def _main():
# Set random seed for training
seed = 8964
tf.set_random_seed(seed)
np.random.seed(seed)
random.seed(seed)
pp = pprint.PrettyPrinter()
print_flags_dict = {}
for key in FLAGS.__flags.keys():
print_flags_dict[key] = getattr(FLAGS, key)
pp.pprint(print_flags_dict)
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
assert os.path.exists(FLAGS.train_dir), "Specify training data directory!"
assert os.path.exists(FLAGS.val_dir), "Specify validation data directory!"
trl = Learner()
trl.train(FLAGS)
def run_second(cfg):
start_time = time.time()
try:
if cfg.hardware.seed.get('secondary', None) is None:
seed = random.randint(0, sys.maxint)
cfg.hardware.seed.secondary = seed
random.seed(cfg.hardware.seed.secondary)
data = datafile.load_data(cfg.hardware.primaryfile)
date, primcfg, history = data['date'], data['cfg'], data['history']
cfg.update(primcfg, overwrite=False)
primcfg = primcfg.copy(deep=True)
bot = robots.build_robot(cfg.sim)
# this should not change
assert cfg.goals.motor.m_feats == bot.m_feats, 'expected {}, got {}'.format(
cfg.goals.motor.m_feats, bot.m_feats)
assert cfg.goals.motor.m_bounds == bot.m_bounds
# sensors might change
cfg.goals.effect.s_feats = bot.s_feats
cfg.goals.effect.s_bounds = bot.s_bounds
## Configuration ##
cfg.goals.effect.res = 10
## Bias loading ##
grid_expl = GridExplorer(primcfg.goals.effect.s_feats,
cfg=primcfg.goals)
dataset = []
for order, goal, _, effect in history:
effect = [
min(b_max, max(b_min, e_i))
for e_i, (b_min,
b_max) in zip(effect, primcfg.goals.effect.s_bounds)
]
grid_expl.add_effect(effect, goal=goal)
dataset.append((order, effect))
bs = bias.IMBias(dataset, effectexplorer=grid_expl)
## Instanciation of core modules ##
#s_explorer = bias.GridBias(bs, cfg.effect.cfg.interest.max_deriv, s_feats, cfg = cfg.effect.cfg)
s_explorer = BoundedRandomExplorer(bot.s_feats, cfg.goals)
m_babble = MotorBabble(bot.m_feats, bot.m_bounds, cfg.goals)
m_explorer = bias.MotorBias(bs, m_babble, cfg.goals)
guide = Guide(bot.m_feats,
bot.s_feats,
bot.m_bounds,
cfg.goals,
goalexplorer=s_explorer,
motorbabble=m_explorer)
learner = Learner(bot.m_feats,
bot.s_feats,
bot.m_bounds,
fwd='ES-LWLR',
inv='L-BFGS-B')
cfg.freeze()
print(cfg.makeReport())
## Running learning ##
for i in xrange(cfg.exp.trials):
exp.learn(bot, guide, learner)
gfx.print_progress(i + 1,
cfg.exp.trials,
prefix='{}learning... {}'.format(
gfx.purple, gfx.cyan),
suffix=gfx.end,
eta=1)
print ''
exp.save_data(cfg, guide)
except Exception:
import traceback
traceback.print_exc()
finally:
bot.close()
print("ran for {}s.".format(int(time.time() - start_time)))
def run_tests(cfg):
start_time = time.time()
try:
cfg.update(defaultcfg, overwrite=False)
data = datafile.load_data(cfg.hardware.datafile)
date, expcfg, history = data['date'], data['cfg'], data['history']
cfg.test.setdefault('max', value=len(history))
cfg.test.max = min(len(history), cfg.test.max)
cfg.update(expcfg, overwrite=False)
bot = robots.build_robot(cfg.sim)
assert cfg.goals.effect.s_feats == bot.s_feats
assert cfg.goals.effect.s_bounds == bot.s_bounds
assert cfg.goals.motor.m_feats == bot.m_feats
assert cfg.goals.motor.m_bounds == bot.m_bounds
learner = Learner(bot.m_feats,
bot.s_feats,
bot.m_bounds,
fwd='ES-LWLR',
inv='L-BFGS-B')
if cfg.test.testsetfile is None:
testset = exp.create_tests(bot.s_bounds,
cfg.test.res,
fixed_dim=cfg.test.fixed_dim)
else:
testset = datafile.load_testset(cfg.test.testsetfile)
print(cfg.makeReport())
ticks, results, averages, stds = [], [], [], []
for counter, (order, goal, _, effect) in enumerate(history):
if (counter) % cfg.test.freq == 0:
ticks.append(counter)
result = exp.test_results(bot, learner, testset)
avg, std = exp.test_perfs(testset, result)
results.append(result)
averages.append(avg)
stds.append(std)
print("{}{:5d}: {}{:6.4f}{}".format(gfx.purple, counter,
gfx.cyan, avg, gfx.end))
if counter > cfg.test.max:
break
learner.add_xy(order, effect)
exp.save_test(cfg, date, tuple(testset), tuple(ticks), tuple(results),
tuple(averages), tuple(stds))
except Exception:
import traceback
traceback.print_exc()
finally:
bot.close()
print("ran for {}s.".format(int(time.time() - start_time)))
def run_first(cfg):
start_time = time.time()
try:
if cfg.hardware.seed.get('primary', None) is None:
seed = random.randint(0, sys.maxint)
cfg.hardware.seed.primary = seed
random.seed(cfg.hardware.seed.primary)
bot = robots.build_robot(cfg.sim)
## Configuration ##
goalscfg = treedict.TreeDict()
goalscfg.effect.explorer = BoundedRandomExplorer
cfg.goals.update(goalscfg, overwrite = False)
cfg.goals.effect.s_feats = s_feats = bot.s_feats
cfg.goals.effect.s_bounds = s_bounds = bot.s_bounds
cfg.goals.motor.m_feats = m_feats = bot.m_feats
cfg.goals.motor.m_bounds = m_bounds = bot.m_bounds
## Instanciation of core modules ##
s_explorer = BoundedRandomExplorer(s_feats, cfg.goals)
if 'explorename' in cfg.goals.effect:
if cfg.goals.effect.explorername == 'cluster':
s_explorer = cluster.ClusterExplorer(s_feats, cfg.goals)
elif cfg.goals.effect.explorername == 'cluster_im':
s_explorer = cluster.ClusterIMExplorer(s_feats, cfg.goals)
m_explorer = MotorBabble(bot.m_feats, bot.m_bounds, cfg.goals)
guide = Guide(m_feats, s_feats, m_bounds, cfg.goals,
goalexplorer = s_explorer, motorbabble = m_explorer)
learner = Learner(m_feats, s_feats, m_bounds, fwd = 'ES-LWLR', inv = 'L-BFGS-B')
cfg.freeze()
print(cfg.makeReport())
## Runing learning ##
for i in xrange(cfg.exp.trials):
exp.learn(bot, guide, learner)
if not cfg.sim.verbose:
gfx.print_progress(i+1, cfg.exp.trials,
prefix = '{}learning... {}'.format(gfx.purple, gfx.cyan),
suffix = gfx.end, eta = 1)
print ''
exp.save_data(cfg, guide)
except Exception:
import traceback
traceback.print_exc()
finally:
bot.close()
print("ran for {}s.".format(int(time.time() - start_time)))
def create_learner(self):
learner = Learner(self, self.generate_id())
self.learners.append(learner)