validation_data=(val_x, val_y),
epochs=NUM_EPOCHS,
verbose=0,
callbacks=[ExponentialMovingAverage()])
test_df = std_scaler.transform(test_df)
raw_val_prob = model.predict(val_x)
raw_test_prob = model.predict(test_df)
test_pred = np.expm1(raw_test_prob)
return results.history, test_pred, val_idx, raw_val_prob, raw_test_prob
if __name__ == '__main__':
start_time = time.time()
preprocess_dict = preprocess_csv()
(train_ids, test_ids, processed_train_df, processed_test_df) = [
preprocess_dict[key] for key in
['train_ids', 'test_ids', 'processed_train', 'processed_test']
]
folds = split_to_folds(processed_train_df,
num_folds=NUM_FOLDS,
seed=SEED,
shuffle=False)
all_fold_results = []
all_fold_preds = []
with Pool(NUM_FOLDS) as p:
combined_results = p.starmap(
train_fold, ((curr_fold, processed_train_df, processed_test_df)
test_df_1.astype('float32') *
np.random.normal(1., scale=(0.1 / 0.9), size=test_df_1.shape),
test_rolling.astype('float32') * np.random.normal(
1., scale=(0.1 / 0.9), size=test_rolling.shape)
])[0])
raw_test_prob = np.mean(np.stack(agg_test_prob, axis=0), axis=0)
test_pred = np.expm1(np.dot(raw_test_prob, supports))
return results.history, test_pred, val_idx, raw_val_prob, raw_test_prob
if __name__ == '__main__':
start_time = time.time()
preprocess_dict = preprocess_csv(rolling_macro={
'min_unique': 20,
'lookback_period': 12,
'monthly_resampling': True
})
(train_ids, test_ids, processed_train_df, processed_test_df, train_rolling,
test_rolling) = [
preprocess_dict[key] for key in [
'train_ids', 'test_ids', 'processed_train', 'processed_test',
'train_rolling', 'test_rolling'
]
]
# generate distribution labels
generate_target_partial = partial(generate_target_dist,
num_bins=NUM_BINS,
low=LOW,