def test_metrics(self):
dcg = DCGSlateMetric()
ndcg = NDCGSlateMetric(SlateItemValues([1.0, 2.5, 2.0, 3.0, 1.5, 0.0]))
item_rewards = SlateItemValues([2.0, 1.0, 0.0, 3.0, 1.5, 2.5])
slate = Slate([SlateItem(1), SlateItem(3), SlateItem(2)])
reward = dcg(slate.slots, slate.slot_values(item_rewards))
self.assertAlmostEqual(reward, 5.416508275)
reward = ndcg(slate.slots, slate.slot_values(item_rewards))
self.assertAlmostEqual(reward, 0.473547669)
slate = Slate([SlateItem(5), SlateItem(0), SlateItem(4)])
reward = dcg(slate.slots, slate.slot_values(item_rewards))
self.assertAlmostEqual(reward, 7.463857073)
reward = ndcg(slate.slots, slate.slot_values(item_rewards))
self.assertAlmostEqual(reward, 0.652540703)
log_queries = load_logged_queries(params["test_data"])
slots = SlateSlots(MAX_POSITION)
episodes = []
for qid, qs in sorted(log_queries.items(),
key=lambda i: len(i[1]),
reverse=True):
log_query = qs[0]
context = SlateContext(SlateQuery((qid, *(log_query.query_terms))),
slots)
log_item_rewards = log_training_dataset.predict_item(
log_query.query_id, log_query.query_terms)
log_item_probs = SlateItemProbabilities(log_item_rewards.values)
tgt_item_rewards = tgt_model.item_rewards(context)
tgt_item_probs = SlateItemProbabilities(tgt_item_rewards.values)
gt_item_rewards = gt_model.item_rewards(context)
metric = NDCGSlateMetric(gt_item_rewards)
samples = []
for q in qs:
slate = make_slate(slots, q.list)
samples.append(
LogSample(
slate,
slate.slot_values(gt_item_rewards),
SlateSlotValues(q.position_relevances),
))
episodes.append(
LogEpisode(
context,
metric,
samples,
None,
def evaluate(
experiments: Iterable[Tuple[Iterable[SlateEstimator], int]],
log_dataset: TrainingDataset,
log_distribution: RewardDistribution,
tgt_dataset: TrainingDataset,
tgt_distribution: RewardDistribution,
log_queries: Sequence[TrainingQuery],
slate_size: int,
item_size: int,
metric_func: str,
max_num_workers: int,
device=None,
):
log_length = len(log_queries)
slots = SlateSlots(slate_size)
logging.info("Generating log...")
st = time.perf_counter()
tasks = []
total_samples = 0
for estimators, num_samples in experiments:
samples = []
if num_samples * 10 > log_length:
logging.warning(f"not enough log data, needs {num_samples * 10}")
continue
query_choices = np.random.choice(log_length,
num_samples,
replace=False)
for i in query_choices:
q = log_queries[i]
context = SlateContext(SlateQuery((q.query_id, *(q.query_terms))),
slots)
url_relevances = q.url_relevances
if len(url_relevances) > item_size:
url_relevances = {
k: v
for k, v in sorted(url_relevances.items(),
key=lambda item: item[1])[:item_size]
}
items = url_relevances.keys()
log_item_rewards = log_dataset.item_relevances(
q.query_id, q.query_terms, items)
log_item_probs = log_distribution(log_item_rewards)
tgt_item_rewards = tgt_dataset.item_relevances(
q.query_id, q.query_terms, items)
tgt_item_probs = tgt_distribution(tgt_item_rewards)
tgt_slot_expectation = tgt_item_probs.slot_item_expectations(slots)
gt_item_rewards = SlateItemValues(url_relevances)
if metric_func == "dcg":
metric = DCGSlateMetric(device=device)
elif metric_func == "err":
metric = ERRSlateMetric(4.0, device=device)
else:
metric = NDCGSlateMetric(gt_item_rewards, device=device)
slot_weights = metric.slot_weights(slots)
if tgt_item_probs.is_deterministic:
tgt_slate_prob = 1.0
log_slate = tgt_item_probs.sample_slate(slots)
else:
tgt_slate_prob = float("nan")
log_slate = log_item_probs.sample_slate(slots)
log_slate_prob = log_item_probs.slate_probability(log_slate)
log_rewards = log_slate.slot_values(gt_item_rewards)
log_reward = metric.calculate_reward(slots, log_rewards, None,
slot_weights)
gt_slot_rewards = tgt_slot_expectation.expected_rewards(
gt_item_rewards)
gt_reward = metric.calculate_reward(slots, gt_slot_rewards, None,
slot_weights)
samples.append(
LogSample(
context,
metric,
log_slate,
log_reward,
log_slate_prob,
None,
log_item_probs,
tgt_slate_prob,
None,
tgt_item_probs,
gt_reward,
slot_weights,
))
total_samples += 1
tasks.append((estimators, SlateEstimatorInput(samples)))
dt = time.perf_counter() - st
logging.info(f"Generating log done: {total_samples} samples in {dt}s")
logging.info("start evaluating...")
st = time.perf_counter()
evaluator = Evaluator(tasks, max_num_workers)
Evaluator.report_results(evaluator.evaluate())
logging.info(f"evaluating done in {time.perf_counter() - st}s")
def evalute_all(
dataset: MSLRDatasets,
slate_size: int,
log_trainer: Trainer,
tgt_trainer: Trainer,
tgt_deterministic: bool,
num_episodes: int,
num_samples: int,
):
print(
f"Run: {log_trainer.name}, {tgt_trainer.name}"
f"[{'deterministic' if tgt_deterministic else 'stochastic'}]",
flush=True,
)
logging.info("Preparing models and policies...")
st = time.process_time()
log_trainer.load_model(
os.path.join(dataset.folder,
log_trainer.name + "_anchor_url_features.pickle"))
log_pred = log_trainer.predict(dataset.anchor_url_features)
log_model = TrainedModel(log_pred.scores)
log_policy = MSLRPolicy(log_pred.scores, False, 1.0)
tgt_trainer.load_model(
os.path.join(dataset.folder,
tgt_trainer.name + "_body_features.pickle"))
tgt_pred = tgt_trainer.predict(dataset.body_features)
tgt_model = TrainedModel(tgt_pred.scores)
tgt_policy = MSLRPolicy(tgt_pred.scores, tgt_deterministic, 1.0)
dt = time.process_time() - st
logging.info(f"Preparing models and policies done: {dt}s")
logging.info("Generating log...")
st = time.process_time()
slots = SlateSlots(slate_size)
queries = dataset.queries
episodes = []
for q in queries:
query = SlateQuery(q)
items = SlateItems([SlateItem(i) for i in range(q[2].item())])
if len(items) < slate_size:
logging.warning(f"Number of items ({len(items)}) less than "
f"number of slots ({slate_size})")
continue
context = SlateContext(query, slots, items)
log_item_probs = log_policy(context)
log_item_rewards = log_model.item_rewards(context)
tgt_item_probs = tgt_policy(context)
metric = NDCGSlateMetric(log_item_rewards)
samples = []
for _ in range(num_samples):
slate = log_item_probs.sample_slate(slots)
samples.append(
LogSample(slate, slate.slot_values(log_item_rewards)))
episodes.append(
LogEpisode(context, metric, samples, None, log_item_probs, None,
tgt_item_probs))
if len(episodes) >= num_episodes:
break
dt = time.process_time() - st
logging.info(f"Generating log done: {len(episodes)} samples in {dt}s")
input = SlateEstimatorInput(episodes, tgt_model, log_model)
evaluate(DMEstimator(device=device), input)
def evaluate(
experiments: Iterable[Tuple[Iterable[SlateEstimator], int]],
dataset: MSLRDatasets,
slate_size: int,
item_size: int,
metric_func: str,
log_trainer: Trainer,
log_distribution: RewardDistribution,
log_features: str,
tgt_trainer: Trainer,
tgt_distribution: RewardDistribution,
tgt_features: str,
dm_features: str,
max_num_workers: int,
device=None,
):
assert slate_size < item_size
print(
f"Evaluate All:"
f" slate_size={slate_size}, item_size={item_size}, metric={metric_func}"
f", Log=[{log_trainer.name}, {log_distribution}, {log_features}]"
f", Target=[{tgt_trainer.name}, {tgt_distribution}, {tgt_features}]"
f", DM=[{dm_features}]"
f", Workers={max_num_workers}, device={device}",
flush=True,
)
logging.info("Preparing models and policies...")
st = time.perf_counter()
log_trainer.load_model(
os.path.join(dataset.folder,
log_trainer.name + "_all_" + log_features + ".pickle"))
# calculate behavior model scores
log_pred = log_trainer.predict(getattr(dataset, log_features))
tgt_trainer.load_model(
os.path.join(dataset.folder,
tgt_trainer.name + "_all_" + tgt_features + ".pickle"))
# calculate target model scores
tgt_pred = tgt_trainer.predict(getattr(dataset, tgt_features))
dm_train_features = getattr(dataset, dm_features)
slots = SlateSlots(slate_size)
dt = time.perf_counter() - st
logging.info(f"Preparing models and policies done: {dt}s")
total_samples = 0
for _, num_samples in experiments:
total_samples += num_samples
logging.info(f"Generating log: total_samples={total_samples}")
st = time.perf_counter()
tasks = []
samples_generated = 0
total_queries = dataset.queries.shape[0]
for estimators, num_samples in experiments:
samples = []
for _ in range(num_samples):
# randomly sample a query
q = dataset.queries[random.randrange(total_queries)]
doc_size = int(q[2])
if doc_size < item_size:
# skip if number of docs is less than item_size
continue
si = int(q[1])
ei = si + doc_size
# using top item_size docs for logging
log_scores, item_choices = log_pred.scores[si:ei].sort(
dim=0, descending=True)
log_scores = log_scores[:item_size]
item_choices = item_choices[:item_size]
log_item_probs = log_distribution(SlateItemValues(log_scores))
tgt_scores = tgt_pred.scores[si:ei][item_choices].detach().clone()
tgt_item_probs = tgt_distribution(SlateItemValues(tgt_scores))
tgt_slot_expectation = tgt_item_probs.slot_item_expectations(slots)
gt_item_rewards = SlateItemValues(
dataset.relevances[si:ei][item_choices])
gt_rewards = tgt_slot_expectation.expected_rewards(gt_item_rewards)
if metric_func == "dcg":
metric = DCGSlateMetric(device=device)
elif metric_func == "err":
metric = ERRSlateMetric(4.0, device=device)
else:
metric = NDCGSlateMetric(gt_item_rewards, device=device)
query = SlateQuery((si, ei))
context = SlateContext(query, slots, item_choices)
slot_weights = metric.slot_weights(slots)
gt_reward = metric.calculate_reward(slots, gt_rewards, None,
slot_weights)
if tgt_item_probs.is_deterministic:
tgt_slate_prob = 1.0
log_slate = tgt_item_probs.sample_slate(slots)
log_reward = gt_reward
else:
tgt_slate_prob = float("nan")
log_slate = log_item_probs.sample_slate(slots)
log_rewards = log_slate.slot_values(gt_item_rewards)
log_reward = metric.calculate_reward(slots, log_rewards, None,
slot_weights)
log_slate_prob = log_item_probs.slate_probability(log_slate)
item_features = SlateItemFeatures(
dm_train_features[si:ei][item_choices])
sample = LogSample(
context,
metric,
log_slate,
log_reward,
log_slate_prob,
None,
log_item_probs,
tgt_slate_prob,
None,
tgt_item_probs,
gt_reward,
slot_weights,
None,
item_features,
)
samples.append(sample)
samples_generated += 1
if samples_generated % 1000 == 0:
logging.info(
f" samples generated: {samples_generated}, {100 * samples_generated / total_samples:.1f}%"
)
tasks.append((estimators, SlateEstimatorInput(samples)))
dt = time.perf_counter() - st
logging.info(f"Generating log done: {total_samples} samples in {dt}s")
logging.info("start evaluating...")
st = time.perf_counter()
evaluator = Evaluator(tasks, max_num_workers)
Evaluator.report_results(evaluator.evaluate())
logging.info(f"evaluating done in {time.perf_counter() - st}s")