Spaces:
Sleeping
Sleeping
| from ddpg import Agent | |
| import gymnasium as gym | |
| import numpy as np | |
| import matplotlib.pyplot as plt | |
| import torch | |
| import argparse | |
| from captum.attr import (IntegratedGradients) | |
| class TrainingLoop: | |
| def __init__(self): | |
| pass | |
| def train(self): | |
| env = gym.make( | |
| "LunarLander-v2", | |
| continuous = True, | |
| gravity = -10.0, | |
| render_mode = None | |
| ) | |
| agent = Agent(alpha=0.000025, beta=0.00025, input_dims=[8], tau=0.001, env=env, batch_size=64, layer1_size=400, layer2_size=300, n_actions=4) | |
| agent.load_models() | |
| np.random.seed(0) | |
| score_history = [] | |
| for i in range(1000): | |
| done = False | |
| score = 0 | |
| obs, _ = env.reset() | |
| while not done: | |
| act = agent.choose_action(obs) | |
| new_state, reward, terminated, truncated, info = env.step(act) | |
| done = terminated or truncated | |
| agent.remember(obs, act, reward, new_state, int(done)) | |
| agent.learn() | |
| score += reward | |
| obs = new_state | |
| score_history.append(score) | |
| print("episode", i, "score %.2f" % score, "100 game average %.2f" % np.mean(score_history[-100:])) | |
| if i % 25 == 0: | |
| agent.save_models() | |
| def load_trained(self): | |
| env = gym.make( | |
| "LunarLanderContinuous-v2", | |
| render_mode = "human" | |
| ) | |
| agent = Agent(alpha=0.000025, beta=0.00025, input_dims=[8], tau=0.001, env=env, batch_size=64, layer1_size=400, layer2_size=300, n_actions=4) | |
| agent.load_models() | |
| np.random.seed(0) | |
| score_history = [] | |
| for i in range(50): | |
| done = False | |
| score = 0 | |
| obs, _ = env.reset() | |
| while not done: | |
| act = agent.choose_action(obs) | |
| new_state, reward, terminated, truncated, info = env.step(act) | |
| done = terminated or truncated | |
| score += reward | |
| obs = new_state | |
| score_history.append(score) | |
| print("episode", i, "score %.2f" % score, "100 game average %.2f" % np.mean(score_history[-100:])) | |
| # Model Explainability | |
| from captum.attr import (IntegratedGradients) | |
| def _collect_running_baseline_average(self, num_iterations: int) -> torch.Tensor: | |
| env = gym.make( | |
| "LunarLanderContinuous-v2", | |
| render_mode = None | |
| ) | |
| agent = Agent(alpha=0.000025, beta=0.00025, input_dims=[8], tau=0.001, env=env, batch_size=64, layer1_size=400, layer2_size=300, n_actions=4) | |
| agent.load_models() | |
| torch.manual_seed(0) | |
| sum_obs = torch.zeros(8) | |
| for i in range(num_iterations): | |
| done = False | |
| score = 0 | |
| obs, _ = env.reset() | |
| sum_obs += obs | |
| print(f"Baseline on interation #{i}: {obs}") | |
| while not done: | |
| act = agent.choose_action(obs, attribution=None, baseline=None) | |
| new_state, reward, terminated, truncated, info = env.step(act) | |
| done = terminated or truncated | |
| score += reward | |
| obs = new_state | |
| return sum_obs / num_iterations | |
| def explain_trained(self, option: str, num_iterations :int = 10) -> None: | |
| baseline_options = { | |
| "1": torch.zeros(8), | |
| "2": self._collect_running_baseline_average(num_iterations), | |
| } | |
| baseline = baseline_options[option] | |
| env = gym.make( | |
| "LunarLanderContinuous-v2", | |
| render_mode = "human" | |
| ) | |
| agent = Agent(alpha=0.000025, beta=0.00025, input_dims=[8], tau=0.001, env=env, batch_size=64, layer1_size=400, layer2_size=300, n_actions=4) | |
| agent.load_models() | |
| ig = IntegratedGradients(agent.actor) | |
| np.random.seed(0) | |
| score_history = [] | |
| for i in range(50): | |
| done = False | |
| score = 0 | |
| obs, _ = env.reset() | |
| while not done: | |
| act = agent.choose_action(obs, attribution=ig, baseline=baseline) | |
| new_state, reward, terminated, truncated, info = env.step(act) | |
| done = terminated or truncated | |
| score += reward | |
| obs = new_state | |
| score_history.append(score) | |
| print("episode", i, "score %.2f" % score, "100 game average %.2f" % np.mean(score_history[-100:])) | |