def identify_normalization_parameters(
table_spec: TableSpec,
column_name: str,
preprocessing_options: PreprocessingOptions,
seed: Optional[int] = None,
) -> Dict[int, NormalizationParameters]:
""" Get normalization parameters """
sqlCtx = get_spark_session()
df = sqlCtx.sql(f"SELECT * FROM {table_spec.table_name}")
df = create_normalization_spec_spark(df, column_name,
preprocessing_options.num_samples,
seed)
rows = df.collect()
normalization_processor = normalization_helper(
max_unique_enum_values=preprocessing_options.max_unique_enum_values,
quantile_size=preprocessing_options.quantile_size,
quantile_k2_threshold=preprocessing_options.quantile_k2_threshold,
skip_box_cox=preprocessing_options.skip_box_cox,
skip_quantiles=preprocessing_options.skip_quantiles,
feature_overrides=preprocessing_options.feature_overrides,
whitelist_features=preprocessing_options.whitelist_features,
assert_whitelist_feature_coverage=preprocessing_options.
assert_whitelist_feature_coverage,
)
return normalization_processor(rows)
def timeline_operator(pkl_path: str, input_table_spec: TableSpec):
""" Loads a pandas parquet, converts to pyspark, and uploads df to Hive.
Then call the timeline operator.
"""
pd_df = pd.read_pickle(pkl_path)
spark = get_spark_session()
df = spark.createDataFrame(pd_df)
input_name = f"{input_table_spec.table_name}{PRE_TIMELINE_SUFFIX}"
df.createTempView(input_name)
output_name = input_table_spec.table_name
include_possible_actions = "possible_actions" in pd_df
arg = {
"startDs": "2019-01-01",
"endDs": "2019-01-01",
"addTerminalStateRow": True,
"inputTableName": input_name,
"outputTableName": output_name,
"includePossibleActions": include_possible_actions,
"percentileFunction": "percentile_approx",
"rewardColumns": ["reward", "metrics"],
"extraFeatureColumns": [],
}
call_spark_class(spark, class_name="Timeline", args=json.dumps(arg))
def upload_to_hive(pkl_path: str, input_table_spec: TableSpec):
""" Loads a pandas parquet, converts to pyspark, and uploads df to Hive. """
pd_df = pd.read_pickle(pkl_path)
spark = get_spark_session()
df = spark.createDataFrame(pd_df)
tbl_name = f"{input_table_spec.table_name}{PRE_TIMELINE_SUFFIX}"
df.write.mode("overwrite").saveAsTable(tbl_name)
def query_data(
input_table_spec: TableSpec,
actions: List[str],
custom_reward_expression: Optional[str] = None,
sample_range: Optional[Tuple[float, float]] = None,
multi_steps: Optional[int] = None,
gamma: Optional[float] = None,
) -> Dataset:
""" Perform reward calculation, hashing mdp + subsampling and
other preprocessing such as sparse2dense.
"""
sqlCtx = get_spark_session()
df = sqlCtx.sql(f"SELECT * FROM {input_table_spec.table_name}")
states = infer_states_names(df, multi_steps)
metrics = infer_metrics_names(df, multi_steps)
df = set_reward_col_as_reward(
df,
custom_reward_expression=custom_reward_expression,
multi_steps=multi_steps,
gamma=gamma,
)
df = hash_mdp_id_and_subsample(df, sample_range=sample_range)
df = perform_preprocessing(
df, states=states, actions=actions, metrics=metrics, multi_steps=multi_steps
)
df = select_relevant_columns(df)
df.write.mode("overwrite").parquet(input_table_spec.output_dataset.parquet_url)
return input_table_spec.output_dataset
def calc_custom_reward(df, custom_reward_expression: str):
sqlCtx = get_spark_session()
# create a temporary table for running sql
temp_table_name = "_tmp_calc_reward_df"
temp_reward_name = "_tmp_reward_col"
df.createOrReplaceTempView(temp_table_name)
df = sqlCtx.sql(
f"SELECT *, CAST(COALESCE({custom_reward_expression}, 0) AS FLOAT)"
f" as {temp_reward_name} FROM {temp_table_name}")
return df.drop("reward").withColumnRenamed(temp_reward_name, "reward")
def query_data(
input_table_spec: TableSpec,
discrete_action: bool,
actions: Optional[List[str]] = None,
include_possible_actions=True,
custom_reward_expression: Optional[str] = None,
sample_range: Optional[Tuple[float, float]] = None,
multi_steps: Optional[int] = None,
gamma: Optional[float] = None,
) -> Dataset:
""" Perform reward calculation, hashing mdp + subsampling and
other preprocessing such as sparse2dense.
"""
sqlCtx = get_spark_session()
# pyre-fixme[16]: `TableSpec` has no attribute `table_name`.
# pyre-fixme[16]: `TableSpec` has no attribute `table_name`.
df = sqlCtx.sql(f"SELECT * FROM {input_table_spec.table_name}")
df = set_reward_col_as_reward(
df,
custom_reward_expression=custom_reward_expression,
multi_steps=multi_steps,
gamma=gamma,
)
df = hash_mdp_id_and_subsample(df, sample_range=sample_range)
df = misc_column_preprocessing(df, multi_steps=multi_steps)
df = state_and_metrics_sparse2dense(
df,
states=infer_states_names(df, multi_steps),
metrics=infer_metrics_names(df, multi_steps),
multi_steps=multi_steps,
)
if discrete_action:
assert include_possible_actions
assert actions is not None, "in discrete case, actions must be given."
df = discrete_action_preprocessing(df,
actions=actions,
multi_steps=multi_steps)
else:
actions = infer_action_names(df, multi_steps)
df = parametric_action_preprocessing(
df,
actions=actions,
multi_steps=multi_steps,
include_possible_actions=include_possible_actions,
)
df = select_relevant_columns(
df,
discrete_action=discrete_action,
include_possible_actions=include_possible_actions,
)
return upload_as_parquet(df)
def timeline_operator(input_table_spec: TableSpec):
""" Call the timeline operator. """
input_name = f"{input_table_spec.table_name}{PRE_TIMELINE_SUFFIX}"
output_name = input_table_spec.table_name
arg = {
"startDs": "2019-01-01",
"endDs": "2019-01-01",
"addTerminalStateRow": True,
"inputTableName": input_name,
"outputTableName": output_name,
"includePossibleActions": True,
"percentileFunction": "percentile_approx",
"rewardColumns": ["reward", "metrics"],
"extraFeatureColumns": [],
}
input_json = json.dumps(arg)
spark = get_spark_session()
spark._jvm.com.facebook.spark.rl.Timeline.main(input_json)
def upload_as_parquet(df) -> Dataset:
""" Generate a random parquet. Fails if cannot generate a non-existent name. """
# get a random tmp name and check if it exists
sqlCtx = get_spark_session()
success = False
for _ in range(MAX_UPLOAD_PARQUET_TRIES):
suffix = rand_string(length=UPLOAD_PARQUET_TMP_SUFFIX_LEN)
rand_name = f"tmp_parquet_{suffix}"
if not sqlCtx.catalog._jcatalog.tableExists(rand_name):
success = True
break
if not success:
raise Exception(f"Failed to find name after {MAX_UPLOAD_PARQUET_TRIES} tries.")
# perform the write
df.write.mode("errorifexists").format("parquet").saveAsTable(rand_name)
parquet_url = get_table_url(rand_name)
logger.info(f"Saved parquet to {parquet_url}")
return Dataset(parquet_url=parquet_url)
def get_table_row_count(parquet_url: str):
spark = get_spark_session()
return spark.read.parquet(parquet_url).count()
