add eval_race.py for race competition evaluation

Author: gferraro2019

duckiesim/rl/eval_race.py

@@ -0,0 +1,82 @@
+import torch
+import gymnasium as gym
+import numpy as np
+from munchausen import QNetwork
+
+# import plotext as plt
+import cv2
+
+
+def evaluate_with_rendering(model_path, env_id, num_episodes=10, seed=0, tau_soft=1.0):
+    """
+    Évalue le modèle sauvegardé avec rendu visuel, en utilisant une politique softmax.
+
+    Args:
+        model_path (str): Le chemin vers le fichier du modèle sauvegardé.
+        env_id (str): L'identifiant de l'environnement.
+        num_episodes (int): Le nombre d'épisodes à évaluer.
+        seed (int): La graine initiale pour reproduire les résultats.
+        tau_soft (float): Température pour le softmax.
+    """
+    # Charger le modèle
+    env = gym.make(env_id)
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model = QNetwork(env.action_space, env.observation_space).to(device)
+    model.load_state_dict(torch.load(model_path))
+    model.eval()
+    avg_counter_checkpts =0
+    avg_steps=0
+    avg_return =0
+    # Évaluation
+    for episode in range(num_episodes):
+        obs, _ = env.reset(seed=seed + episode)
+        # obs, _ = env.reset(seed=58)
+
+        done = False
+        total_reward = 0
+
+        print(f"\nEpisode {episode + 1}/{num_episodes}")
+
+        while not done:
+            obs_tensor = torch.Tensor(obs).unsqueeze(0).to(device)
+
+            with torch.no_grad():
+                q_values = model(obs_tensor)
+                policy = torch.softmax(q_values / tau_soft, dim=-1)
+                # print('policy', policy)
+                # argmax
+                action = torch.argmax(policy, dim=-1).item()
+                # action = torch.multinomial(policy, num_samples=1).item()
+                # print('action', action)
+            img = env.render()
+            cv2.imshow("Duckietown", img)
+            cv2.waitKey(1)
+            obs, reward, done, _, _ = env.step(action)
+            # print('obs', obs.shape)
+            total_reward += reward
+            print(f"step {env.unwrapped.episodic_length} reward {reward}")
+        
+        avg_counter_checkpts += env.env.env.env.race.counter_checkpt
+        avg_steps += env.env.env.env.episodic_length
+        avg_return += total_reward
+
+        print(f"Récompense totale de l'épisode : {total_reward}")
+
+    print(f"the Avg. N. Checkpoints:{avg_counter_checkpts/num_episodes}")
+    print(f"the Avg. N. Steps:{avg_steps/num_episodes}")
+    print(f"the Avg. Return:{avg_return/num_episodes}")
+
+
+    env.close()
+
+
+if __name__ == "__main__":
+    # model_path /home/dcas/g.ferraro/Desktop/models= "/home/p.le-tolguenec/Documents/duckietown_rl_course/model/exp_1/munchausen_430404_1746.0135750578204.pt"  # Adapter au chemin réel
+    model_path = (
+        "/home/dcas/g.ferraro/Desktop/models/munchausen_1000000_1648.556805028793.pt"
+    )
+
+    env_id = "DuckietownDiscrete-v0"
+
+    evaluate_with_rendering(model_path, env_id, num_episodes=20, seed=42, tau_soft=0.05)