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
150
151
152
153
154
155
156
157
158
159
160
161
162

use super::base_q_network::BaseQFunction;
use crate::misc::weight_initializer::{he_init, xavier_init};
use tch::nn::{Init, Linear, LinearConfig, Module};
use tch::{nn, no_grad, Tensor};

pub struct FCQNetwork {
    layers: Vec<Linear>,
    n_input_channels: i64,
    action_size: i64,
    n_hidden_layers: usize,
    n_hidden_channels: i64,
}

impl FCQNetwork {
    pub fn new(
        vs: &nn::VarStore,
        n_input_channels: i64,
        action_size: i64,
        n_hidden_layers: usize,
        n_hidden_channels: Option<i64>,
    ) -> Self {
        let root = vs.root();
        let mut layers: Vec<Linear> = Vec::new();
        let n_hidden_channels = n_hidden_channels.unwrap_or(256);

        layers.push(nn::linear(
            &root,
            n_input_channels,
            n_hidden_channels,
            LinearConfig {
                ws_init: he_init(n_input_channels),
                bs_init: Some(Init::Const(0.0)),
                bias: true,
            },
        ));
        for _ in 0..n_hidden_layers {
            layers.push(nn::linear(
                &root,
                n_hidden_channels,
                n_hidden_channels,
                LinearConfig {
                    ws_init: he_init(n_hidden_channels),
                    bs_init: Some(Init::Const(0.0)),
                    bias: true,
                },
            ));
        }
        layers.push(nn::linear(
            &root,
            n_hidden_channels,
            action_size,
            LinearConfig {
                ws_init: xavier_init(n_hidden_channels, action_size),
                bs_init: Some(Init::Const(0.0)),
                bias: true,
            },
        ));

        FCQNetwork {
            layers,
            n_input_channels,
            action_size,
            n_hidden_layers,
            n_hidden_channels,
        }
    }
}

impl BaseQFunction for FCQNetwork {
    fn forward(&self, x: &Tensor) -> Tensor {
        let mut h = x.view([-1, self.n_input_channels]);
        for i in 0..self.layers.len() {
            h = self.layers[i].forward(&h);
            if i < self.layers.len() - 1 {
                h = h.relu();
            }
        }
        h.view([-1, self.action_size])
    }

    fn is_cuda(&self) -> bool {
        false
    }

    fn clone(&self) -> Box<dyn BaseQFunction> {
        let vs = nn::VarStore::new(tch::Device::Cpu);
        let mut cloned_network = FCQNetwork::new(
            &vs,
            self.n_input_channels,
            self.action_size,
            self.n_hidden_layers,
            Some(self.n_hidden_channels),
        );

        no_grad(|| {
            for (cloned_layer, original_layer) in cloned_network.layers.iter_mut().zip(&self.layers)
            {
                cloned_layer.ws.copy_(&original_layer.ws);
                if let Some(ref mut cloned_bs) = cloned_layer.bs {
                    if let Some(ref original_bs) = &original_layer.bs {
                        cloned_bs.copy_(original_bs);
                    }
                }
            }
        });

        Box::new(cloned_network)
    }
}

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

    #[test]
    fn test_fcqnetwork_forward() {
        let vs = nn::VarStore::new(Device::Cpu);
        let n_input_channels = 4;
        let action_size = 2;
        let n_hidden_layers = 2;
        let n_hidden_channels = Some(64);

        let network = FCQNetwork::new(
            &vs,
            n_input_channels,
            action_size,
            n_hidden_layers,
            n_hidden_channels,
        );

        let input = Tensor::randn([1, n_input_channels], (tch::Kind::Float, Device::Cpu));
        let output = network.forward(&input);

        assert_eq!(output.size(), vec![1, action_size]);
    }

    #[test]
    fn test_fcqnetwork_clone() {
        let vs = nn::VarStore::new(Device::Cpu);
        let n_input_channels = 4;
        let action_size = 2;
        let n_hidden_layers = 2;
        let n_hidden_channels = Some(64);

        let network = FCQNetwork::new(
            &vs,
            n_input_channels,
            action_size,
            n_hidden_layers,
            n_hidden_channels,
        );
        let cloned_network = network.clone();

        let input = Tensor::randn([1, n_input_channels], (tch::Kind::Float, Device::Cpu));
        let output_original = network.forward(&input);
        let output_cloned = cloned_network.forward(&input);

        assert_eq!(output_original.size(), output_cloned.size());
        assert!(output_original.allclose(&output_cloned, 1e-6, 1e-6, false));
    }
}