make_prediction.py

import os
import pickle
import time
import numpy as np

from nn import *
from boosting import *
from config import models_config, DIR
from data_preprocessing_utils import Preprocessor

PREPROCESSORS = ['models/preprocessor_3']  # must be "models/preprocessor_%d_%d"
MODELS = ['models/cb_3_0_2019-11-13-00-09.pth']
MODEL_TYPE = 'cb'

if __name__ == '__main__':
    for MODEL, PREPROCESSOR in zip(MODELS, PREPROCESSORS):
        print('Making prediciton for model %s with preprocessor %s' % (MODEL[7:], PREPROCESSOR[15:]))

        init_config = models_config[MODEL_TYPE]['init_config']

        preprocessor = pickle.load(open(os.path.join(DIR, PREPROCESSOR), 'rb'))

        if MODEL_TYPE == 'nn':
            scaler = Scaler(preprocessor)
            init_config['scaler'] = scaler
            TrainerClass = NetTrainer
        elif MODEL_TYPE == 'cb':
            TrainerClass = CatBoostTrainer

        trainer = TrainerClass(None, None, **init_config)
        trainer.load_model(os.path.join(DIR, MODEL))
        submission = np.array([]).reshape(-1, 2)
        t0 = time.time()
        submission = trainer.predict(preprocessor.df.loc[preprocessor.prod_idx], submission)
        print('Generated predictions in %d seconds' % (time.time() - t0))
        print('Samples prediction:', submission[10, 1], 'for row', submission[10, 0])
        print('Number of NaNs in prediction:', np.isnan(submission[:, 1]).sum())
        with open(os.path.join(DIR, 'results/prediction_%s.pkl' % MODEL[7:]), 'wb') as f:
            pickle.dump(submission, f)