OverflowError: math range error

[Shivaraj Mulimani]

I am trying train using MQ2007-list dataset.

with open('/home/shivaraj/Downloads/MQ2007-list/Fold1/train.txt') as trainfile,
open('/home/shivaraj/Downloads/MQ2007-list/Fold1/vali.txt') as valifile,
open('/home/shivaraj/Downloads/MQ2007-list/Fold1/test.txt') as evalfile:
TX, Ty, Tqids, T_ = pyltr.data.letor.read_dataset(trainfile)
VX, Vy, Vqids, V_ = pyltr.data.letor.read_dataset(valifile)
EX, Ey, Eqids, E_ = pyltr.data.letor.read_dataset(evalfile)

metric = pyltr.metrics.NDCG(k=10)

Only needed if you want to perform validation (early stopping & trimming)

monitor = pyltr.models.monitors.ValidationMonitor(
VX, Vy, Vqids, metric=metric, stop_after=250)

model = pyltr.models.LambdaMART(
metric=metric,
n_estimators=1000,
learning_rate=0.02,
max_features=0.5,
query_subsample=0.5,
max_leaf_nodes=10,
min_samples_leaf=64,
verbose=1,
)

model.fit(TX, Ty, Tqids, monitor=monitor)

This error log--

OverflowError Traceback (most recent call last)
in
16 )
17
---> 18 model.fit(TX, Ty, Tqids, monitor=monitor)

~/miniconda3/envs/smartDB/lib/python3.6/site-packages/pyltr/models/lambdamart.py in fit(self, X, y, qids, monitor)
199
200 n_stages = self.fit_stages(X, y, qids, y_pred,
--> 201 random_state, begin_at_stage, monitor)
202
203 if n_stages < self.estimators.shape[0]:

~/miniconda3/envs/smartDB/lib/python3.6/site-packages/pyltr/models/lambdamart.py in _fit_stages(self, X, y, qids, y_pred, random_state, begin_at_stage, monitor)
406 y_pred = self._fit_stage(i, X, y, qids, y_pred, sample_weight,
407 sample_mask, query_groups_to_use,
--> 408 random_state)
409
410 train_total_score, oob_total_score = 0.0, 0.0

~/miniconda3/envs/smartDB/lib/python3.6/site-packages/pyltr/models/lambdamart.py in _fit_stage(self, i, X, y, qids, y_pred, sample_weight, sample_mask, query_groups, random_state)
332 for qid, a, b, _ in query_groups:
333 lambdas, deltas = self._calc_lambdas_deltas(qid, y[a:b],
--> 334 y_pred[a:b])
335 all_lambdas[a:b] = lambdas
336 all_deltas[a:b] = deltas

~/miniconda3/envs/smartDB/lib/python3.6/site-packages/pyltr/models/lambdamart.py in _calc_lambdas_deltas(self, qid, y, y_pred)
267 actual = y[positions]
268
--> 269 swap_deltas = self.metric.calc_swap_deltas(qid, actual)
270 max_k = self.metric.max_k()
271 if max_k is None or ns < max_k:

~/miniconda3/envs/smartDB/lib/python3.6/site-packages/pyltr/metrics/dcg.py in calc_swap_deltas(self, qid, targets, coeff)
33 for j in range(i + 1, n_targets):
34 deltas[i, j] = coeff *
---> 35 (self._gain_fn(targets[i]) - self._gain_fn(targets[j])) *
36 (self._get_discount(j) - self._get_discount(i))
37

~/miniconda3/envs/smartDB/lib/python3.6/site-packages/pyltr/metrics/gains.py in _exp2_gain(x)
16
17 def _exp2_gain(x):
---> 18 return math.exp(x * _LOG2) - 1.0
19
20

OverflowError: math range error

[Jonathan Shore]

Did you find a workaround for this? I am having the same problem as well ...

[Jerry Ma]

Hello,

I don't have the full picture on what scores are included in the dataset; however, I'm guessing one of two possibilities:

• Missing values in the target (solution: drop querysets with missing target values)
• Large (1000+) values for the target (solution: gain_type='identity' in the NDCG constructor)

[Jonathan Shore]

One could restructure the problem to be in log space, avoiding the exponential. This is how many numerical algorithms deal with exponentials (and potential overflow).

[Jerry Ma]

That's indeed what the above bit of code does :)