class Trainer:
def __init__(self, game_dir):
self.agent = CustomAgent()
self.env = Environment(game_dir)
def train(self):
self.start_time = time()
for epoch_no in range(1, self.agent.nb_epochs + 1):
for game_no in tqdm(range(len(self.env.games))):
obs, infos = self.env.reset()
self.agent.train()
scores = [0] * len(obs)
dones = [False] * len(obs)
steps = [0] * len(obs)
while not all(dones):
# Increase step counts.
steps = [
step + int(not done)
for step, done in zip(steps, dones)
]
commands = self.agent.act(obs, scores, dones, infos)
obs, scores, dones, infos = self.env.step(commands)
# Let the agent know the game is done.
self.agent.act(obs, scores, dones, infos)
score = sum(scores) / self.agent.batch_size
score, possible_points, percentage = get_points(
score, infos['extra.walkthrough'][0])
print('Score: {}/{}'.format(score, possible_points))
def train(game_files):
agent = CustomAgent()
requested_infos = agent.select_additional_infos()
_validate_requested_infos(requested_infos)
env_id = textworld.gym.register_games(game_files, requested_infos,
max_episode_steps=agent.max_nb_steps_per_episode,
name="training")
env_id = textworld.gym.make_batch(env_id, batch_size=agent.batch_size, parallel=True)
env = gym.make(env_id)
for epoch_no in range(1, agent.nb_epochs + 1):
stats = {
"scores": [],
"steps": [],
}
for game_no in tqdm(range(len(game_files))):
obs, infos = env.reset()
agent.train()
scores = [0] * len(obs)
dones = [False] * len(obs)
steps = [0] * len(obs)
while not all(dones):
# Increase step counts.
steps = [step + int(not done) for step, done in zip(steps, dones)]
commands = agent.act(obs, scores, dones, infos)
obs, scores, dones, infos = env.step(commands)
# Let the agent knows the game is done.
agent.act(obs, scores, dones, infos)
stats["scores"].extend(scores)
stats["steps"].extend(steps)
score = sum(stats["scores"]) / agent.batch_size
steps = sum(stats["steps"]) / agent.batch_size
print("Epoch: {:3d} | {:2.1f} pts | {:4.1f} steps".format(epoch_no, score, steps))
agent.finish(final=True)
def __init__(self, game_dir):
self.agent = CustomAgent()
self.env = Environment(game_dir)
def __init__(self, game_dir):
self.agent = CustomAgent(verbose=True)
self.env = Environment(game_dir)
# Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# even the metrics!
memory = SequentialMemory(limit=100, window_length=1)
random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=.15,
mu=0.,
sigma=.1)
# agent = DDPGAgent(nb_actions=nb_actions, actor=actor, critic=critic, critic_action_input=action_input,
# memory=memory, nb_steps_warmup_critic=100, nb_steps_warmup_actor=100,
# random_process=random_process, gamma=.99, target_model_update=1e-3, processor=Processor())
agent = CustomAgent(nb_actions=nb_actions,
actor=actor,
critic=critic,
critic_action_input=action_input,
memory=memory,
nb_steps_warmup_critic=100,
nb_steps_warmup_actor=100,
random_process=random_process,
gamma=.99,
target_model_update=1e-3,
processor=Processor())
agent.compile([Adam(lr=1e-4), Adam(lr=1e-3)], metrics=['mae'])
# agent.load_weights('ddpg_init_experiences_Ship_Env_()_()_weights.h5f')
steps = 20000
for i in range(10):
# agent.fit(env, nb_steps=steps, visualize=False, verbose=1, log_interval=5)
agent.fit(env,
nb_steps=steps,
visualize=False,
verbose=1,
import gym
from pathlib import Path
from pprint import pprint as print
import textworld
from custom_agent import CustomAgent
from train import _validate_requested_infos
# games
#train_dir = Path('/home/nik-96/Documents/datasets/tw_train_data/')
#games = [str(file.relative_to(".")) for file in train_dir.iterdir() if file.suffix == '.ulx'][0]
agent = CustomAgent()
requested_infos = agent.select_additional_infos()
_validate_requested_infos(requested_infos)
env_id = textworld.gym.register_games([
'/home/nik-96/Documents/datasets/tw_train_data/train/tw-cooking-recipe2+cut+drop+go12-XQ7oC7pxS1OPsy6D.ulx'
],
requested_infos,
max_episode_steps=agent.
max_nb_steps_per_episode,
name="training")
env_id = textworld.gym.make_batch(env_id,
batch_size=agent.batch_size,
parallel=True)
env = gym.make(env_id)
state, infos = env.reset()
print(infos)