Spaces:
Sleeping
Sleeping
File size: 4,840 Bytes
569299e ec3a146 f6f3371 ec3a146 569299e f6f3371 ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 569299e ec3a146 f6f3371 569299e ec3a146 569299e ec3a146 569299e f6f3371 569299e f6f3371 ec3a146 569299e f6f3371 ec3a146 f6f3371 ec3a146 569299e ec3a146 569299e f6f3371 ec3a146 569299e f6f3371 ec3a146 569299e ec3a146 569299e f6f3371 ec3a146 569299e ec3a146 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 |
from ddpg import Agent
import gymnasium as gym
import numpy as np
import matplotlib.pyplot as plt
import torch
from captum.attr import (IntegratedGradients)
class TrainingLoop:
def __init__(self, env_spec, output_path='./output/', seed=0, **kwargs):
assert env_spec in gym.envs.registry.keys()
defaults = {
"continuous": True,
"gravity": -10.0,
"render_mode": None
}
defaults.update(**kwargs)
self.env = gym.make(
env_spec,
**defaults
)
torch.manual_seed(seed)
self.agent = None
self.output_path = output_path
# TODO: spec-to-hyperparameters look-up
def create_agent(self, alpha=0.000025, beta=0.00025, input_dims=[8], tau=0.001, batch_size=64, layer1_size=400, layer2_size=300, n_actions=4):
self.agent = Agent(alpha=alpha, beta=beta, input_dims=input_dims, tau=tau, env=self.env, batch_size=batch_size, layer1_size=layer1_size, layer2_size=layer2_size, n_actions=n_actions)
def train(self):
assert self.agent is not None
self.agent.load_models()
score_history = []
for i in range(1000):
done = False
score = 0
obs, _ = self.env.reset()
while not done:
act = self.agent.choose_action(obs)
new_state, reward, terminated, truncated, info = self.env.step(act)
done = terminated or truncated
self.agent.remember(obs, act, reward, new_state, int(done))
self.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:
self.agent.save_models()
self.env.close()
def load_trained(self):
assert self.agent is not None
self.agent.load_models()
score_history = []
for i in range(50):
done = False
score = 0
obs, _ = self.env.reset()
while not done:
act = self.agent.choose_action(obs)
new_state, reward, terminated, truncated, info = self.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:]))
self.env.close()
# Model Explainability
from captum.attr import (IntegratedGradients)
def _collect_running_baseline_average(self, num_iterations: int) -> torch.Tensor:
assert self.agent is not None
print("--------- Collecting running baseline average ----------")
self.agent.load_models()
sum_obs = torch.zeros(8)
for i in range(num_iterations):
done = False
score = 0
obs, _ = self.env.reset()
sum_obs += obs
# print(f"Baseline on interation #{i}: {obs}")
while not done:
act = self.agent.choose_action(obs, baseline=None)
new_state, reward, terminated, truncated, info = self.env.step(act)
done = terminated or truncated
score += reward
obs = new_state
print(f"Baseline collected: {sum_obs / num_iterations}")
self.env.close()
return sum_obs / num_iterations
def explain_trained(self, option: str, num_iterations :int = 10) -> None:
assert self.agent is not None
baseline_options = {
"1": torch.zeros(8),
"2": self._collect_running_baseline_average(num_iterations),
}
baseline = baseline_options[option]
print("\n\n\n\n--------- Performing Attributions -----------")
self.agent.load_models()
print(self.agent.actor)
ig = IntegratedGradients(self.agent.actor)
self.agent.ig = ig
score_history = []
for i in range(50):
done = False
score = 0
obs, _ = self.env.reset()
while not done:
act = self.agent.choose_action(observation=obs, baseline=baseline)
new_state, reward, terminated, truncated, info = self.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:]))
self.env.close()
return self.agent.attributions
|