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

use super::ActorConfig;
use crate::{
    model::{ModelBase, SubModel},
    opt::{Optimizer, OptimizerConfig},
    util::OutDim,
};
use anyhow::{Context, Result};
use log::{info, trace};
use serde::{de::DeserializeOwned, Serialize};
use std::{
    path::Path,
};
use tch::{nn, Device, Tensor};

#[allow(clippy::upper_case_acronyms)]
/// Represents a stochastic policy for SAC agents.
pub struct Actor<P>
where
    P: SubModel<Output = (Tensor, Tensor)>,
    P::Config: DeserializeOwned + Serialize + OutDim,
{
    device: Device,
    var_store: nn::VarStore,

    // Dimension of the action vector.
    pub(super) out_dim: i64,

    // Action-value function
    pi: P,

    // Optimizer
    opt_config: OptimizerConfig,
    opt: Optimizer,
}

impl<P> Actor<P>
where
    P: SubModel<Output = (Tensor, Tensor)>,
    P::Config: DeserializeOwned + Serialize + OutDim,
{
    /// Constructs [Actor].
    pub fn build(config: ActorConfig<P::Config>, device: Device) -> Result<Actor<P>> {
        let pi_config = config.pi_config.context("pi_config is not set.")?;
        let out_dim = pi_config.get_out_dim();
        let opt_config = config.opt_config;
        let var_store = nn::VarStore::new(device);
        let pi = P::build(&var_store, pi_config);

        Ok(Actor::_build(
            device, out_dim, opt_config, pi, var_store, None,
        ))
    }

    fn _build(
        device: Device,
        out_dim: i64,
        opt_config: OptimizerConfig,
        pi: P,
        mut var_store: nn::VarStore,
        var_store_src: Option<&nn::VarStore>,
    ) -> Self {
        // Optimizer
        let opt = opt_config.build(&var_store).unwrap();

        // Copy var_store
        if let Some(var_store_src) = var_store_src {
            var_store.copy(var_store_src).unwrap();
        }

        Self {
            device,
            out_dim,
            opt_config,
            var_store,
            opt,
            pi,
        }
    }

    /// Outputs the parameters of Gaussian distribution given an observation.
    pub fn forward(&self, x: &P::Input) -> (Tensor, Tensor) {
        let (mean, std) = self.pi.forward(&x);
        debug_assert_eq!(mean.size().as_slice()[1], self.out_dim);
        debug_assert_eq!(std.size().as_slice()[1], self.out_dim);
        (mean, std)
    }
}

impl<P> Clone for Actor<P>
where
    P: SubModel<Output = (Tensor, Tensor)>,
    P::Config: DeserializeOwned + Serialize + OutDim,
{
    fn clone(&self) -> Self {
        let device = self.device;
        let out_dim = self.out_dim;
        let opt_config = self.opt_config.clone();
        let var_store = nn::VarStore::new(device);
        let pi = self.pi.clone_with_var_store(&var_store);

        Self::_build(
            device,
            out_dim,
            opt_config,
            pi,
            var_store,
            Some(&self.var_store),
        )
    }
}

impl<P> ModelBase for Actor<P>
where
    P: SubModel<Output = (Tensor, Tensor)>,
    P::Config: DeserializeOwned + Serialize + OutDim,
{
    fn backward_step(&mut self, loss: &Tensor) {
        self.opt.backward_step(loss);
    }

    fn get_var_store_mut(&mut self) -> &mut nn::VarStore {
        &mut self.var_store
    }

    fn get_var_store(&self) -> &nn::VarStore {
        &self.var_store
    }

    fn save<T: AsRef<Path>>(&self, path: T) -> Result<()> {
        self.var_store.save(&path)?;
        info!("Save actor to {:?}", path.as_ref());
        let vs = self.var_store.variables();
        for (name, _) in vs.iter() {
            trace!("Save variable {}", name);
        }
        Ok(())
    }

    fn load<T: AsRef<Path>>(&mut self, path: T) -> Result<()> {
        self.var_store.load(&path)?;
        info!("Load actor from {:?}", path.as_ref());
        Ok(())
    }
}