refacto: bot directories

bot/src/burnrl/dqn/dqn_model.rs

@@ -0,0 +1,216 @@
+use crate::burnrl::dqn::utils::soft_update_linear;
+use crate::burnrl::environment::TrictracEnvironment;
+use burn::module::Module;
+use burn::nn::{Linear, LinearConfig};
+use burn::optim::AdamWConfig;
+use burn::tensor::activation::relu;
+use burn::tensor::backend::{AutodiffBackend, Backend};
+use burn::tensor::Tensor;
+use burn_rl::agent::DQN;
+use burn_rl::agent::{DQNModel, DQNTrainingConfig};
+use burn_rl::base::{Action, ElemType, Environment, Memory, Model, State};
+use std::fmt;
+use std::time::SystemTime;
+
+#[derive(Module, Debug)]
+pub struct Net<B: Backend> {
+    linear_0: Linear<B>,
+    linear_1: Linear<B>,
+    linear_2: Linear<B>,
+}
+
+impl<B: Backend> Net<B> {
+    #[allow(unused)]
+    pub fn new(input_size: usize, dense_size: usize, output_size: usize) -> Self {
+        Self {
+            linear_0: LinearConfig::new(input_size, dense_size).init(&Default::default()),
+            linear_1: LinearConfig::new(dense_size, dense_size).init(&Default::default()),
+            linear_2: LinearConfig::new(dense_size, output_size).init(&Default::default()),
+        }
+    }
+
+    fn consume(self) -> (Linear<B>, Linear<B>, Linear<B>) {
+        (self.linear_0, self.linear_1, self.linear_2)
+    }
+}
+
+impl<B: Backend> Model<B, Tensor<B, 2>, Tensor<B, 2>> for Net<B> {
+    fn forward(&self, input: Tensor<B, 2>) -> Tensor<B, 2> {
+        let layer_0_output = relu(self.linear_0.forward(input));
+        let layer_1_output = relu(self.linear_1.forward(layer_0_output));
+
+        relu(self.linear_2.forward(layer_1_output))
+    }
+
+    fn infer(&self, input: Tensor<B, 2>) -> Tensor<B, 2> {
+        self.forward(input)
+    }
+}
+
+impl<B: Backend> DQNModel<B> for Net<B> {
+    fn soft_update(this: Self, that: &Self, tau: ElemType) -> Self {
+        let (linear_0, linear_1, linear_2) = this.consume();
+
+        Self {
+            linear_0: soft_update_linear(linear_0, &that.linear_0, tau),
+            linear_1: soft_update_linear(linear_1, &that.linear_1, tau),
+            linear_2: soft_update_linear(linear_2, &that.linear_2, tau),
+        }
+    }
+}
+
+#[allow(unused)]
+const MEMORY_SIZE: usize = 8192;
+
+pub struct DqnConfig {
+    pub min_steps: f32,
+    pub max_steps: usize,
+    pub num_episodes: usize,
+    pub dense_size: usize,
+    pub eps_start: f64,
+    pub eps_end: f64,
+    pub eps_decay: f64,
+
+    pub gamma: f32,
+    pub tau: f32,
+    pub learning_rate: f32,
+    pub batch_size: usize,
+    pub clip_grad: f32,
+}
+
+impl fmt::Display for DqnConfig {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        let mut s = String::new();
+        s.push_str(&format!("min_steps={:?}\n", self.min_steps));
+        s.push_str(&format!("max_steps={:?}\n", self.max_steps));
+        s.push_str(&format!("num_episodes={:?}\n", self.num_episodes));
+        s.push_str(&format!("dense_size={:?}\n", self.dense_size));
+        s.push_str(&format!("eps_start={:?}\n", self.eps_start));
+        s.push_str(&format!("eps_end={:?}\n", self.eps_end));
+        s.push_str(&format!("eps_decay={:?}\n", self.eps_decay));
+        s.push_str(&format!("gamma={:?}\n", self.gamma));
+        s.push_str(&format!("tau={:?}\n", self.tau));
+        s.push_str(&format!("learning_rate={:?}\n", self.learning_rate));
+        s.push_str(&format!("batch_size={:?}\n", self.batch_size));
+        s.push_str(&format!("clip_grad={:?}\n", self.clip_grad));
+        write!(f, "{s}")
+    }
+}
+
+impl Default for DqnConfig {
+    fn default() -> Self {
+        Self {
+            min_steps: 250.0,
+            max_steps: 2000,
+            num_episodes: 1000,
+            dense_size: 256,
+            eps_start: 0.9,
+            eps_end: 0.05,
+            eps_decay: 1000.0,
+
+            gamma: 0.999,
+            tau: 0.005,
+            learning_rate: 0.001,
+            batch_size: 32,
+            clip_grad: 100.0,
+        }
+    }
+}
+
+type MyAgent<E, B> = DQN<E, B, Net<B>>;
+
+#[allow(unused)]
+pub fn run<E: Environment + AsMut<TrictracEnvironment>, B: AutodiffBackend>(
+    conf: &DqnConfig,
+    visualized: bool,
+) -> DQN<E, B, Net<B>> {
+    // ) -> impl Agent<E> {
+    let mut env = E::new(visualized);
+    // env.as_mut().min_steps = conf.min_steps;
+    env.as_mut().max_steps = conf.max_steps;
+
+    let model = Net::<B>::new(
+        <<E as Environment>::StateType as State>::size(),
+        conf.dense_size,
+        <<E as Environment>::ActionType as Action>::size(),
+    );
+
+    let mut agent = MyAgent::new(model);
+
+    // let config = DQNTrainingConfig::default();
+    let config = DQNTrainingConfig {
+        gamma: conf.gamma,
+        tau: conf.tau,
+        learning_rate: conf.learning_rate,
+        batch_size: conf.batch_size,
+        clip_grad: Some(burn::grad_clipping::GradientClippingConfig::Value(
+            conf.clip_grad,
+        )),
+    };
+
+    let mut memory = Memory::<E, B, MEMORY_SIZE>::default();
+
+    let mut optimizer = AdamWConfig::new()
+        .with_grad_clipping(config.clip_grad.clone())
+        .init();
+
+    let mut policy_net = agent.model().as_ref().unwrap().clone();
+
+    let mut step = 0_usize;
+
+    for episode in 0..conf.num_episodes {
+        let mut episode_done = false;
+        let mut episode_reward: ElemType = 0.0;
+        let mut episode_duration = 0_usize;
+        let mut state = env.state();
+        let mut now = SystemTime::now();
+
+        while !episode_done {
+            let eps_threshold = conf.eps_end
+                + (conf.eps_start - conf.eps_end) * f64::exp(-(step as f64) / conf.eps_decay);
+            let action =
+                DQN::<E, B, Net<B>>::react_with_exploration(&policy_net, state, eps_threshold);
+            let snapshot = env.step(action);
+
+            episode_reward +=
+                <<E as Environment>::RewardType as Into<ElemType>>::into(snapshot.reward().clone());
+
+            memory.push(
+                state,
+                *snapshot.state(),
+                action,
+                snapshot.reward().clone(),
+                snapshot.done(),
+            );
+
+            if config.batch_size < memory.len() {
+                policy_net =
+                    agent.train::<MEMORY_SIZE>(policy_net, &memory, &mut optimizer, &config);
+            }
+
+            step += 1;
+            episode_duration += 1;
+
+            if snapshot.done() || episode_duration >= conf.max_steps {
+                let envmut = env.as_mut();
+                let goodmoves_ratio = ((envmut.goodmoves_count as f32 / episode_duration as f32)
+                    * 100.0)
+                    .round() as u32;
+                println!(
+                    "{{\"episode\": {episode}, \"reward\": {episode_reward:.4}, \"steps count\": {episode_duration}, \"epsilon\": {eps_threshold:.3}, \"goodmoves\": {}, \"ratio\": {}%, \"rollpoints\":{}, \"duration\": {}}}",
+                    envmut.goodmoves_count,
+                    goodmoves_ratio,
+                    envmut.pointrolls_count,
+                    now.elapsed().unwrap().as_secs(),
+                );
+                if goodmoves_ratio < 5 && 10 < episode {}
+                env.reset();
+                episode_done = true;
+                now = SystemTime::now();
+            } else {
+                state = *snapshot.state();
+            }
+        }
+    }
+    agent
+}

bot/src/burnrl/dqn/main.rs

@@ -0,0 +1,54 @@
+use bot::burnrl::dqn::{
+    dqn_model,
+    utils::{demo_model, load_model, save_model},
+};
+use bot::burnrl::environment;
+use burn::backend::{Autodiff, NdArray};
+use burn_rl::agent::DQN;
+use burn_rl::base::ElemType;
+
+type Backend = Autodiff<NdArray<ElemType>>;
+type Env = environment::TrictracEnvironment;
+
+fn main() {
+    // println!("> Entraînement");
+
+    // See also MEMORY_SIZE in dqn_model.rs : 8192
+    let conf = dqn_model::DqnConfig {
+        //                   defaults
+        num_episodes: 50,  // 40
+        min_steps: 1000.0, // 1000 min of max steps by episode (mise à jour par la fonction)
+        max_steps: 1000,   // 1000 max steps by episode
+        dense_size: 256,   // 128  neural network complexity (default 128)
+        eps_start: 0.9,    // 0.9  epsilon initial value (0.9 => more exploration)
+        eps_end: 0.05,     // 0.05
+        // eps_decay higher = epsilon decrease slower
+        // used in : epsilon = eps_end + (eps_start - eps_end) * e^(-step / eps_decay);
+        // epsilon is updated at the start of each episode
+        eps_decay: 2000.0, // 1000 ?
+
+        gamma: 0.9999, // 0.999 discount factor. Plus élevé = encourage stratégies à long terme
+        tau: 0.0005, // 0.005 soft update rate. Taux de mise à jour du réseau cible. Plus bas = adaptation
+        // plus lente moins sensible aux coups de chance
+        learning_rate: 0.001, // 0.001 taille du pas. Bas : plus lent, haut : risque de ne jamais
+        // converger
+        batch_size: 128, // 32 nombre d'expériences passées sur lesquelles pour calcul de l'erreur moy.
+        clip_grad: 70.0, // 100 limite max de correction à apporter au gradient (default 100)
+    };
+    println!("{conf}----------");
+    let agent = dqn_model::run::<Env, Backend>(&conf, false); //true);
+
+    let valid_agent = agent.valid();
+
+    println!("> Sauvegarde du modèle de validation");
+
+    let path = "bot/models/burnrl_dqn".to_string();
+    save_model(valid_agent.model().as_ref().unwrap(), &path);
+
+    println!("> Chargement du modèle pour test");
+    let loaded_model = load_model(conf.dense_size, &path);
+    let loaded_agent = DQN::new(loaded_model.unwrap());
+
+    println!("> Test avec le modèle chargé");
+    demo_model(loaded_agent);
+}

bot/src/burnrl/dqn/mod.rs

@@ -0,0 +1,2 @@
+pub mod dqn_model;
+pub mod utils;

bot/src/burnrl/dqn/utils.rs

@@ -0,0 +1,112 @@
+use crate::burnrl::dqn::dqn_model;
+use crate::burnrl::environment::{TrictracAction, TrictracEnvironment};
+use crate::training_common::get_valid_action_indices;
+use burn::backend::{ndarray::NdArrayDevice, NdArray};
+use burn::module::{Module, Param, ParamId};
+use burn::nn::Linear;
+use burn::record::{CompactRecorder, Recorder};
+use burn::tensor::backend::Backend;
+use burn::tensor::cast::ToElement;
+use burn::tensor::Tensor;
+use burn_rl::agent::{DQNModel, DQN};
+use burn_rl::base::{Action, ElemType, Environment, State};
+
+pub fn save_model(model: &dqn_model::Net<NdArray<ElemType>>, path: &String) {
+    let recorder = CompactRecorder::new();
+    let model_path = format!("{path}.mpk");
+    println!("Modèle de validation sauvegardé : {model_path}");
+    recorder
+        .record(model.clone().into_record(), model_path.into())
+        .unwrap();
+}
+
+pub fn load_model(dense_size: usize, path: &String) -> Option<dqn_model::Net<NdArray<ElemType>>> {
+    let model_path = format!("{path}.mpk");
+    // println!("Chargement du modèle depuis : {model_path}");
+
+    CompactRecorder::new()
+        .load(model_path.into(), &NdArrayDevice::default())
+        .map(|record| {
+            dqn_model::Net::new(
+                <TrictracEnvironment as Environment>::StateType::size(),
+                dense_size,
+                <TrictracEnvironment as Environment>::ActionType::size(),
+            )
+            .load_record(record)
+        })
+        .ok()
+}
+
+pub fn demo_model<B: Backend, M: DQNModel<B>>(agent: DQN<TrictracEnvironment, B, M>) {
+    let mut env = TrictracEnvironment::new(true);
+    let mut done = false;
+    while !done {
+        // let action = match infer_action(&agent, &env, state) {
+        let action = match infer_action(&agent, &env) {
+            Some(value) => value,
+            None => break,
+        };
+        // Execute action
+        let snapshot = env.step(action);
+        done = snapshot.done();
+    }
+}
+
+fn infer_action<B: Backend, M: DQNModel<B>>(
+    agent: &DQN<TrictracEnvironment, B, M>,
+    env: &TrictracEnvironment,
+) -> Option<TrictracAction> {
+    let state = env.state();
+    // Get q-values
+    let q_values = agent
+        .model()
+        .as_ref()
+        .unwrap()
+        .infer(state.to_tensor().unsqueeze());
+    // Get valid actions
+    let valid_actions_indices = get_valid_action_indices(&env.game);
+    if valid_actions_indices.is_empty() {
+        return None; // No valid actions, end of episode
+    }
+    // Set non valid actions q-values to lowest
+    let mut masked_q_values = q_values.clone();
+    let q_values_vec: Vec<f32> = q_values.into_data().into_vec().unwrap();
+    for (index, q_value) in q_values_vec.iter().enumerate() {
+        if !valid_actions_indices.contains(&index) {
+            masked_q_values = masked_q_values.clone().mask_fill(
+                masked_q_values.clone().equal_elem(*q_value),
+                f32::NEG_INFINITY,
+            );
+        }
+    }
+    // Get best action (highest q-value)
+    let action_index = masked_q_values.argmax(1).into_scalar().to_u32();
+    let action = TrictracAction::from(action_index);
+    Some(action)
+}
+
+fn soft_update_tensor<const N: usize, B: Backend>(
+    this: &Param<Tensor<B, N>>,
+    that: &Param<Tensor<B, N>>,
+    tau: ElemType,
+) -> Param<Tensor<B, N>> {
+    let that_weight = that.val();
+    let this_weight = this.val();
+    let new_weight = this_weight * (1.0 - tau) + that_weight * tau;
+
+    Param::initialized(ParamId::new(), new_weight)
+}
+
+pub fn soft_update_linear<B: Backend>(
+    this: Linear<B>,
+    that: &Linear<B>,
+    tau: ElemType,
+) -> Linear<B> {
+    let weight = soft_update_tensor(&this.weight, &that.weight, tau);
+    let bias = match (&this.bias, &that.bias) {
+        (Some(this_bias), Some(that_bias)) => Some(soft_update_tensor(this_bias, that_bias, tau)),
+        _ => None,
+    };
+
+    Linear::<B> { weight, bias }
+}