relayrl_algorithms 0.4.1

Independent / Multi-Agent Proximal Policy Optimization (PPO) Algorithms for the RelayRL 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

use burn_tensor::backend::Backend;
use burn_tensor::{BasicOps, Float, Tensor, TensorKind};
#[cfg(feature = "tch-backend")]
use relayrl_types::data::tensor::TchDType;
use relayrl_types::data::tensor::{DType, NdArrayDType};
use relayrl_types::prelude::tensor::relayrl::{BackendMatcher, SupportedTensorBackend};

use super::error::NeuralNetworkError;
use super::generic_mlp::GenericMlp;
use super::traits::{NeuralNetwork, NeuralNetworkForward, NeuralNetworkSpec, WeightProvider};
use super::types::{ActivationKind, LayerSpecs};

/// A critic network wrapping a `GenericMlp` and enforcing a single f32 output — the value estimate.
#[derive(Clone, Debug)]
pub struct ValueFunction<
    B: Backend + BackendMatcher<Backend = B>,
    KindIn: TensorKind<B> + BasicOps<B>,
>(GenericMlp<B, KindIn, Float>);

impl<B: Backend + BackendMatcher<Backend = B>, KindIn: TensorKind<B> + BasicOps<B>>
    ValueFunction<B, KindIn>
{
    /// Wraps a `GenericMlp` as a value function, erroring unless its output is a single f32.
    pub fn new(vf_mlp: GenericMlp<B, KindIn, Float>) -> Result<Self, NeuralNetworkError> {
        match (vf_mlp.output_dtype(), vf_mlp.output_dim()) {
            (DType::NdArray(NdArrayDType::F32), 1) => Ok(Self(vf_mlp)),
            #[cfg(feature = "tch-backend")]
            (DType::Tch(TchDType::F32), 1) => Ok(Self(vf_mlp)),
            _ => Err(NeuralNetworkError::UnsupportedOutputParams(
                vf_mlp.output_dtype().to_string(),
                vf_mlp.output_dim().to_string(),
            )),
        }
    }

    /// Builds a value function from a fresh MLP with the given hidden sizes and activation.
    pub fn new_generic_mlp(
        input_dim: usize,
        input_dtype: DType,
        hidden_sizes: &[usize],
        activation: ActivationKind<B>,
        device: &B::Device,
    ) -> Result<Self, NeuralNetworkError> {
        let output_dype: DType = match B::get_supported_backend() {
            SupportedTensorBackend::NdArray => DType::NdArray(NdArrayDType::F32),
            #[cfg(feature = "tch-backend")]
            SupportedTensorBackend::Tch => DType::Tch(TchDType::F32),
            _ => {
                return Err(NeuralNetworkError::BackendUnavailable(
                    match B::get_supported_backend() {
                        SupportedTensorBackend::NdArray => "NdArray",
                        #[cfg(feature = "tch-backend")]
                        SupportedTensorBackend::Tch => "Tch",
                        _ => "None",
                    }
                    .to_string(),
                ));
            }
        };
        Self::new(GenericMlp::new(
            input_dim,
            input_dtype,
            hidden_sizes,
            1,
            output_dype,
            activation,
            device,
        ))
    }

    /// Builds a value function from a default-sized MLP for the given input.
    pub fn new_default_mlp(
        input_dim: usize,
        input_dtype: DType,
        device: &B::Device,
    ) -> Result<Self, NeuralNetworkError> {
        Ok(Self(GenericMlp::default(
            input_dim,
            input_dtype,
            1,
            DType::NdArray(NdArrayDType::F32),
            device,
        )))
    }

    /// Returns the value network's per-layer weight specs for model export.
    pub fn get_vf_layer_specs(&self) -> LayerSpecs {
        self.0.get_layer_specs()
    }
}

impl<
    B: Backend + BackendMatcher<Backend = B>,
    KindIn: TensorKind<B> + BasicOps<B>,
    KindOut: TensorKind<B> + BasicOps<B>,
> NeuralNetworkSpec<B, KindIn, KindOut> for ValueFunction<B, KindIn>
{
    fn input_dim(&self) -> &usize {
        self.0.input_dim()
    }

    fn input_dtype(&self) -> &DType {
        self.0.input_dtype()
    }

    fn output_dim(&self) -> &usize {
        self.0.output_dim()
    }

    fn output_dtype(&self) -> &DType {
        self.0.output_dtype()
    }
}

impl<B: Backend + BackendMatcher<Backend = B>, KindIn: TensorKind<B> + BasicOps<B>>
    NeuralNetworkForward<B, KindIn, Float> for ValueFunction<B, KindIn>
where
    KindIn: BasicOps<B>,
{
    fn forward<const IN_D: usize, const OUT_D: usize>(
        &self,
        input: Tensor<B, IN_D, KindIn>,
    ) -> Tensor<B, OUT_D, Float>
    where
        KindIn: BasicOps<B>,
    {
        self.0.forward(input)
    }
}