henri/trictrac
1
1
Fork 0
Code Issues 14 Pull requests Actions Packages Projects 1 Releases Wiki Activity

wip

Browse source
This commit is contained in:
Henri Bourcereau 2025-06-28 21:34:44 +02:00
parent cf93255f03
commit f05094b2d4
3 changed files with 150 additions and 75 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

96
bot/src/bin/train_dqn_full.rs
View file

@ -1,5 +1,5 @@
use bot::strategy::burn_dqn_agent::{BurnDqnAgent, DqnConfig, Experience};
use bot::strategy::burn_environment::{TrictracEnvironment, TrictracAction};
use bot::strategy::burn_environment::{TrictracAction, TrictracEnvironment};
use bot::strategy::dqn_common::get_valid_actions;
use burn_rl::base::Environment;
use std::env;
@ -80,7 +80,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
// Configuration DQN
let config = DqnConfig {
state_size: 36,
action_size: 1000, // Espace d'actions réduit via contexte
action_size: 1252, // Espace d'actions réduit via contexte
hidden_size: 256,
learning_rate: 0.001,
gamma: 0.99,
@ -94,6 +94,8 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
// Créer l'agent et l'environnement
let mut agent = BurnDqnAgent::new(config);
let mut optimizer = AdamConfig::new().init();
let mut env = TrictracEnvironment::new(true);
// Variables pour les statistiques
@ -114,7 +116,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
loop {
step += 1;
let current_state = snapshot.state;
let current_state = snapshot.state();
// Obtenir les actions valides selon le contexte du jeu
let valid_actions = get_valid_actions(&env.game);
@ -127,22 +129,31 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
let valid_indices: Vec<usize> = (0..valid_actions.len()).collect();
// Sélectionner une action avec l'agent DQN
let action_index = agent.select_action(&current_state.data.iter().map(|&x| x as f32).collect::<Vec<_>>(), &valid_indices);
let action = TrictracAction { index: action_index as u32 };
let action_index = agent.select_action(
&current_state
.data
.iter()
.map(|&x| x as f32)
.collect::<Vec<_>>(),
&valid_indices,
);
let action = TrictracAction {
index: action_index as u32,
};
// Exécuter l'action
snapshot = env.step(action);
episode_reward += snapshot.reward;
episode_reward += snapshot.reward();
// Préparer l'expérience pour l'agent
let experience = Experience {
state: current_state.data.iter().map(|&x| x as f32).collect(),
action: action_index,
reward: snapshot.reward,
reward: snapshot.reward(),
next_state: if snapshot.terminated {
None
} else {
Some(snapshot.state.data.iter().map(|&x| x as f32).collect())
Some(snapshot.state().data.iter().map(|&x| x as f32).collect())
},
done: snapshot.terminated,
};
@ -151,7 +162,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
agent.add_experience(experience);
// Entraîner l'agent
if let Some(loss) = agent.train_step() {
if let Some(loss) = agent.train_step(optimizer) {
episode_loss += loss;
loss_count += 1;
}
@ -163,7 +174,11 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
}
// Calculer la loss moyenne de l'épisode
let avg_loss = if loss_count > 0 { episode_loss / loss_count as f32 } else { 0.0 };
let avg_loss = if loss_count > 0 {
episode_loss / loss_count as f32
} else {
0.0
};
// Sauvegarder les statistiques
total_rewards.push(episode_reward);
@ -172,9 +187,12 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
// Affichage des statistiques
if episode % save_every == 0 {
let avg_reward = total_rewards.iter().rev().take(save_every).sum::<f32>() / save_every as f32;
let avg_length = episode_lengths.iter().rev().take(save_every).sum::<usize>() / save_every;
let avg_episode_loss = losses.iter().rev().take(save_every).sum::<f32>() / save_every as f32;
let avg_reward =
total_rewards.iter().rev().take(save_every).sum::<f32>() / save_every as f32;
let avg_length =
episode_lengths.iter().rev().take(save_every).sum::<usize>() / save_every;
let avg_episode_loss =
losses.iter().rev().take(save_every).sum::<f32>() / save_every as f32;
println!("Episode {} | Avg Reward: {:.3} | Avg Length: {} | Avg Loss: {:.6} | Epsilon: {:.3} | Buffer: {}",
episode, avg_reward, avg_length, avg_episode_loss, agent.get_epsilon(), agent.get_buffer_size());
@ -187,8 +205,14 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
println!(" → Modèle sauvegardé : {}", checkpoint_path);
}
} else if episode % 10 == 0 {
println!("Episode {} | Reward: {:.3} | Length: {} | Loss: {:.6} | Epsilon: {:.3}",
episode, episode_reward, step, avg_loss, agent.get_epsilon());
println!(
"Episode {} | Reward: {:.3} | Length: {} | Loss: {:.6} | Epsilon: {:.3}",
episode,
episode_reward,
step,
avg_loss,
agent.get_epsilon()
);
}
}
@ -199,18 +223,41 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
// Statistiques finales
println!();
println!("=== Résultats de l'entraînement ===");
let final_avg_reward = total_rewards.iter().rev().take(100.min(episodes)).sum::<f32>() / 100.min(episodes) as f32;
let final_avg_length = episode_lengths.iter().rev().take(100.min(episodes)).sum::<usize>() / 100.min(episodes);
let final_avg_loss = losses.iter().rev().take(100.min(episodes)).sum::<f32>() / 100.min(episodes) as f32;
let final_avg_reward = total_rewards
.iter()
.rev()
.take(100.min(episodes))
.sum::<f32>()
/ 100.min(episodes) as f32;
let final_avg_length = episode_lengths
.iter()
.rev()
.take(100.min(episodes))
.sum::<usize>()
/ 100.min(episodes);
let final_avg_loss =
losses.iter().rev().take(100.min(episodes)).sum::<f32>() / 100.min(episodes) as f32;
println!("Récompense moyenne (100 derniers épisodes) : {:.3}", final_avg_reward);
println!("Longueur moyenne (100 derniers épisodes) : {}", final_avg_length);
println!("Loss moyenne (100 derniers épisodes) : {:.6}", final_avg_loss);
println!(
"Récompense moyenne (100 derniers épisodes) : {:.3}",
final_avg_reward
);
println!(
"Longueur moyenne (100 derniers épisodes) : {}",
final_avg_length
);
println!(
"Loss moyenne (100 derniers épisodes) : {:.6}",
final_avg_loss
);
println!("Epsilon final : {:.3}", agent.get_epsilon());
println!("Taille du buffer final : {}", agent.get_buffer_size());
// Statistiques globales
let max_reward = total_rewards.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
let max_reward = total_rewards
.iter()
.cloned()
.fold(f32::NEG_INFINITY, f32::max);
let min_reward = total_rewards.iter().cloned().fold(f32::INFINITY, f32::min);
println!("Récompense max : {:.3}", max_reward);
println!("Récompense min : {:.3}", min_reward);
@ -220,7 +267,10 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
println!("Modèle final sauvegardé : {}", final_path);
println!();
println!("Pour utiliser le modèle entraîné :");
println!(" cargo run --bin=client_cli -- --bot burn_dqn:{}_final,dummy", model_path);
println!(
" cargo run --bin=client_cli -- --bot burn_dqn:{}_final,dummy",
model_path
);
Ok(())
}

39
bot/src/strategy/burn_dqn_agent.rs
View file

@ -1,12 +1,13 @@
use burn::module::AutodiffModule;
use burn::tensor::backend::AutodiffBackend;
use burn::{
backend::{ndarray::NdArrayDevice, Autodiff, NdArray},
nn::{Linear, LinearConfig, loss::MseLoss},
module::Module,
tensor::Tensor,
optim::{AdamConfig, Optimizer},
nn::{loss::MseLoss, Linear, LinearConfig},
optim::{GradientsParams, Optimizer},
record::{CompactRecorder, Recorder},
tensor::Tensor,
};
use rand::Rng;
use serde::{Deserialize, Serialize};
use std::collections::VecDeque;
@ -26,7 +27,12 @@ pub struct DqnNetwork<B: burn::prelude::Backend> {
impl<B: burn::prelude::Backend> DqnNetwork<B> {
/// Crée un nouveau réseau DQN
pub fn new(input_size: usize, hidden_size: usize, output_size: usize, device: &B::Device) -> Self {
pub fn new(
input_size: usize,
hidden_size: usize,
output_size: usize,
device: &B::Device,
) -> Self {
let fc1 = LinearConfig::new(input_size, hidden_size).init(device);
let fc2 = LinearConfig::new(hidden_size, hidden_size).init(device);
let fc3 = LinearConfig::new(hidden_size, output_size).init(device);
@ -94,7 +100,6 @@ pub struct BurnDqnAgent {
device: MyDevice,
q_network: DqnNetwork<MyBackend>,
target_network: DqnNetwork<MyBackend>,
optimizer: burn::optim::Adam<MyBackend>,
replay_buffer: VecDeque<Experience>,
epsilon: f32,
step_count: usize,
@ -119,14 +124,11 @@ impl BurnDqnAgent {
&device,
);
let optimizer = AdamConfig::new().init();
Self {
config: config.clone(),
device,
q_network,
target_network,
optimizer,
replay_buffer: VecDeque::new(),
epsilon: config.epsilon,
step_count: 0,
@ -146,11 +148,11 @@ impl BurnDqnAgent {
}
// Exploitation : choisir la meilleure action selon le Q-network
let state_tensor = Tensor::<MyBackend, 2>::from_floats([state], &self.device);
let state_tensor = Tensor::<MyBackend, 2>::from_floats(state, &self.device);
let q_values = self.q_network.forward(state_tensor);
// Convertir en vecteur pour traitement
let q_data = q_values.into_data().convert::<f32>().value;
let q_data = q_values.into_data().convert::<f32>().into_vec().unwrap();
// Trouver la meilleure action parmi les actions valides
let mut best_action = valid_actions[0];
@ -175,7 +177,10 @@ impl BurnDqnAgent {
}
/// Entraîne le réseau sur un batch d'expériences
pub fn train_step(&mut self) -> Option<f32> {
pub fn train_step<B: AutodiffBackend, M: AutodiffModule<B>>(
&mut self,
optimizer: &mut impl Optimizer<M, B>,
) -> Option<f32> {
if self.replay_buffer.len() < self.config.batch_size {
return None;
}
@ -197,12 +202,14 @@ impl BurnDqnAgent {
let loss = MseLoss::new().forward(
current_q_values,
target_q_values,
burn::nn::loss::Reduction::Mean
burn::nn::loss::Reduction::Mean,
);
// Backpropagation (version simplifiée)
let grads = loss.backward();
self.q_network = self.optimizer.step(self.config.learning_rate, self.q_network, grads);
// Gradients linked to each parameter of the model.
// let grads = GradientsParams::from_grads(grads, &self.q_network);
self.q_network = optimizer.step(self.config.learning_rate, self.q_network, grads);
// Mise à jour du réseau cible
self.step_count += 1;
@ -258,7 +265,9 @@ impl BurnDqnAgent {
}
/// Charge un modèle pour l'inférence
pub fn load_model_for_inference(path: &str) -> Result<(DqnNetwork<InferenceBackend>, DqnConfig), Box<dyn std::error::Error>> {
pub fn load_model_for_inference(
path: &str,
) -> Result<(DqnNetwork<InferenceBackend>, DqnConfig), Box<dyn std::error::Error>> {
// Charger la configuration
let config_path = format!("{}_config.json", path);
let config_json = std::fs::read_to_string(config_path)?;

16
doc/refs/claudeAIquestionOnlyRust.md
View file

@ -250,3 +250,19 @@ claude-3-5-haiku: 18.8k input, 443 output, 0 cache read, 0 cache write
claude-sonnet: 10 input, 666 output, 0 cache read, 245.6k cache write
Mais pourtant 2 millions indiqués dans la page usage : <https://console.anthropic.com/usage>, et 7.88 dollars de consommés sur <https://console.anthropic.com/cost>.
I just had a claude code session in which I kept having this error, even if the agent didn't seem to read a lot of files : API Error (429 {"type":"error","error":{"type":"rate_limit_error","message":"This request would exceed the rate limit for your organization (813e6b21-ec6f-44c3-a7f0-408244105e5c) of 20,000 input tokens per minute.
at the end of the session the token usage and cost indicated was this :
Total cost: $0.95
Total duration (API): 1h 24m 22.8s
Total duration (wall): 1h 43m 3.5s
Total code changes: 746 lines added, 0 lines removed
Token usage by model:
claude-3-5-haiku: 18.8k input, 443 output, 0 cache read, 0 cache write
claude-sonnet: 10 input, 666 output, 0 cache read, 245.6k cache write
but the usage on the /usage page was 2,073,698 token in, and the cost on the /cost page was $7.90.
When looking at the costs csv file, it seems that it is the "input cache write 5m" that consumed nearly all the tokens ( $7,71 ). Is it a bug ?