examples/keras_spark_rossmann_run.py [182:282]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - df = df.withColumn('Promo2Days', F.when(df.Promo2SinceYear > 1900, F.greatest(F.lit(0), F.least(F.lit(25 * 7), F.datediff(df.Date, df.Promo2Since)))) .otherwise(0)) df = df.withColumn('Promo2Weeks', (df.Promo2Days / 7).cast(T.IntegerType())) # Check that we did not lose any rows through inner joins. assert num_rows == df.count(), 'lost rows in joins' return df def build_vocabulary(df, cols): vocab = {} for col in cols: values = [r[0] for r in df.select(col).distinct().collect()] col_type = type([x for x in values if x is not None][0]) default_value = col_type() vocab[col] = sorted(values, key=lambda x: x or default_value) return vocab def cast_columns(df, cols): for col in cols: df = df.withColumn(col, F.coalesce(df[col].cast(T.FloatType()), F.lit(0.0))) return df def lookup_columns(df, vocab): def lookup(mapping): def fn(v): return mapping.index(v) return F.udf(fn, returnType=T.IntegerType()) for col, mapping in vocab.items(): df = df.withColumn(col, lookup(mapping)(df[col])) return df if args.sample_rate: train_csv = train_csv.sample(withReplacement=False, fraction=args.sample_rate) test_csv = test_csv.sample(withReplacement=False, fraction=args.sample_rate) # Prepare data frames from CSV files. train_df = prepare_df(train_csv).cache() test_df = prepare_df(test_csv).cache() # Add elapsed times from holidays & promos, the data spanning training & test datasets. elapsed_cols = ['Promo', 'StateHoliday', 'SchoolHoliday'] elapsed = add_elapsed(train_df.select('Date', 'Store', *elapsed_cols) .unionAll(test_df.select('Date', 'Store', *elapsed_cols)), elapsed_cols) # Join with elapsed times. train_df = train_df \ .join(elapsed, ['Date', 'Store']) \ .select(train_df['*'], *[prefix + col for prefix in ['Before', 'After'] for col in elapsed_cols]) test_df = test_df \ .join(elapsed, ['Date', 'Store']) \ .select(test_df['*'], *[prefix + col for prefix in ['Before', 'After'] for col in elapsed_cols]) # Filter out zero sales. train_df = train_df.filter(train_df.Sales > 0) print('===================') print('Prepared data frame') print('===================') train_df.show() categorical_cols = [ 'Store', 'State', 'DayOfWeek', 'Year', 'Month', 'Day', 'Week', 'CompetitionMonthsOpen', 'Promo2Weeks', 'StoreType', 'Assortment', 'PromoInterval', 'CompetitionOpenSinceYear', 'Promo2SinceYear', 'Events', 'Promo', 'StateHoliday', 'SchoolHoliday' ] continuous_cols = [ 'CompetitionDistance', 'Max_TemperatureC', 'Mean_TemperatureC', 'Min_TemperatureC', 'Max_Humidity', 'Mean_Humidity', 'Min_Humidity', 'Max_Wind_SpeedKm_h', 'Mean_Wind_SpeedKm_h', 'CloudCover', 'trend', 'trend_DE', 'BeforePromo', 'AfterPromo', 'AfterStateHoliday', 'BeforeStateHoliday', 'BeforeSchoolHoliday', 'AfterSchoolHoliday' ] all_cols = categorical_cols + continuous_cols # Select features. train_df = train_df.select(*(all_cols + ['Sales', 'Date'])).cache() test_df = test_df.select(*(all_cols + ['Id', 'Date'])).cache() # Build vocabulary of categorical columns. vocab = build_vocabulary(train_df.select(*categorical_cols) .unionAll(test_df.select(*categorical_cols)).cache(), categorical_cols) # Cast continuous columns to float & lookup categorical columns. train_df = cast_columns(train_df, continuous_cols + ['Sales']) train_df = lookup_columns(train_df, vocab) test_df = cast_columns(test_df, continuous_cols) test_df = lookup_columns(test_df, vocab) # Split into training & validation. # Test set is in 2015, use the same period in 2014 from the training set as a validation set. test_min_date = test_df.agg(F.min(test_df.Date)).collect()[0][0] test_max_date = test_df.agg(F.max(test_df.Date)).collect()[0][0] - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - examples/keras_spark_rossmann_estimator.py [187:287]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - df = df.withColumn('Promo2Days', F.when(df.Promo2SinceYear > 1900, F.greatest(F.lit(0), F.least(F.lit(25 * 7), F.datediff(df.Date, df.Promo2Since)))) .otherwise(0)) df = df.withColumn('Promo2Weeks', (df.Promo2Days / 7).cast(T.IntegerType())) # Check that we did not lose any rows through inner joins. assert num_rows == df.count(), 'lost rows in joins' return df def build_vocabulary(df, cols): vocab = {} for col in cols: values = [r[0] for r in df.select(col).distinct().collect()] col_type = type([x for x in values if x is not None][0]) default_value = col_type() vocab[col] = sorted(values, key=lambda x: x or default_value) return vocab def cast_columns(df, cols): for col in cols: df = df.withColumn(col, F.coalesce(df[col].cast(T.FloatType()), F.lit(0.0))) return df def lookup_columns(df, vocab): def lookup(mapping): def fn(v): return mapping.index(v) return F.udf(fn, returnType=T.IntegerType()) for col, mapping in vocab.items(): df = df.withColumn(col, lookup(mapping)(df[col])) return df if args.sample_rate: train_csv = train_csv.sample(withReplacement=False, fraction=args.sample_rate) test_csv = test_csv.sample(withReplacement=False, fraction=args.sample_rate) # Prepare data frames from CSV files. train_df = prepare_df(train_csv).cache() test_df = prepare_df(test_csv).cache() # Add elapsed times from holidays & promos, the data spanning training & test datasets. elapsed_cols = ['Promo', 'StateHoliday', 'SchoolHoliday'] elapsed = add_elapsed(train_df.select('Date', 'Store', *elapsed_cols) .unionAll(test_df.select('Date', 'Store', *elapsed_cols)), elapsed_cols) # Join with elapsed times. train_df = train_df \ .join(elapsed, ['Date', 'Store']) \ .select(train_df['*'], *[prefix + col for prefix in ['Before', 'After'] for col in elapsed_cols]) test_df = test_df \ .join(elapsed, ['Date', 'Store']) \ .select(test_df['*'], *[prefix + col for prefix in ['Before', 'After'] for col in elapsed_cols]) # Filter out zero sales. train_df = train_df.filter(train_df.Sales > 0) print('===================') print('Prepared data frame') print('===================') train_df.show() categorical_cols = [ 'Store', 'State', 'DayOfWeek', 'Year', 'Month', 'Day', 'Week', 'CompetitionMonthsOpen', 'Promo2Weeks', 'StoreType', 'Assortment', 'PromoInterval', 'CompetitionOpenSinceYear', 'Promo2SinceYear', 'Events', 'Promo', 'StateHoliday', 'SchoolHoliday' ] continuous_cols = [ 'CompetitionDistance', 'Max_TemperatureC', 'Mean_TemperatureC', 'Min_TemperatureC', 'Max_Humidity', 'Mean_Humidity', 'Min_Humidity', 'Max_Wind_SpeedKm_h', 'Mean_Wind_SpeedKm_h', 'CloudCover', 'trend', 'trend_DE', 'BeforePromo', 'AfterPromo', 'AfterStateHoliday', 'BeforeStateHoliday', 'BeforeSchoolHoliday', 'AfterSchoolHoliday' ] all_cols = categorical_cols + continuous_cols # Select features. train_df = train_df.select(*(all_cols + ['Sales', 'Date'])).cache() test_df = test_df.select(*(all_cols + ['Id', 'Date'])).cache() # Build vocabulary of categorical columns. vocab = build_vocabulary(train_df.select(*categorical_cols) .unionAll(test_df.select(*categorical_cols)).cache(), categorical_cols) # Cast continuous columns to float & lookup categorical columns. train_df = cast_columns(train_df, continuous_cols + ['Sales']) train_df = lookup_columns(train_df, vocab) test_df = cast_columns(test_df, continuous_cols) test_df = lookup_columns(test_df, vocab) # Split into training & validation. # Test set is in 2015, use the same period in 2014 from the training set as a validation set. test_min_date = test_df.agg(F.min(test_df.Date)).collect()[0][0] test_max_date = test_df.agg(F.max(test_df.Date)).collect()[0][0] - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -