<f32>);

    /// Mock action for testing represented as a vector of i32. Can call `batch()` and `unbatch` on it.
    #[derive(Clone)]
    pub(crate) struct MockAction(pub Vec<i32>);

    /// Mock action distribution for testing represented as a vector of i32. Can call `batch()` and `unbatch` on it.
    #[derive(Clone)]
    pub(crate) struct MockActionDistribution(Vec<f32>);

    #[derive(Clone)]
    pub(crate) struct MockActionContext;

    /// Mock policy state for testing represented as an arbitrary `usize` that has no effect on the policy.
    #[derive(Clone)]
    pub(crate) struct MockPolicyState;

    #[derive(Clone, Record)]
    pub(crate) struct MockRecord {
        item: usize,
    }

    impl PolicyState<TestBackend> for MockPolicyState {
        type Record = MockRecord;

        fn into_record(self) -> Self::Record {
            MockRecord { item: 0 }
        }

        fn load_record(&self, _record: Self::Record) -> Self {
            self.clone()
        }
    }

    impl Batchable for MockObservation {
        fn batch(items: Vec<Self>) -> Self {
            MockObservation(items.iter().flat_map(|m| m.0.clone()).collect())
        }

        fn unbatch(self) -> Vec<Self> {
            vec![MockObservation(self.0)]
        }
    }

    impl Batchable for MockAction {
        fn batch(items: Vec<Self>) -> Self {
            MockAction(items.iter().flat_map(|m| m.0.clone()).collect())
        }

        fn unbatch(self) -> Vec<Self> {
            let mut actions = vec![];
            for a in self.0 {
                actions.push(MockAction(vec![a]));
            }
            actions
        }
    }

    impl Batchable for MockActionDistribution {
        fn batch(items: Vec<Self>) -> Self {
            MockActionDistribution(items.iter().flat_map(|m| m.0.clone()).collect())
        }

        fn unbatch(self) -> Vec<Self> {
            let mut dists = vec![];
            for _ in self.0 {
                dists.push(MockActionDistribution(vec![0.]));
            }
            dists
        }
    }
}