Bug fix: small values of max_bin cause program to crash

[btrotta]

Fixes a minor bug where setting max_bin < 4 can cause lightgbm to crash. Below is an example to reproduce the bug. I think the problem is that in bin.cpp, the method FindMaxBinWithZeroAsOneBin sometimes attempts to create more bins than allowed by max_bin.

import numpy as np
import lightgbm as lgb

np.random.seed(0)
y = np.random.choice([0, 1], 100)
X = np.zeros((100, 1))
X[:30, 0] = -1
X[30:60, 0] = 1
X[60:, 0] = 2
X[0] = np.nan
params = {'objective': 'binary', 'seed': 0, 'min_data_in_leaf': 1, 'max_bin': 2}  # causes crash in training
#params = {'objective': 'binary', 'seed': 0, 'min_data_in_leaf': 1, 'max_bin': 3}  # works normally
lgb_x = lgb.Dataset(X, label=y)
est = lgb.train(params, lgb_x)

# another example, max_bin is 3 and X contains nan
X[0, 0] = np.nan
params = {'objective': 'binary', 'seed': 0, 'min_data_in_leaf': 1, 'max_bin': 3}  # causes crash
#params = {'objective': 'binary', 'seed': 0, 'min_data_in_leaf': 1, 'max_bin': 4}  # works
lgb_x = lgb.Dataset(X, label=y)
est = lgb.train(params, lgb_x)

[btrotta]

I have changed an existing test to account for the new binning behavior. The method FindMaxBinWithZeroAsOneBin now will always create a zero bin if max_bin >=3. (I assume this is the intended behavior, based on the method name; let me know if not.) This means it needs to have bin upper bounds -kZeroThreshold, kZeroThreshold, inf even if there are no negative values. This reduces the number of distinct predicted values in the test I modified.

Here is a description of the current behavior of FindMaxBinWithZeroAsOneBin whenmax_bin <= 3:

• If max_bin = 1, returns [inf]
• If max_bin = 2 and there are no negative data values, returns [kZeroThreshold, inf]
• if max_bin = 2 and there is at least one negative data value, returns [-kZeroThreshold, inf]
• if max_bin = 3, returns [-kZeroThreshold, kZeroThreshold, inf]