def test1():
n_types = 10
n_labels = 4
print
print '# test sequence labeler on mod data with LOLS'
data = macarico.util.make_sequence_mod_data(20, 6, n_types, n_labels)
data = [Example(x, y, n_labels) for x, y in data]
tRNN = TransitionRNN([RNNFeatures(n_types)], [AttendAt()], n_labels)
policy = LinearPolicy(tRNN, n_labels)
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
p_rollin_ref = stochastic(ExponentialAnnealing(0.9))
p_rollout_ref = stochastic(ExponentialAnnealing(0.9))
macarico.util.trainloop(
training_data=data[:len(data) // 2],
dev_data=data[len(data) // 2:],
policy=policy,
learning_alg=lambda ex: LOLS.lols(ex, HammingLoss,
HammingLossReference(), policy,
p_rollin_ref, p_rollout_ref),
losses=HammingLoss(),
optimizer=optimizer,
run_per_epoch=[p_rollin_ref.step, p_rollout_ref.step],
train_eval_skip=1,
)
def test2():
# aggrevate
print
print '# test sequence labeler on mod data with AggreVaTe'
n_types = 10
n_labels = 4
data = macarico.util.make_sequence_mod_data(100, 5, n_types, n_labels)
data = [Example(x, y, n_labels) for x, y in data]
tRNN = TransitionRNN(
[RNNFeatures(n_types)],
[AttendAt()],
n_labels,
)
policy = LinearPolicy(tRNN, n_labels)
p_rollin_ref = stochastic(ExponentialAnnealing(0.99))
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
macarico.util.trainloop(
training_data=data[:len(data) // 2],
dev_data=data[len(data) // 2:],
policy=policy,
Learner=lambda: AggreVaTe(HammingLossReference(), policy, p_rollin_ref
),
losses=HammingLoss(),
optimizer=optimizer,
run_per_epoch=[p_rollin_ref.step],
n_epochs=4,
train_eval_skip=1,
)
def test0():
print
print '# test sequence labeler on mod data with DAgger'
n_types = 10
n_labels = 4
data = [
Example(x, y, n_labels)
for x, y in macarico.util.make_sequence_mod_data(
100, 5, n_types, n_labels)
]
tRNN = Actor([RNNFeatures(n_types, output_field='mytok_rnn')],
[AttendAt(field='mytok_rnn')], n_labels)
policy = LinearPolicy(tRNN, n_labels)
p_rollin_ref = stochastic(ExponentialAnnealing(0.99))
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
macarico.util.trainloop(
training_data=data[:len(data) // 2],
dev_data=data[len(data) // 2:],
policy=policy,
Learner=lambda: DAgger(HammingLossReference(), policy, p_rollin_ref),
losses=HammingLoss(),
optimizer=optimizer,
run_per_epoch=[p_rollin_ref.step],
n_epochs=4,
train_eval_skip=1,
)
def test_wsj():
print
print '# test on wsj subset'
from macarico.data import nlp_data
tr,de,te,vocab,label_id = \
nlp_data.read_wsj_pos('data/wsj.pos', n_tr=50, n_de=50, n_te=0)
n_types = len(vocab)
n_labels = len(label_id)
print 'n_train: %s, n_dev: %s, n_test: %s' % (len(tr), len(de), len(te))
print 'n_types: %s, n_labels: %s' % (n_types, n_labels)
tRNN = TransitionRNN([RNNFeatures(n_types, rnn_type='RNN')], [AttendAt()],
n_labels)
policy = LinearPolicy(tRNN, n_labels)
p_rollin_ref = stochastic(ExponentialAnnealing(0.9))
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
macarico.util.trainloop(
training_data=tr,
dev_data=de,
policy=policy,
Learner=lambda: DAgger(HammingLossReference(), policy, p_rollin_ref),
# Learner = lambda: MaximumLikelihood(HammingLossReference(), policy),
losses=HammingLoss(),
optimizer=optimizer,
run_per_epoch=[p_rollin_ref.step],
n_epochs=10,
# train_eval_skip = None,
)
def test1(learning_method, exploration):
print
print '# testing learning_method=%d exploration=%d' % (learning_method,
exploration)
print
n_types = 10
n_labels = 4
data = macarico.util.make_sequence_mod_data(100, 6, n_types, n_labels)
data = [Example(x, y, n_labels) for x, y in data]
tRNN = TransitionRNN([RNNFeatures(n_types)], [AttendAt()], n_labels)
policy = LinearPolicy(tRNN, n_labels)
optimizer = torch.optim.Adam(policy.parameters(), lr=0.001)
p_rollin_ref = stochastic(ExponentialAnnealing(0.9))
p_rollout_ref = stochastic(ExponentialAnnealing(0.99999))
macarico.util.trainloop(
training_data=data[:len(data) // 2],
dev_data=data[len(data) // 2:],
policy=policy,
Learner=lambda: BanditLOLS(
HammingLossReference(),
policy,
p_rollin_ref,
p_rollout_ref,
learning_method, # LEARN_IPS, LEARN_DR, LEARN_BIASED
exploration,
),
losses=HammingLoss(),
optimizer=optimizer,
run_per_epoch=[p_rollin_ref.step, p_rollout_ref.step],
train_eval_skip=10,
)
def test_restore(n_types, n_labels, data, model):
actor = TransitionRNN([RNNFeatures(n_types)], [AttendAt()], n_labels)
policy = LinearPolicy(actor, n_labels)
print 'evaluating new model: %g' % \
macarico.util.evaluate(data, policy, HammingLoss())
policy.load_state_dict(model)
print 'evaluating restored model: %g' % \
macarico.util.evaluate(data, policy, HammingLoss())
def test1(LEARNER=LearnerOpts.DAGGER):
print
print 'Running test 1 with learner=%s' % LEARNER
print '======================================================='
n_states = 3
n_actions = 2
tRNN = TransitionRNN([mdp.MDPFeatures(n_states, noise_rate=0.5)],
[AttendAt(lambda _: 0, 's')], n_actions)
policy = LinearPolicy(tRNN, n_actions)
p_rollin_ref = stochastic(ExponentialAnnealing(0.99))
p_rollout_ref = stochastic(ExponentialAnnealing(1))
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
test_mdp, pi_ref = make_ross_mdp()
if LEARNER == LearnerOpts.DAGGER:
learner = lambda: DAgger(pi_ref, policy, p_rollin_ref)
elif LEARNER == LearnerOpts.TWISTED:
learner = lambda: TwistedDAgger(pi_ref, policy, p_rollin_ref)
elif LEARNER == LearnerOpts.MAXLIK:
learner = lambda: MaximumLikelihood(pi_ref, policy)
elif LEARNER == LearnerOpts.AGGREVATE:
learner = lambda: AggreVaTe(pi_ref, policy, p_rollin_ref)
elif LEARNER == LearnerOpts.LOLS:
learner = None
losses = []
for epoch in xrange(101):
optimizer.zero_grad()
if learner is not None:
l = learner()
env = test_mdp.mk_env()
res = env.run_episode(l)
loss = mdp.MDPLoss()(test_mdp, env)
l.update(loss)
elif LEARNER == LearnerOpts.LOLS:
lols(test_mdp, mdp.MDPLoss, pi_ref, policy, p_rollin_ref,
p_rollout_ref)
optimizer.step()
p_rollin_ref.step()
p_rollout_ref.step()
env = test_mdp.mk_env()
res = env.run_episode(policy)
loss = mdp.MDPLoss()(test_mdp, env)
losses.append(loss)
if epoch % 20 == 0:
print epoch, sum(losses[-100:]) / len(losses[-100:]), '\t', res
def run_gridworld(ex, actor):
policy = LinearPolicy(actor(), 4)
baseline = EWMA(0.8)
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
losses = []
for epoch in xrange(2001):
optimizer.zero_grad()
learner = Reinforce(policy, baseline)
env = ex.mk_env()
res = env.run_episode(learner)
loss = GridLoss()(ex, env)
losses.append(loss)
if epoch % 100 == 0:
print sum(losses[-10:]) / len(losses[-10:]), '\t', res
learner.update(loss)
optimizer.step()
def build_policy_bag(features_bag, n_actions, loss_fn, n_layers, hidden_dim):
return [
LinearPolicy(features,
n_actions,
loss_fn=loss_fn,
n_layers=n_layers,
hidden_dim=hidden_dim) for features in features_bag
]
def run_environment(ex, actor, lossfn, rl_alg=None, n_epochs=201, lr=0.01):
if rl_alg is None:
baseline = EWMA(0.8)
rl_alg = lambda policy: Reinforce(policy, baseline)
policy = LinearPolicy(actor(), ex.n_actions, n_layers=1)
optimizer = torch.optim.Adam(policy.parameters(), lr=lr)
losses = []
for epoch in xrange(n_epochs):
optimizer.zero_grad()
learner = rl_alg(policy)
#learner = AdvantageActorCritic(policy, baseline)
env = ex.mk_env()
res = env.run_episode(learner) # , epoch % 5000 == 0)
loss = lossfn(ex, env)
losses.append(loss)
if epoch % 20 == 0:
print(epoch, '\t',
sum(losses[-500:]) / len(losses[-500:]), '\t', res)
learner.update(loss)
optimizer.step()
def run_train(n_types, n_labels, data):
actor = TransitionRNN([RNNFeatures(n_types)], [AttendAt()], n_labels)
policy = LinearPolicy(actor, n_labels)
print 'training'
_, model = macarico.util.trainloop(
training_data=data[:len(data) // 2],
dev_data=data[len(data) // 2:],
policy=policy,
Learner=lambda: MaximumLikelihood(HammingLossReference(), policy),
losses=HammingLoss(),
optimizer=torch.optim.Adam(policy.parameters(), lr=0.01),
n_epochs=2,
train_eval_skip=1,
returned_parameters='best',
)
print 'evaluating learned model: %g' % \
macarico.util.evaluate(data, policy, HammingLoss())
policy.load_state_dict(model)
print 'evaluating learned model: %g' % \
macarico.util.evaluate(data, policy, HammingLoss())
return model
def test1(use_bootstrap):
n_types = 10
n_labels = 4
print
print '# test sequence labeler on mod data with Reslope and', (
'bootstrap' if use_bootstrap else 'boltzmann'), 'exploration'
data = macarico.util.make_sequence_mod_data(3000, 6, n_types, n_labels)
data = [Example(x, y, n_labels) for x, y in data]
if not use_bootstrap:
tRNN = TransitionRNN([RNNFeatures(n_types)], [AttendAt()], n_labels)
policy = LinearPolicy(tRNN, n_labels)
else:
rnns = [
TransitionRNN([RNNFeatures(n_types)], [AttendAt()],
n_labels,
h_name='h%d' % i) for i in xrange(5)
]
policy = BootstrapPolicy(rnns, n_labels)
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
p_ref = stochastic(ExponentialAnnealing(0.9))
macarico.util.trainloop(
training_data = data[:2048],
dev_data = data[2048:],
policy = policy,
Learner = lambda: Reslope(HammingLossReference(), policy, p_ref,
exploration=BanditLOLS.EXPLORE_BOOTSTRAP if use_bootstrap else \
BanditLOLS.EXPLORE_BOLTZMANN
),
losses = HammingLoss(),
optimizer = optimizer,
run_per_epoch = [p_ref.step],
train_eval_skip = 1,
bandit_evaluation = True,
n_epochs = 1,
)
def run_ppo(ex, actor, loss_fn, eps, learner_type):
print(learner_type)
print('Eps: ', eps)
policy = LinearPolicy(actor(), ex.n_actions, n_layers=1)
baseline = EWMA(0.8)
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
losses = []
n_episodes = 10000
for episode in range(n_episodes):
dy.renew_cg()
if learner_type == 'ppo':
learner = PPO(policy, baseline, eps)
elif learner_type == 'reinforce':
learner = Reinforce(policy, baseline)
env = ex.mk_env()
env.run_episode(learner)
loss = loss_fn(ex, env)
losses.append(loss)
if episode % 5 == 0:
print('episode: ', episode, 'loss:',
sum(losses[-500:]) / len(losses[-500:]))
learner.update(loss)
optimizer.update()
def test1(task=0, LEARNER=LearnerOpts.DAGGER):
print
print 'Running test 1 (v%d) with learner=%s' % (task, LEARNER)
print '======================================================='
if task == 0:
print 'Sequence reversal task, easy version'
data = macarico.util.make_sequence_reversal_data(100, 5, 5)
foci = [AttendAt(lambda s: s.N - s.n - 1)]
elif task == 1:
print 'Sequence reversal task, hard version'
data = macarico.util.make_sequence_reversal_data(1000, 5, 5)
foci = [AttendAt()]
elif task == 2:
print 'Sequence reversal task, multi-focus version'
data = macarico.util.make_sequence_reversal_data(100, 5, 5)
foci = [AttendAt(), AttendAt(lambda s: s.N - s.n - 1)]
elif task == 3:
print 'Memoryless task, add-one mod K'
data = macarico.util.make_sequence_mod_data(50, 5, 10, 3)
foci = [AttendAt()]
elif task == 4:
print 'Matti-style data'
data = make_matti_data(1000, 20, 2, 0.05)
foci = [AttendAt()]
n_types = 1 + max({x for X, _ in data for x in X})
n_labels = 1 + max({y for _, Y in data for y in Y})
data = [Example(x, y, n_labels) for x, y in data]
random.shuffle(data)
m = len(data) // 2
train = data[:m]
dev = data[m:]
print 'n_train: %s, n_dev: %s' % (len(train), len(dev))
print 'n_types: %s, n_labels: %s' % (n_types, n_labels)
print 'learner:', LEARNER
print
tRNN = Actor([RNNFeatures(n_types)], foci, n_labels)
policy = LinearPolicy(tRNN, n_labels)
baseline = EWMA(0.8)
p_rollin_ref = stochastic(ExponentialAnnealing(0.5))
p_rollout_ref = stochastic(ExponentialAnnealing(0.5))
if LEARNER == LearnerOpts.AC:
from macarico.lts.reinforce import AdvantageActorCritic, LinearValueFn
baseline = LinearValueFn(policy.features)
policy.vfa = baseline # adds params to policy via nn.module
optimizer = torch.optim.Adam(policy.parameters(), lr=0.01)
if LEARNER == LearnerOpts.DAGGER:
learner = lambda: DAgger(HammingLossReference(), policy, p_rollin_ref)
elif LEARNER == LearnerOpts.TWISTED:
learner = lambda: TwistedDAgger(HammingLossReference(), policy,
p_rollin_ref)
elif LEARNER == LearnerOpts.MAXLIK:
learner = lambda: MaximumLikelihood(HammingLossReference(), policy)
elif LEARNER == LearnerOpts.AC:
learner = lambda: AdvantageActorCritic(policy, baseline)
elif LEARNER == LearnerOpts.REINFORCE:
learner = lambda: Reinforce(policy, baseline)
elif LEARNER == LearnerOpts.BANDITLOLS:
learner = lambda: BanditLOLS(HammingLossReference(
), policy, p_rollin_ref, p_rollout_ref, BanditLOLS.LEARN_DR, BanditLOLS
.EXPLORE_UNIFORM, baseline)
macarico.util.trainloop(
training_data=train,
dev_data=dev,
policy=policy,
Learner=learner,
losses=HammingLoss(),
optimizer=optimizer,
run_per_epoch=[p_rollin_ref.step, p_rollout_ref.step],
n_epochs=10,
train_eval_skip=1,
)