def evaluate_saved_model():
args = parse_a2c_args()
#TROUVER COMMENT UTILISER LE GPU
device = torch.device("cpu") #"cuda" if torch.cuda.is_available() else
env = DoomEnvironment(args, is_train=True)
print(env.num_actions)
obs_shape = (3, args.screen_height, args.screen_width)
policy = CNNPolicy(obs_shape, args).to(device)
args.scenario_dir = "scenarios_transfer_learning/big_combs_test/"
checkpoint = torch.load("0.pth.tar",
map_location=lambda storage, loc: storage)
policy.load_state_dict(checkpoint['model'])
policy.eval()
assert args.model_checkpoint, 'No model checkpoint found'
assert os.path.isfile(
args.model_checkpoint), 'The model could not be loaded'
# This lambda stuff is required otherwise it will try and load on GPU
for i in range(1, 64): #for i in range(args.num_mazes_test):
env = DoomEnvironment(args, idx=i, is_train=False)
movie_name = 'videos/3 - 12-04-21 - base/big_combs_test{:0004}.mp4'.format(
i)
print('Creating movie {}'.format(movie_name))
make_movie(policy, env, movie_name, args)
def evaluate_saved_model():
args = parse_a2c_args()
#TROUVER COMMENT UTILISER LE GPU
device = torch.device("cpu") #"cuda" if torch.cuda.is_available() else
env = DoomEnvironment(args, is_train=True)
print(env.num_actions)
obs_shape = (3, args.screen_height, args.screen_width)
policy = CNNPolicy(obs_shape, args).to(device)
name = "Final on enclave"
args.scenario_dir = "scenarios_transfer_learning/scenes/"
checkpoint = torch.load("final.pth.tar",
map_location=lambda storage, loc: storage)
policy.load_state_dict(checkpoint['model'])
policy.eval()
assert args.model_checkpoint, 'No model checkpoint found'
assert os.path.isfile(
args.model_checkpoint), 'The model could not be loaded'
# This lambda stuff is required otherwise it will try and load on GPU
args.scenario = "custom_scenario000.cfg"
env = DoomEnvironment(args, is_train=False)
movie_name = '/home/adam/Bureau/Visuels/0 - Rollout faits main/{}.mp4'.format(
name)
print('Creating movie {}'.format(movie_name))
make_movie(policy, env, movie_name, args)
def evaluate_saved_model():
args = parse_a2c_args()
# TROUVER COMMENT UTILISER LE GPU
device = torch.device("cpu") # "cuda" if torch.cuda.is_available() else
env = DoomEnvironment(args, is_train=False)
print(env.num_actions)
obs_shape = (3, args.screen_height, args.screen_width)
policy = CNNPolicy(obs_shape, args).to(device)
results = []
for model in range(0, 2):
checkpoint = torch.load(str(model) + ".pth.tar",
map_location=lambda storage, loc: storage)
policy.load_state_dict(checkpoint['model'])
policy.eval()
assert args.model_checkpoint, 'No model checkpoint found'
assert os.path.isfile(
args.model_checkpoint), 'The model could not be loaded'
# This lambda stuff is required otherwise it will try and load on GPU
results.append([])
for i in range(args.num_mazes_test):
env = DoomEnvironment(args, idx=i, is_train=False)
results[model].append(get_results(policy, env, args))
print(i)
success_rate = 0
average_reward = 0
average_time = 0
for res in results[model]:
if res[1] < 525:
success_rate += 1
average_time += res[1]
average_reward += res[0]
if success_rate != 0:
average_time /= success_rate
success_rate /= args.num_mazes_test
average_reward /= args.num_mazes_test
print(success_rate, average_reward, average_time)
time_diff = 0
finished_levels = 0
for i in range(args.num_mazes_test):
if results[1][i][1] < 525:
finished_levels += 1
time_diff += results[1][i][1] - results[0][i][1]
print(time_diff / finished_levels)
def evaluate_saved_model():
args = parse_a2c_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
env = DoomEnvironment(args, is_train=True)
print(env.num_actions)
obs_shape = (3, args.screen_height, args.screen_width)
policy = CNNPolicy(obs_shape, args).to(device)
checkpoint = torch.load(args.model_checkpoint, map_location=lambda storage, loc: storage)
policy.load_state_dict(checkpoint['model'])
policy.eval()
assert args.model_checkpoint, 'No model checkpoint found'
assert os.path.isfile(args.model_checkpoint), 'The model could not be loaded'
# This lambda stuff is required otherwise it will try and load on GPU
for i in range(args.num_mazes_test):
env = DoomEnvironment(args, idx=i, is_train=True)
movie_name = 'videos/rollout_{:0004}.mp4'.format(i)
print('Creating movie {}'.format(movie_name))
make_movie(policy, env, movie_name, args)
else:
return obs
def step(self, actions):
new_obs = []
rewards = []
dones = []
infos = []
for action, pipe in zip(actions, self.parent_pipes):
pipe.send(action)
for pipe in self.parent_pipes:
obs, reward, done, info = pipe.recv()
new_obs.append(self.prep_obs(obs))
rewards.append(reward)
dones.append(done)
infos.append(info)
return np.stack(new_obs), rewards, dones, infos
if __name__ == '__main__':
args = parse_a2c_args()
args.scenario_dir = '../scenarios/'
mp_test_envs = MultiEnv(args.simulator, args.num_environments, 1, args)
mp_test_envs.reset()
def evaluate_saved_model():
args = parse_a2c_args()
args2 = parse_a2c_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
num_updates = int(
args.num_frames) // args.num_steps // args.num_environments
# Writer will output to ./runs/ directory by default
writer = torch.utils.tensorboard.SummaryWriter()
train_envs = MultiEnv(args.simulator,
args.num_environments,
args,
is_train=True)
# Création des environnements de test des niveaux classiques
args2.scenario_dir = "scenarios_transfer_learning/mazes_classic_test/"
args2.scenario = "custom_scenario_test{:003}.cfg"
classic_test_envs = MultiEnv(args.simulator,
args.num_environments,
args2,
is_train=False)
# Création des environnements de test des niveaux peignes
args2.scenario_dir = "scenarios_transfer_learning/little_combs_test/"
little_combs_test_envs = MultiEnv(args.simulator,
args.num_environments,
args2,
is_train=False)
args2.scenario_dir = "scenarios_transfer_learning/medium_combs_test/"
medium_combs_test_envs = MultiEnv(args.simulator,
args.num_environments,
args2,
is_train=False)
obs_shape = train_envs.obs_shape
policy = CNNPolicy(obs_shape, args).to(device)
agent = A2CAgent(policy,
args.hidden_size,
value_weight=args.value_loss_coef,
entropy_weight=args.entropy_coef,
num_steps=args.num_steps,
num_parallel=args.num_environments,
gamma=args.gamma,
lr=args.learning_rate,
opt_alpha=args.alpha,
opt_momentum=args.momentum,
max_grad_norm=args.max_grad_norm)
obs = little_combs_test_envs.reset()
num_checkpoints = 355
for j in range(num_checkpoints):
if j % 8 == 0:
checkpoint_filename = '/home/adam/Bureau/Transfer Learning/FINAL/checkpoint_{}.pth.tar'.format(
str(j + 1))
agent.load_model(checkpoint_filename)
total_num_steps = (j + 1) * args.num_environments * args.num_steps
mean_rewards_classic, game_times_classic = agent.evaluate(
classic_test_envs, j, total_num_steps)
mean_rewards_little, game_times_little = agent.evaluate(
little_combs_test_envs, j, total_num_steps)
mean_rewards_medium, game_times_medium = agent.evaluate(
medium_combs_test_envs, j, total_num_steps)
writer.add_scalar("Reward classic levels", mean_rewards_classic,
(j + 1) * 100)
writer.add_scalar("Reward little combs levels",
mean_rewards_little, (j + 1) * 100)
writer.add_scalar("Reward medium combs levels",
mean_rewards_medium, (j + 1) * 100)
print(j)
def train():
args = parse_a2c_args()
args2 = parse_a2c_args()
output_dir = initialize_logging(args)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
num_updates = int(
args.num_frames) // args.num_steps // args.num_environments
# Create the train and test environments with Multiple processes
train_envs = MultiEnv(args.simulator,
args.num_environments,
args,
is_train=True)
#Création des environnements de test des niveaux classiques
args2.scenario_dir = "scenarios_transfer_learning/mazes_classic_test/"
args2.scenario = "custom_scenario_test{:003}.cfg"
classic_test_envs = MultiEnv(args.simulator,
args.num_environments,
args2,
is_train=False)
#Création des environnements de test des niveaux peignes
args2.scenario_dir = "scenarios_transfer_learning/little_combs_test/"
little_combs_test_envs = MultiEnv(args.simulator,
args.num_environments,
args2,
is_train=False)
args2.scenario_dir = "scenarios_transfer_learning/medium_combs_test/"
medium_combs_test_envs = MultiEnv(args.simulator,
args.num_environments,
args2,
is_train=False)
test_envs = MultiEnv(args.simulator,
args.num_environments,
args,
is_train=False)
# Writer will output to ./runs/ directory by default
writer = torch.utils.tensorboard.SummaryWriter()
obs_shape = train_envs.obs_shape
# The agent's policy network and training algorithm A2C
policy = CNNPolicy(obs_shape, args).to(device)
agent = A2CAgent(policy,
args.hidden_size,
value_weight=args.value_loss_coef,
entropy_weight=args.entropy_coef,
num_steps=args.num_steps,
num_parallel=args.num_environments,
gamma=args.gamma,
lr=args.learning_rate,
opt_alpha=args.alpha,
opt_momentum=args.momentum,
max_grad_norm=args.max_grad_norm)
start_j = 0
if args.reload_model:
checkpoint_idx = args.reload_model.split(',')[1]
checkpoint_filename = '{}models/base_line.pth.tar'.format(output_dir)
agent.load_model(checkpoint_filename)
start_j = 0 #(int(checkpoint_idx) // args.num_steps // args.num_environments) + 1
obs = train_envs.reset()
start = time.time()
nb_of_saves = 0
for j in range(start_j, num_updates):
print("------", j / num_updates * 100, "-------")
# Test des performances du modèle
if not args.skip_eval and j % args.eval_freq == 0:
total_num_steps = (j + 1) * args.num_environments * args.num_steps
mean_rewards_classic, game_times_classic = agent.evaluate(
classic_test_envs, j, total_num_steps)
mean_rewards_little, game_times_little = agent.evaluate(
little_combs_test_envs, j, total_num_steps)
mean_rewards_medium, game_times_medium = agent.evaluate(
medium_combs_test_envs, j, total_num_steps)
# succes_classic = sum([1 if i!=525 else 0 for i in game_times_classic])/16
# succes_little = sum([1 if i!=525 else 0 for i in game_times_little])/16
# succes_medium = sum([1 if i!=525 else 0 for i in game_times_medium])/16
writer.add_scalar("Reward classic levels", mean_rewards_classic, j)
writer.add_scalar("Reward little combs levels",
mean_rewards_little, j)
writer.add_scalar("Reward medium combs levels",
mean_rewards_medium, j)
# writer.add_scalar("Success rate classic levels", succes_classic, j)
# writer.add_scalar("Success rate little combs levels", succes_little, j)
# writer.add_scalar("Success rate medium combs levels", succes_medium, j)
for step in range(args.num_steps):
action = agent.get_action(obs, step)
obs, reward, done, info = train_envs.step(action)
agent.add_rewards_masks(reward, done, step)
report = agent.update(obs)
if j % args.log_interval == 0:
end = time.time()
total_num_steps = (j + 1) * args.num_environments * args.num_steps
save_num_steps = (start_j) * args.num_environments * args.num_steps
FPS = int((total_num_steps - save_num_steps) / (end - start)),
logging.info(report.format(j, total_num_steps, FPS))
if j % args.model_save_rate == 0:
nb_of_saves += 1
agent.save_policy2(nb_of_saves, args, output_dir)
# cancel the env processes
train_envs.cancel()
test_envs.cancel()
def evaluate_saved_model(models, models_dir):
args = parse_a2c_args()
# TROUVER COMMENT UTILISER LE GPU
device = torch.device("cpu") # "cuda" if torch.cuda.is_available() else
#création des environnements
#création des environnements little-combs
little_comb_env = []
for i in range(50):
args.scenario_dir = "scenarios_transfer_learning/little_combs_test/"
args.scenario = "custom_scenario_test{:003}.cfg"
little_comb_env.append(DoomEnvironment(args, idx=i, is_train=False))
#création des environnements big-combs
big_comb_env = []
for i in range(50):
args.scenario_dir = "scenarios_transfer_learning/big_combs_test/"
args.scenario = "custom_scenario_test{:003}.cfg"
big_comb_env.append(DoomEnvironment(args, idx=i, is_train=False))
#création des environnements classic
classic_env = []
for i in range(50):
args.scenario_dir = "scenarios_transfer_learning/mazes_classic_test/"
args.scenario = "custom_scenario_test{:003}.cfg"
classic_env.append(DoomEnvironment(args, idx=i, is_train=False))
#création des environnements other levels
medium_comb_env = []
for i in range(16):
args.scenario_dir = "scenarios_transfer_learning/medium_combs_test/"
args.scenario = "custom_scenario_test{:003}.cfg"
medium_comb_env.append(DoomEnvironment(args, idx=i, is_train=False))
obs_shape = (3, args.screen_height, args.screen_width)
policy = CNNPolicy(obs_shape, args).to(device)
resultat = []
for model in models:
checkpoint = torch.load(models_dir + "/checkpoint_" + str(model) +
".pth.tar",
map_location=lambda storage, loc: storage)
policy.load_state_dict(checkpoint['model'])
policy.eval()
resultat.append(model)
assert args.model_checkpoint, 'No model checkpoint found'
assert os.path.isfile(
args.model_checkpoint), 'The model could not be loaded'
# This lambda stuff is required otherwise it will try and load on GPU
# evaluation sur les niveaux classiques
results = []
for i in range(50):
env = classic_env[i]
results.append(get_results(policy, env, args))
print("Classic levels evaluation is done")
success_rate = 0
average_reward = 0
for res in results:
if res[1] < 525:
success_rate += 1
average_reward += res[0]
success_rate /= args.num_mazes_test
average_reward /= args.num_mazes_test
resultat.append([success_rate, average_reward])
# evaluation sur les little combs
results = []
for i in range(50):
env = little_comb_env[i]
results.append(get_results(policy, env, args))
print("Little combs evaluation is done")
success_rate = 0
average_reward = 0
for res in results:
if res[1] < 525:
success_rate += 1
average_reward += res[0]
success_rate /= args.num_mazes_test
average_reward /= args.num_mazes_test
resultat.append([success_rate, average_reward])
# evaluation sur les big combs
results = []
for i in range(50):
env = big_comb_env[i]
results.append(get_results(policy, env, args))
print("Big combs evaluation is done")
success_rate = 0
average_reward = 0
for res in results:
if res[1] < 525:
success_rate += 1
average_reward += res[0]
success_rate /= args.num_mazes_test
average_reward /= args.num_mazes_test
resultat.append([success_rate, average_reward])
# evaluation sur les autres niveaux
results = []
for i in range(16):
env = medium_comb_env[i]
results.append(get_results(policy, env, args))
print("Other levels evaluation is done")
success_rate = 0
average_reward = 0
for res in results:
if res[1] < 525:
success_rate += 1
average_reward += res[0]
success_rate /= args.num_mazes_test
average_reward /= args.num_mazes_test
resultat.append([success_rate, average_reward])
print("Checkpoint " + str(model) + " has been evaluated")
print(resultat)
def train():
args = parse_a2c_args()
output_dir = initialize_logging(args)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
num_updates = int(args.num_frames) // args.num_steps // args.num_environments
# Create the train and test environments with Multiple processes
train_envs = MultiEnv(args.simulator, args.num_environments, args, is_train=True)
test_envs = MultiEnv(args.simulator, args.num_environments, args, is_train=False)
obs_shape = train_envs.obs_shape
# The agent's policy network and training algorithm A2C
policy = CNNPolicy(obs_shape, args).to(device)
agent = A2CAgent(policy,
args.hidden_size,
value_weight=args.value_loss_coef,
entropy_weight=args.entropy_coef,
num_steps=args.num_steps,
num_parallel=args.num_environments,
gamma=args.gamma,
lr=args.learning_rate,
opt_alpha=args.alpha,
opt_momentum=args.momentum,
max_grad_norm=args.max_grad_norm)
start_j = 0
if args.reload_model:
checkpoint_idx = args.reload_model.split(',')[1]
checkpoint_filename = '{}models/checkpoint_{}.pth.tar'.format(output_dir, checkpoint_idx)
agent.load_model(checkpoint_filename)
start_j = (int(checkpoint_idx) // args.num_steps // args.num_environments) + 1
obs = train_envs.reset()
start = time.time()
for j in range(start_j, num_updates):
if not args.skip_eval and j % args.eval_freq == 0:
total_num_steps = (j + 1) * args.num_environments * args.num_steps
mean_rewards, game_times = agent.evaluate(test_envs, j, total_num_steps)
logging.info(mean_rewards)
logging.info(game_times)
for step in range(args.num_steps):
action = agent.get_action(obs, step)
obs, reward, done, info = train_envs.step(action)
agent.add_rewards_masks(reward, done, step)
report = agent.update(obs)
if j % args.log_interval == 0:
end = time.time()
total_num_steps = (j + 1) * args.num_environments * args.num_steps
save_num_steps = (start_j) * args.num_environments * args.num_steps
FPS = int((total_num_steps - save_num_steps) / (end - start)),
logging.info(report.format(j, total_num_steps, FPS))
if j % args.model_save_rate == 0:
total_num_steps = (j + 1) * args.num_environments * args.num_steps
agent.save_policy(total_num_steps, args, output_dir)
# cancel the env processes
train_envs.cancel()
test_envs.cancel()