def __init__(self,
action_names,
gamma,
model,
metrics_to_score=None,
device=None) -> None:
super().__init__(action_names, gamma, model, metrics_to_score)
self._device = device
self.ope_dm_estimator = OPEstimatorAdapter(
DMEstimator(device=self._device))
self.ope_ips_estimator = OPEstimatorAdapter(
IPSEstimator(device=self._device))
self.ope_dr_estimator = OPEstimatorAdapter(
DoublyRobustEstimator(device=self._device))
self.ope_seq_dr_estimator = SequentialOPEstimatorAdapter(
SeqDREstimator(device=self._device), gamma, device=self._device)
self.ope_seq_weighted_dr_estimator = SequentialOPEstimatorAdapter(
SeqDREstimator(weighted=True, device=self._device),
gamma,
device=self._device,
)
self.ope_seq_magic_estimator = SequentialOPEstimatorAdapter(
MAGICEstimator(device=self._device), gamma)
def test_switch_equal_to_ips(self):
"""
Switch with tau set at the max value should be equal to IPS
"""
# Setting the base to 1 will cause all candidates to be the maximum threshold
SwitchEstimator.EXP_BASE = 1
switch = SwitchEstimator(rmax=1.0).evaluate(self.bandit_input)
ips = IPSEstimator().evaluate(self.bandit_input)
self.assertAlmostEqual(ips.estimated_reward, switch.estimated_reward)
def __init__(self,
action_names,
gamma,
model,
metrics_to_score=None,
device=None) -> None:
super().__init__(action_names, gamma, model, metrics_to_score)
self._device = device
self.ope_dm_estimator = OPEstimatorAdapter(
DMEstimator(device=self._device))
self.ope_ips_estimator = OPEstimatorAdapter(
IPSEstimator(device=self._device))
self.ope_dr_estimator = OPEstimatorAdapter(
DoublyRobustEstimator(device=self._device))
def test_seq2slate_eval_data_page(self):
"""
Create 3 slate ranking logs and evaluate using Direct Method, Inverse
Propensity Scores, and Doubly Robust.
The logs are as follows:
state: [1, 0, 0], [0, 1, 0], [0, 0, 1]
indices in logged slates: [3, 2], [3, 2], [3, 2]
model output indices: [2, 3], [3, 2], [2, 3]
logged reward: 4, 5, 7
logged propensities: 0.2, 0.5, 0.4
predicted rewards on logged slates: 2, 4, 6
predicted rewards on model outputted slates: 1, 4, 5
predicted propensities: 0.4, 0.3, 0.7
When eval_greedy=True:
Direct Method uses the predicted rewards on model outputted slates.
Thus the result is expected to be (1 + 4 + 5) / 3
Inverse Propensity Scores would scale the reward by 1.0 / logged propensities
whenever the model output slate matches with the logged slate.
Since only the second log matches with the model output, the IPS result
is expected to be 5 / 0.5 / 3
Doubly Robust is the sum of the direct method result and propensity-scaled
reward difference; the latter is defined as:
1.0 / logged_propensities * (logged reward - predicted reward on logged slate)
* Indicator(model slate == logged slate)
Since only the second logged slate matches with the model outputted slate,
the DR result is expected to be (1 + 4 + 5) / 3 + 1.0 / 0.5 * (5 - 4) / 3
When eval_greedy=False:
Only Inverse Propensity Scores would be accurate. Because it would be too
expensive to compute all possible slates' propensities and predicted rewards
for Direct Method.
The expected IPS = (0.4 / 0.2 * 4 + 0.3 / 0.5 * 5 + 0.7 / 0.4 * 7) / 3
"""
batch_size = 3
state_dim = 3
src_seq_len = 2
tgt_seq_len = 2
candidate_dim = 2
reward_net = FakeSeq2SlateRewardNetwork()
seq2slate_net = FakeSeq2SlateTransformerNet()
src_seq = torch.eye(candidate_dim).repeat(batch_size, 1, 1)
tgt_out_idx = torch.LongTensor([[3, 2], [3, 2], [3, 2]])
tgt_out_seq = src_seq[
torch.arange(batch_size).repeat_interleave(tgt_seq_len),
tgt_out_idx.flatten() - 2, ].reshape(batch_size, tgt_seq_len,
candidate_dim)
ptb = rlt.PreprocessedTrainingBatch(
training_input=rlt.PreprocessedRankingInput(
state=rlt.FeatureData(float_features=torch.eye(state_dim)),
src_seq=rlt.FeatureData(float_features=src_seq),
tgt_out_seq=rlt.FeatureData(float_features=tgt_out_seq),
src_src_mask=torch.ones(batch_size, src_seq_len, src_seq_len),
tgt_out_idx=tgt_out_idx,
tgt_out_probs=torch.tensor([0.2, 0.5, 0.4]),
slate_reward=torch.tensor([4.0, 5.0, 7.0]),
),
extras=rlt.ExtraData(
sequence_number=torch.tensor([0, 0, 0]),
mdp_id=np.array(["0", "1", "2"]),
),
)
edp = EvaluationDataPage.create_from_tensors_seq2slate(
seq2slate_net, reward_net, ptb.training_input, eval_greedy=True)
logger.info(
"---------- Start evaluating eval_greedy=True -----------------")
doubly_robust_estimator = OPEstimatorAdapter(DoublyRobustEstimator())
dm_estimator = OPEstimatorAdapter(DMEstimator())
ips_estimator = OPEstimatorAdapter(IPSEstimator())
switch_estimator = OPEstimatorAdapter(SwitchEstimator())
switch_dr_estimator = OPEstimatorAdapter(SwitchDREstimator())
doubly_robust = doubly_robust_estimator.estimate(edp)
inverse_propensity = ips_estimator.estimate(edp)
direct_method = dm_estimator.estimate(edp)
# Verify that Switch with low exponent is equivalent to IPS
switch_ips = switch_estimator.estimate(edp, exp_base=1)
# Verify that Switch with no candidates is equivalent to DM
switch_dm = switch_estimator.estimate(edp, candidates=0)
# Verify that SwitchDR with low exponent is equivalent to DR
switch_dr_dr = switch_dr_estimator.estimate(edp, exp_base=1)
# Verify that SwitchDR with no candidates is equivalent to DM
switch_dr_dm = switch_dr_estimator.estimate(edp, candidates=0)
logger.info(f"{direct_method}, {inverse_propensity}, {doubly_robust}")
avg_logged_reward = (4 + 5 + 7) / 3
self.assertAlmostEqual(direct_method.raw, (1 + 4 + 5) / 3, delta=1e-6)
self.assertAlmostEqual(direct_method.normalized,
direct_method.raw / avg_logged_reward,
delta=1e-6)
self.assertAlmostEqual(inverse_propensity.raw, 5 / 0.5 / 3, delta=1e-6)
self.assertAlmostEqual(
inverse_propensity.normalized,
inverse_propensity.raw / avg_logged_reward,
delta=1e-6,
)
self.assertAlmostEqual(doubly_robust.raw,
direct_method.raw + 1 / 0.5 * (5 - 4) / 3,
delta=1e-6)
self.assertAlmostEqual(doubly_robust.normalized,
doubly_robust.raw / avg_logged_reward,
delta=1e-6)
self.assertAlmostEqual(switch_ips.raw,
inverse_propensity.raw,
delta=1e-6)
self.assertAlmostEqual(switch_dm.raw, direct_method.raw, delta=1e-6)
self.assertAlmostEqual(switch_dr_dr.raw, doubly_robust.raw, delta=1e-6)
self.assertAlmostEqual(switch_dr_dm.raw, direct_method.raw, delta=1e-6)
logger.info(
"---------- Finish evaluating eval_greedy=True -----------------")
logger.info(
"---------- Start evaluating eval_greedy=False -----------------")
edp = EvaluationDataPage.create_from_tensors_seq2slate(
seq2slate_net, reward_net, ptb.training_input, eval_greedy=False)
doubly_robust_estimator = OPEstimatorAdapter(DoublyRobustEstimator())
dm_estimator = OPEstimatorAdapter(DMEstimator())
ips_estimator = OPEstimatorAdapter(IPSEstimator())
doubly_robust = doubly_robust_estimator.estimate(edp)
inverse_propensity = ips_estimator.estimate(edp)
direct_method = dm_estimator.estimate(edp)
self.assertAlmostEqual(
inverse_propensity.raw,
(0.4 / 0.2 * 4 + 0.3 / 0.5 * 5 + 0.7 / 0.4 * 7) / 3,
delta=1e-6,
)
self.assertAlmostEqual(
inverse_propensity.normalized,
inverse_propensity.raw / avg_logged_reward,
delta=1e-6,
)
logger.info(
"---------- Finish evaluating eval_greedy=False -----------------")
description="Read command line parameters.")
parser.add_argument("-p", "--parameters", help="Path to config file.")
args = parser.parse_args(sys.argv[1:])
with open(args.parameters, "r") as f:
params = json.load(f)
if "dataset" not in params:
raise Exception('Please define "dataset" in config file')
random.seed(1234)
np.random.seed(1234)
torch.random.manual_seed(1234)
dataset = UCIMultiClassDataset(params["dataset"])
log_trainer = LogisticRegressionTrainer()
log_epsilon = 0.1
tgt_trainer = SGDClassifierTrainer()
tgt_epsilon = 0.1
dm_trainer = DecisionTreeTrainer()
experiments = [(
(
DMEstimator(DecisionTreeTrainer()),
IPSEstimator(),
DoublyRobustEstimator(DecisionTreeTrainer()),
),
1000,
) for _ in range(100)]
evaluate_all(experiments, dataset, log_trainer, log_epsilon, tgt_trainer,
tgt_epsilon, 0)
logs = []
for i in range(num_epsidoes):
train_choices = random.sample(range(num_total_samples), num_sample)
samples = []
for i in train_choices:
context = MultiClassContext(i)
logged_action, logged_dist = log_policy(context)
logged_reward = log_model(context)[logged_action]
target_action, target_dist = target_policy(context)
samples.append(
LogSample(
context,
logged_action,
logged_dist,
logged_reward,
target_action,
target_dist,
))
logs.append(Log(samples))
input = BanditsEstimatorInput(action_space, logs, target_model, gt_model)
result = DMEstimator().evaluate(input)
logging.info(f"DM result: {result}")
result = IPSEstimator().evaluate(input)
logging.info(f"IPS result: {result}")
result = DoublyRobustEstimator().evaluate(input)
logging.info(f"DR result: {result}")