Different results when running multiple jobs in parallel

[4Freye]

I discovered an error when setting n_jobs == anything other than -1. Notably the predictions are slightly different than when running only one job (the default).

Until this issue is resolved I reccommend that n_jobs be set to 1 when using cross_val_fit_predict, cross_val_fit, and cross_val_predict. Otherwise the predicted values may not be optimal.

from sklearn.metrics import mean_squared_error
import numpy as np
import pandas as pd

# create dataframe
np.random.seed(1)
df = pd.DataFrame(np.random.random((40,10)))
df['entity'] = np.repeat(['A','B','C', 'D'], 10)
df['time'] = list(range(10)) * 4
df.set_index(['entity','time'], inplace=True)

# specify model 
model = RandomForestRegressor(n_estimators=10, random_state=1)

# run panel split: initialize, generate test labels, and fit and predict on data with and without parallel
ps = PanelSplit(periods = pd.Series(df.index.get_level_values('time')), n_splits = 5)
pred_df = ps.gen_test_labels(df.iloc[:, 0].reset_index())
pred_df['pred'], _ = ps.cross_val_fit_predict(model, X= df.iloc[:, 1:], y=df.iloc[:, 0])
pred_df['pred_parallel'], _ = ps.cross_val_fit_predict(model, X= df.iloc[:, 1:], y=df.iloc[:, 0], n_jobs = -1)

# see if the predictions are the same in parallel and not. The output is False.
print(mean_squared_error(pred_df[0], pred_df['pred']) == mean_squared_error(pred_df[0], pred_df['pred_parallel']))```

[4Freye]

Closed as it was fixed when this issue was resolved