kumoai 2.14.0.dev202601051732__cp311-cp311-macosx_11_0_arm64.whl → 2.15.0.dev202601141731__cp311-cp311-macosx_11_0_arm64.whl

kumoai/experimental/rfm/base/sql_sampler.py

@@ -1,8 +1,10 @@
 from abc import abstractmethod
+from collections import defaultdict
 from typing import TYPE_CHECKING, Literal
 
 import numpy as np
 import pandas as pd
+from kumoapi.rfm.context import Subgraph
 from kumoapi.typing import Dtype
 
 from kumoai.experimental.rfm.base import (
@@ -11,11 +13,14 @@ from kumoai.experimental.rfm.base import (
     SamplerOutput,
     SourceColumn,
 )
+from kumoai.experimental.rfm.base.mapper import Mapper
 from kumoai.utils import ProgressLogger, quote_ident
 
 if TYPE_CHECKING:
     from kumoai.experimental.rfm import Graph
 
+EdgeType = tuple[str, str, str]
+
 
 class SQLSampler(Sampler):
     def __init__(
@@ -100,11 +105,28 @@ class SQLSampler(Sampler):
         num_neighbors: list[int],
     ) -> SamplerOutput:
 
-        # TODO Add entity time column to `columns_dict`.
+        # Make sure to always include primary key, foreign key and time columns
+        # during data fetching since these are needed for graph traversal:
+        sample_columns_dict: dict[str, set[str]] = {}
+        for table, columns in columns_dict.items():
+            sample_columns = columns | {
+                foreign_key
+                for foreign_key, _ in self.foreign_key_dict[table]
+            }
+            if primary_key := self.primary_key_dict.get(table):
+                sample_columns |= {primary_key}
+            sample_columns_dict[table] = sample_columns
+        if not isinstance(anchor_time, pd.Series):
+            sample_columns_dict[entity_table_name] |= {
+                self.time_column_dict[entity_table_name]
+            }
+
+        # Sample Entity Table #################################################
+
         df, batch = self._by_pkey(
             table_name=entity_table_name,
-            pkey=entity_pkey,
-            columns=columns_dict[entity_table_name],
+            index=entity_pkey,
+            columns=sample_columns_dict[entity_table_name],
         )
         if len(batch) != len(entity_pkey):
             mask = np.ones(len(entity_pkey), dtype=bool)
@@ -113,23 +135,230 @@ class SQLSampler(Sampler):
                            f"{entity_pkey.iloc[mask].tolist()} do not exist "
                            f"in the '{entity_table_name}' table")
 
+        # Make sure that entities are returned in expected order:
         perm = batch.argsort()
         batch = batch[perm]
         df = df.iloc[perm].reset_index(drop=True)
 
+        # Fill 'entity' anchor times with actual values:
         if not isinstance(anchor_time, pd.Series):
             time_column = self.time_column_dict[entity_table_name]
             anchor_time = df[time_column]
+        assert isinstance(anchor_time, pd.Series)
+
+        # Recursive Neighbor Sampling #########################################
+
+        mapper_dict: dict[str, Mapper] = defaultdict(
+            lambda: Mapper(num_examples=len(entity_pkey)))
+        mapper_dict[entity_table_name].add(
+            pkey=df[self.primary_key_dict[entity_table_name]],
+            batch=batch,
+        )
+
+        dfs_dict: dict[str, list[pd.DataFrame]] = defaultdict(list)
+        dfs_dict[entity_table_name].append(df)
+        batches_dict: dict[str, list[np.ndarray]] = defaultdict(list)
+        batches_dict[entity_table_name].append(batch)
+        num_sampled_nodes_dict: dict[str, list[int]] = defaultdict(
+            lambda: [0] * (len(num_neighbors) + 1))
+        num_sampled_nodes_dict[entity_table_name][0] = len(entity_pkey)
+
+        rows_dict: dict[EdgeType, list[np.ndarray]] = defaultdict(list)
+        cols_dict: dict[EdgeType, list[np.ndarray]] = defaultdict(list)
+        num_sampled_edges_dict: dict[EdgeType, list[int]] = defaultdict(
+            lambda: [0] * len(num_neighbors))
+
+        # The start index of data frame slices of the previous hop:
+        offset_dict: dict[str, int] = defaultdict(int)
+
+        for hop, neighbors in enumerate(num_neighbors):
+            if neighbors == 0:
+                break  # Abort early.
+
+            for table in list(num_sampled_nodes_dict.keys()):
+                # Only sample from tables that have been visited in the
+                # previous hop:
+                if num_sampled_nodes_dict[table][hop] == 0:
+                    continue
+
+                # Collect the slices of data sampled in the previous hop
+                # (but maintain only required key information):
+                cols = [fkey for fkey, _ in self.foreign_key_dict[table]]
+                if table in self.primary_key_dict:
+                    cols.append(self.primary_key_dict[table])
+                dfs = [df[cols] for df in dfs_dict[table][offset_dict[table]:]]
+                df = pd.concat(
+                    dfs,
+                    axis=0,
+                    ignore_index=True,
+                ) if len(dfs) > 1 else dfs[0]
+                batches = batches_dict[table][offset_dict[table]:]
+                batch = (np.concatenate(batches)
+                         if len(batches) > 1 else batches[0])
+                offset_dict[table] = len(batches_dict[table])  # Increase.
+
+                pkey: pd.Series | None = None
+                index: pd.ndarray | None = None
+                if table in self.primary_key_dict:
+                    pkey = df[self.primary_key_dict[table]]
+                    index = mapper_dict[table].get(pkey, batch)
+
+                # Iterate over foreign keys in the current table:
+                for fkey, dst_table in self.foreign_key_dict[table]:
+                    row = mapper_dict[dst_table].get(df[fkey], batch)
+                    mask = row == -1
+                    if mask.any():
+                        key_df = pd.DataFrame({
+                            'fkey': df[fkey],
+                            'batch': batch,
+                        }).iloc[mask]
+                        # Only maintain unique keys per example:
+                        unique_key_df = key_df.drop_duplicates()
+                        # Fully de-duplicate keys across examples:
+                        code, fkey_index = pd.factorize(unique_key_df['fkey'])
+
+                        _df, _batch = self._by_pkey(
+                            table_name=dst_table,
+                            index=fkey_index,
+                            columns=sample_columns_dict[dst_table],
+                        )  # Ensure result is sorted according to input order:
+                        _df = _df.iloc[_batch.argsort()]
+
+                        # Compute valid entries (without dangling foreign keys)
+                        # in `unique_fkey_df`:
+                        _mask = np.full(len(fkey_index), fill_value=False)
+                        _mask[_batch] = True
+                        _mask = _mask[code]
+
+                        # Recontruct unique (key, batch) pairs:
+                        code, _ = pd.factorize(unique_key_df['fkey'][_mask])
+                        _df = _df.iloc[code].reset_index(drop=True)
+                        _batch = unique_key_df['batch'].to_numpy()[_mask]
+
+                        # Register node IDs:
+                        mapper_dict[dst_table].add(
+                            pkey=_df[self.primary_key_dict[dst_table]],
+                            batch=_batch,
+                        )
+                        row[mask] = mapper_dict[dst_table].get(
+                            pkey=key_df['fkey'],
+                            batch=key_df['batch'].to_numpy(),
+                        )  # NOTE `row` may still hold `-1` for dangling fkeys.
+
+                        dfs_dict[dst_table].append(_df)
+                        batches_dict[dst_table].append(_batch)
+                        num_sampled_nodes_dict[dst_table][hop + 1] += (  #
+                            len(_batch))
+
+                    mask = row != -1
+
+                    col = index
+                    if col is None:
+                        start = sum(num_sampled_nodes_dict[table][:hop])
+                        end = sum(num_sampled_nodes_dict[table][:hop + 1])
+                        col = np.arange(start, end)
+
+                    row = row[mask]
+                    col = col[mask]
+
+                    edge_type = (table, fkey, dst_table)
+                    edge_type = Subgraph.rev_edge_type(edge_type)
+                    rows_dict[edge_type].append(row)
+                    cols_dict[edge_type].append(col)
+                    num_sampled_edges_dict[edge_type][hop] = len(col)
+
+                # Iterate over foreign keys that reference the current table:
+                for src_table, fkey in self.rev_foreign_key_dict[table]:
+                    assert pkey is not None and index is not None
+                    _df, _batch = self._by_fkey(
+                        table_name=src_table,
+                        foreign_key=fkey,
+                        index=pkey,
+                        num_neighbors=neighbors,
+                        anchor_time=anchor_time.iloc[batch],
+                        columns=sample_columns_dict[src_table],
+                    )
+
+                    edge_type = (src_table, fkey, table)
+                    cols_dict[edge_type].append(index[_batch])
+                    num_sampled_edges_dict[edge_type][hop] = len(_batch)
+
+                    _batch = batch[_batch]
+                    num_nodes = sum(num_sampled_nodes_dict[src_table])
+                    if src_table in self.primary_key_dict:
+                        _pkey = _df[self.primary_key_dict]
+                        mapper_dict[src_table].add(_pkey, _batch)
+                        row = mapper_dict[src_table].get(_pkey, _batch)
+
+                        # Only preserve unknown rows:
+                        mask = row >= num_nodes  # type: ignore
+                        mask[pd.duplicated(row)] = False
+                        _df = _df.iloc[mask]
+                        _batch = _batch[mask]
+                    else:
+                        row = np.arange(num_nodes, num_nodes + len(_batch))
+
+                    rows_dict[edge_type].append(row)
+                    num_sampled_nodes_dict[src_table][hop + 1] += len(_batch)
+
+                    dfs_dict[src_table].append(_df)
+                    batches_dict[src_table].append(_batch)
+
+        # Post-Processing #####################################################
+
+        df_dict = {
+            table:
+            pd.concat(dfs, axis=0, ignore_index=True)
+            if len(dfs) > 1 else dfs[0]
+            for table, dfs in dfs_dict.items()
+        }
+
+        # Only store unique rows in `df` above a certain threshold:
+        inverse_dict: dict[str, np.ndarray] = {}
+        for table, df in df_dict.items():
+            if table not in self.primary_key_dict:
+                continue
+            unique, index, inverse = np.unique(
+                df_dict[table][self.primary_key_dict[table]],
+                return_index=True,
+                return_inverse=True,
+            )
+            if len(df) > 1.05 * len(unique):
+                df_dict[table] = df.iloc[index].reset_index(drop=True)
+                inverse_dict[table] = inverse
+
+        df_dict = {  # Post-filter column set:
+            table: df[list(columns_dict[table])]
+            for table, df in df_dict.items()
+        }
+        batch_dict = {
+            table: np.concatenate(batches) if len(batches) > 1 else batches[0]
+            for table, batches in batches_dict.items()
+        }
+        row_dict = {
+            edge_type: np.concatenate(rows)
+            for edge_type, rows in rows_dict.items()
+        }
+        col_dict = {
+            edge_type: np.concatenate(cols)
+            for edge_type, cols in cols_dict.items()
+        }
+
+        if len(num_sampled_edges_dict) == 0:  # Single table:
+            num_sampled_nodes_dict = {
+                key: value[:1]
+                for key, value in num_sampled_nodes_dict.items()
+            }
 
         return SamplerOutput(
             anchor_time=anchor_time.astype(int).to_numpy(),
-            df_dict={entity_table_name: df},
-            inverse_dict={},
-            batch_dict={entity_table_name: batch},
-            num_sampled_nodes_dict={entity_table_name: [len(batch)]},
-            row_dict={},
-            col_dict={},
-            num_sampled_edges_dict={},
+            df_dict=df_dict,
+            inverse_dict=inverse_dict,
+            batch_dict=batch_dict,
+            num_sampled_nodes_dict=num_sampled_nodes_dict,
+            row_dict=row_dict,
+            col_dict=col_dict,
+            num_sampled_edges_dict=num_sampled_edges_dict,
         )
 
     # Abstract Methods ########################################################
@@ -138,7 +367,19 @@ class SQLSampler(Sampler):
     def _by_pkey(
         self,
         table_name: str,
-        pkey: pd.Series,
+        index: pd.Series,
+        columns: set[str],
+    ) -> tuple[pd.DataFrame, np.ndarray]:
+        pass
+
+    @abstractmethod
+    def _by_fkey(
+        self,
+        table_name: str,
+        foreign_key: str,
+        index: pd.Series,
+        num_neighbors: int,
+        anchor_time: pd.Series | None,
         columns: set[str],
     ) -> tuple[pd.DataFrame, np.ndarray]:
         pass

kumoai/experimental/rfm/base/table.py

@@ -1,4 +1,3 @@
-import warnings
 from abc import ABC, abstractmethod
 from collections.abc import Sequence
 from functools import cached_property
@@ -20,6 +19,7 @@ from kumoai.experimental.rfm.base import (
     SourceColumn,
     SourceForeignKey,
 )
+from kumoai.experimental.rfm.base.utils import to_datetime
 from kumoai.experimental.rfm.infer import (
     infer_dtype,
     infer_primary_key,
@@ -399,39 +399,39 @@ class Table(ABC):
         r"""Returns a :class:`pandas.DataFrame` object containing metadata
         information about the columns in this table.
 
-        The returned dataframe has columns ``name``, ``dtype``, ``stype``,
-        ``is_primary_key``, ``is_time_column`` and ``is_end_time_column``,
-        which provide an aggregate view of the properties of the columns of
-        this table.
+        The returned dataframe has columns ``"Name"``, ``"Data Type"``,
+        ``"Semantic Type"``, ``"Primary Key"``, ``"Time Column"`` and
+        ``"End Time Column"``, which provide an aggregated view of the
+        properties of the columns of this table.
 
         Example:
             >>> # doctest: +SKIP
             >>> import kumoai.experimental.rfm as rfm
             >>> table = rfm.LocalTable(df=..., name=...).infer_metadata()
             >>> table.metadata
-                name        dtype    stype  is_primary_key  is_time_column  is_end_time_column
-            0   CustomerID  float64  ID     True            False           False
+                Name        Data Type  Semantic Type  Primary Key  Time Column  End Time Column
+            0   CustomerID  float64    ID             True         False        False
         """  # noqa: E501
         cols = self.columns
 
         return pd.DataFrame({
-            'name':
+            'Name':
             pd.Series(dtype=str, data=[c.name for c in cols]),
-            'dtype':
+            'Data Type':
             pd.Series(dtype=str, data=[c.dtype for c in cols]),
-            'stype':
+            'Semantic Type':
             pd.Series(dtype=str, data=[c.stype for c in cols]),
-            'is_primary_key':
+            'Primary Key':
             pd.Series(
                 dtype=bool,
                 data=[self._primary_key == c.name for c in cols],
             ),
-            'is_time_column':
+            'Time Column':
             pd.Series(
                 dtype=bool,
                 data=[self._time_column == c.name for c in cols],
             ),
-            'is_end_time_column':
+            'End Time Column':
             pd.Series(
                 dtype=bool,
                 data=[self._end_time_column == c.name for c in cols],
@@ -623,20 +623,6 @@ class Table(ABC):
         r"""Sanitzes a :class:`pandas.DataFrame` in-place such that its data
         types match table data and semantic type specification.
         """
-        def _to_datetime(ser: pd.Series) -> pd.Series:
-            if not pd.api.types.is_datetime64_any_dtype(ser):
-                with warnings.catch_warnings():
-                    warnings.filterwarnings(
-                        'ignore',
-                        message='Could not infer format',
-                    )
-                    ser = pd.to_datetime(ser, errors='coerce')
-            if isinstance(ser.dtype, pd.DatetimeTZDtype):
-                ser = ser.dt.tz_localize(None)
-            if ser.dtype != 'datetime64[ns]':
-                ser = ser.astype('datetime64[ns]')
-            return ser
-
         def _to_list(ser: pd.Series, dtype: Dtype | None) -> pd.Series:
             if (pd.api.types.is_string_dtype(ser)
                     and dtype in {Dtype.intlist, Dtype.floatlist}):
@@ -667,9 +653,9 @@ class Table(ABC):
             stype = (stype_dict or {}).get(column_name)
 
             if dtype == Dtype.time:
-                df[column_name] = _to_datetime(df[column_name])
+                df[column_name] = to_datetime(df[column_name])
             elif stype == Stype.timestamp:
-                df[column_name] = _to_datetime(df[column_name])
+                df[column_name] = to_datetime(df[column_name])
             elif dtype is not None and dtype.is_list():
                 df[column_name] = _to_list(df[column_name], dtype)
             elif stype == Stype.sequence:

kumoai/experimental/rfm/base/utils.py

@@ -0,0 +1,36 @@
+import warnings
+
+import pandas as pd
+import pyarrow as pa
+
+
+def is_datetime(ser: pd.Series) -> bool:
+    r"""Check whether a :class:`pandas.Series` holds datetime values."""
+    if isinstance(ser.dtype, pd.ArrowDtype):
+        dtype = ser.dtype.pyarrow_dtype
+        return (pa.types.is_timestamp(dtype) or pa.types.is_date(dtype)
+                or pa.types.is_time(dtype))
+
+    return pd.api.types.is_datetime64_any_dtype(ser)
+
+
+def to_datetime(ser: pd.Series) -> pd.Series:
+    """Converts a :class:`pandas.Series` to ``datetime64[ns]`` format."""
+    if isinstance(ser.dtype, pd.ArrowDtype):
+        ser = pd.Series(ser.to_numpy(), index=ser.index, name=ser.name)
+
+    if not pd.api.types.is_datetime64_any_dtype(ser):
+        with warnings.catch_warnings():
+            warnings.filterwarnings(
+                'ignore',
+                message='Could not infer format',
+            )
+            ser = pd.to_datetime(ser, errors='coerce')
+
+    if isinstance(ser.dtype, pd.DatetimeTZDtype):
+        ser = ser.dt.tz_localize(None)
+
+    if ser.dtype != 'datetime64[ns]':
+        ser = ser.astype('datetime64[ns]')
+
+    return ser

kumoai/experimental/rfm/graph.py

@@ -649,28 +649,28 @@ class Graph:
         r"""Returns a :class:`pandas.DataFrame` object containing metadata
         information about the tables in this graph.
 
-        The returned dataframe has columns ``name``, ``primary_key``,
-        ``time_column``, and ``end_time_column``, which provide an aggregate
-        view of the properties of the tables of this graph.
+        The returned dataframe has columns ``"Name"``, ``"Primary Key"``,
+        ``"Time Column"``, and ``"End Time Column"``, which provide an
+        aggregated view of the properties of the tables of this graph.
 
         Example:
             >>> # doctest: +SKIP
             >>> import kumoai.experimental.rfm as rfm
             >>> graph = rfm.Graph(tables=...).infer_metadata()
             >>> graph.metadata  # doctest: +SKIP
-                name   primary_key  time_column  end_time_column
-            0   users      user_id            -                -
+                Name   Primary Key  Time Column  End Time Column
+            0   users  user_id      -            -
         """
         tables = list(self.tables.values())
 
         return pd.DataFrame({
-            'name':
+            'Name':
             pd.Series(dtype=str, data=[t.name for t in tables]),
-            'primary_key':
+            'Primary Key':
             pd.Series(dtype=str, data=[t._primary_key or '-' for t in tables]),
-            'time_column':
+            'Time Column':
             pd.Series(dtype=str, data=[t._time_column or '-' for t in tables]),
-            'end_time_column':
+            'End Time Column':
             pd.Series(
                 dtype=str,
                 data=[t._end_time_column or '-' for t in tables],

kumoai/experimental/rfm/infer/dtype.py

@@ -3,6 +3,8 @@ import pandas as pd
 import pyarrow as pa
 from kumoapi.typing import Dtype
 
+from kumoai.experimental.rfm.base.utils import is_datetime
+
 PANDAS_TO_DTYPE: dict[str, Dtype] = {
     'bool': Dtype.bool,
     'boolean': Dtype.bool,
@@ -34,7 +36,7 @@ def infer_dtype(ser: pd.Series) -> Dtype:
     Returns:
         The data type.
     """
-    if pd.api.types.is_datetime64_any_dtype(ser.dtype):
+    if is_datetime(ser):
         return Dtype.date
     if pd.api.types.is_timedelta64_dtype(ser.dtype):
         return Dtype.timedelta

kumoai/experimental/rfm/infer/time_col.py

@@ -3,6 +3,8 @@ import warnings
 
 import pandas as pd
 
+from kumoai.experimental.rfm.base.utils import to_datetime
+
 
 def infer_time_column(
     df: pd.DataFrame,
@@ -43,11 +45,11 @@ def infer_time_column(
     with warnings.catch_warnings():
         warnings.filterwarnings('ignore', message='Could not infer format')
         min_timestamp_dict = {
-            key: pd.to_datetime(df[key].iloc[:10_000], 'coerce')
+            key: to_datetime(df[key].iloc[:10_000])
             for key in candidates
         }
     min_timestamp_dict = {
-        key: value.min().tz_localize(None)
+        key: value.min()
         for key, value in min_timestamp_dict.items()
     }
     min_timestamp_dict = {