def __init__(self,
content,
display_size,
model_name='1006-1538806261.ckpt',
use_ppo_models=False):
pygame.init()
self.surface = pygame.display.set_mode(display_size, 0, 24)
pygame.display.set_caption('oculomotor')
self.retina = Retina()
self.lip = LIP()
self.vc = VC()
self.pfc = PFC()
self.fef = FEF()
self.bg = BG(model_name=model_name, use_saved_models=use_ppo_models)
self.sc = SC()
self.hp = HP()
self.cb = CB()
self.step = 0
self.episode = 0
self.agent = Agent(retina=self.retina,
lip=self.lip,
vc=self.vc,
pfc=self.pfc,
fef=self.fef,
bg=self.bg,
sc=self.sc,
hp=self.hp,
cb=self.cb)
self.env = Environment(content)
obs = self.env.reset()
self.last_image = obs['screen']
self.last_angle = obs['angle']
self.last_reward = 0
self.last_done = False
self.episode_reward = 0
self.font = pygame.font.Font(None, 20)
self.display_size = display_size
def __init__(self, content, display_size):
pygame.init()
self.surface = pygame.display.set_mode(display_size, 0, 24)
pygame.display.set_caption('oculomotor')
self.retina = Retina()
self.lip = LIP()
self.vc = VC()
self.pfc = PFC()
self.fef = FEF()
self.bg = BG()
self.sc = SC()
self.hp = HP()
self.cb = CB()
self.agent = Agent(
retina=self.retina,
lip=self.lip,
vc=self.vc,
pfc=self.pfc,
fef=self.fef,
bg=self.bg,
sc=self.sc,
hp=self.hp,
cb=self.cb
)
self.env = Environment(content)
self.pfc.load_model('data/pfc_task_detection.pth')
#self.bg.load_model('data/bg_rl.pth')
obs = self.env.reset()
self.last_image = obs['screen']
self.last_angle = obs['angle']
self.last_reward = 0
self.last_done = False
self.episode_reward = 0
self.font = pygame.font.Font(None, 20)
self.display_size = display_size
def calc_fps(self, content_type, with_agent):
content = self.get_content(content_type)
if with_agent:
agent = Agent(
retina=Retina(),
lip=LIP(),
vc=VC(),
pfc=PFC(),
fef=FEF(),
bg=BG(),
sc=SC(),
hp=HP(),
cb=CB()
)
env = Environment(content)
obs = env.reset()
reward = 0
done = False
step_size = 1000
step = 0
start = time.time()
for i in range(step_size):
if with_agent:
image, angle = obs['screen'], obs['angle']
dummy_action = agent(image, angle, reward, done)
dh = np.random.uniform(low=-0.05, high=0.05)
dv = np.random.uniform(low=-0.05, high=0.05)
action = np.array([dh, dv])
obs, reward, done, _ = env.step(action)
step += 1
if done:
obs = env.reset()
elapsed_time = time.time() - start
fps = step_size / elapsed_time
return fps
def __init__(self, content, display_size):
pygame.init()
self.surface = pygame.display.set_mode(display_size, 0, 24)
pygame.display.set_caption('oculomotor')
self.retina = Retina()
self.lip = LIP()
self.vc = VC()
self.pfc = PFC()
self.fef = FEF()
#self.bg = BG() # Train mode
self.bg = BG(train=False, backprop=False) # Test mode
self.bg.agent.load("./results/exp20_after") # Test mode
self.sc = SC()
self.hp = HP()
self.cb = CB()
self.agent = Agent(retina=self.retina,
lip=self.lip,
vc=self.vc,
pfc=self.pfc,
fef=self.fef,
bg=self.bg,
sc=self.sc,
hp=self.hp,
cb=self.cb)
self.env = Environment(content)
obs = self.env.reset()
self.last_image = obs['screen']
self.last_angle = obs['angle']
self.last_reward = 0
self.last_done = False
self.episode_reward = 0
self.font = pygame.font.Font(None, 20)
self.display_size = display_size
def evaluate(logger, log_path):
retina = Retina()
lip = LIP()
vc = VC()
pfc = PFC()
fef = FEF()
bg = BG(use_saved_models=True)
sc = SC()
hp = HP()
cb = CB()
agent = Agent(
retina=retina,
lip=lip,
vc=vc,
pfc=pfc,
fef=fef,
bg=bg,
sc=sc,
hp=hp,
cb=cb
)
#bg.load_model("model.pkl")
total_reward = 0
all_trial_results = []
for i, task in enumerate(tasks):
trial_results, task_reward = task.evaluate(agent)
all_trial_results += trial_results
total_reward += task_reward
logger.log("evaluation_reward", total_reward, i)
# Save result csv files
save_results(all_trial_results, log_path)
print("evaluation finished:")
logger.close()
def train(content, step_size, logger, model_name, use_ppo_models):
retina = Retina()
lip = LIP()
vc = VC()
pfc = PFC()
fef = FEF()
bg = BG(model_name, use_saved_models=use_ppo_models)
sc = SC()
hp = HP()
cb = CB()
agent = Agent(
retina=retina,
lip=lip,
vc=vc,
pfc=pfc,
fef=fef,
bg=bg,
sc=sc,
hp=hp,
cb=cb
)
env = Environment(content)
# If your training code is written inside BG module, please add model load code here like.
#
# bg.load_model("model.pkl")
#
# When runnning with Docker, directory under 'oculomotor/' is volume shared
# with the host, so you can load/save the model data at anywhere under 'oculomotor/' dir.
obs = env.reset()
reward = 0
done = False
episode_reward = 0
episode_count = 0
# Add initial reward log
logger.log("episode_reward", episode_reward, episode_count)
for i in range(step_size):
image, angle = obs['screen'], obs['angle']
# Choose action by the agent's decision
action = agent(image, angle, reward, done)
# Foward environment one step
obs, reward, done, _ = env.step(action)
episode_reward += reward
# TODO: remove this
done = done or i % 180 == 0
if done:
obs = env.reset()
print(
"\033[93m {} step, {} episode reward={} \033[0m".format(
i, episode_count, episode_reward
)
)
# Store log for tensorboard graph
episode_count += 1
logger.log("episode_reward", episode_reward, episode_count)
episode_reward = 0
# Plase add model save code as you like.
if i % 5000 == 0:
bg.save_model()
print("training finished")
logger.close()
def train(content, step_size, logger):
retina = Retina()
lip = LIP()
vc = VC()
pfc = PFC()
fef = FEF()
bg = BG(init_weight_path="./data/bg.pth")
sc = SC()
hp = HP()
cb = CB()
agent = Agent(retina=retina,
lip=lip,
vc=vc,
pfc=pfc,
fef=fef,
bg=bg,
sc=sc,
hp=hp,
cb=cb)
env = Environment(content)
# If your training code is written inside BG module, please add model load code here like.
#
# bg.load_model("model.pkl")
#
# When runnning with Docker, directory under 'oculomotor/' is volume shared
# with the host, so you can load/save the model data at anywhere under 'oculomotor/' dir.
obs = env.reset()
reward = 0
done = False
episode_reward = 0
episode_count = 0
step = 0
# Add initial reward log
logger.log("episode_reward", episode_reward, episode_count)
for i in range(step_size):
image, angle = obs['screen'], obs['angle']
# Choose action by the agent's decision
action = agent(image, angle, reward, done)
# Foward environment one step
obs, reward, done, _ = env.step(action)
episode_reward += reward
step += 1
if step % EPI_THRESHOLD == 0:
done = True
if done:
obs = env.reset()
print("episode reward={}".format(episode_reward))
# Store log for tensorboard graph
episode_count += 1
logger.log("episode_reward", episode_reward, episode_count)
episode_reward = 0
step = 0
# Plase add model save code as you like.
#
# if i % 10 == 0:
# bg.save_model("model.pkl")
print("training finished")
logger.close()
def train(content, step_size, logger, log_path):
starttime = time.time()
retina = Retina()
lip = LIP()
vc = VC()
pfc = PFC()
fef = FEF()
bg = BG(log_path=log_path)
sc = SC()
hp = HP()
cb = CB()
agent = Agent(retina=retina,
lip=lip,
vc=vc,
pfc=pfc,
fef=fef,
bg=bg,
sc=sc,
hp=hp,
cb=cb)
env = Environment(content)
# If your training code is written inside BG module, please add model load code here like.
#
# bg.load_model("model.pkl")
#
# When runnning with Docker, directory under 'oculomotor/' is volume shared
# with the host, so you can load/save the model data at anywhere under 'oculomotor/' dir.
obs = env.reset()
reward = 0
done = False
episode_reward = 0
episode_count = 0
# Add initial reward log
logger.log("episode_reward", episode_reward, episode_count)
#step_size = 10800
while True:
for i in range(step_size):
image, angle = obs['screen'], obs['angle']
# Choose action by the agent's decision
action = agent(image, angle, reward, done)
# Foward environment one step
obs, reward, done, _ = env.step(action)
episode_reward += reward
if done:
print("episode count={}".format(episode_count))
obs = env.reset()
print("episode reward={}".format(episode_reward))
# Store log for tensorboard graph
episode_count += 1
logger.log("episode_reward", episode_reward, episode_count)
episode_reward = 0
# Plase add model save code as you like.
bg.save_model(str(episode_count) + "model")
episode_count += 1
logger.log("episode_reward", episode_reward, episode_count)
episode_reward = 0
bg.save_model(str(episode_count) + "model")
print("training finished")
logger.close()
endtime = time.time()
print('whole time:', endtime - starttime)
