wip (#344)

docs/api-reference/dtypes.md

@@ -18,6 +18,7 @@
         - Categorical
         - String
         - Datetime
+        - Object
       show_root_heading: false
       show_source: false
       show_bases: false

narwhals/__init__.py

@@ -11,6 +11,7 @@
 from narwhals.dtypes import Int16
 from narwhals.dtypes import Int32
 from narwhals.dtypes import Int64
+from narwhals.dtypes import Object
 from narwhals.dtypes import String
 from narwhals.dtypes import UInt8
 from narwhals.dtypes import UInt16
@@ -72,6 +73,7 @@
     "Float64",
     "Float32",
     "Boolean",
+    "Object",
     "Categorical",
     "String",
     "Datetime",

narwhals/_pandas_like/dataframe.py

@@ -135,7 +135,8 @@ def iter_rows(
     @property
     def schema(self) -> dict[str, DType]:
         return {
-            col: translate_dtype(dtype) for col, dtype in self._dataframe.dtypes.items()
+            col: translate_dtype(self._dataframe.loc[:, col])
+            for col in self._dataframe.columns
         }
 
     # --- reshape ---

narwhals/_pandas_like/namespace.py

@@ -32,6 +32,7 @@ class PandasNamespace:
     Float64 = dtypes.Float64
     Float32 = dtypes.Float32
     Boolean = dtypes.Boolean
+    Object = dtypes.Object
     Categorical = dtypes.Categorical
     String = dtypes.String
     Datetime = dtypes.Datetime

narwhals/_pandas_like/series.py

@@ -151,7 +151,7 @@ def shape(self) -> tuple[int]:
 
     @property
     def dtype(self) -> DType:
-        return translate_dtype(self._series.dtype)
+        return translate_dtype(self._series)
 
     def cast(
         self,

narwhals/_pandas_like/utils.py

@@ -373,9 +373,10 @@ def set_axis(obj: T, index: Any, implementation: str) -> T:
         return obj.set_axis(index, axis=0)  # type: ignore[no-any-return, attr-defined]
 
 
-def translate_dtype(dtype: Any) -> DType:
+def translate_dtype(column: Any) -> DType:
     from narwhals import dtypes
 
+    dtype = column.dtype
     if dtype in ("int64", "Int64", "Int64[pyarrow]"):
         return dtypes.Int64()
     if dtype in ("int32", "Int32", "Int32[pyarrow]"):
@@ -412,7 +413,15 @@ def translate_dtype(dtype: Any) -> DType:
     if str(dtype) == "date32[day][pyarrow]":
         return dtypes.Date()
     if dtype == "object":
-        return dtypes.String()
+        if (idx := column.first_valid_index()) is not None and isinstance(
+            column.loc[idx], str
+        ):
+            # Infer based on first non-missing value.
+            # For pandas pre 3.0, this isn't perfect.
+            # After pandas 3.0, pandas has a dedicated string dtype
+            # which is inferred by default.
+            return dtypes.String()
+        return dtypes.Object()
     msg = f"Unknown dtype: {dtype}"  # pragma: no cover
     raise AssertionError(msg)
 

narwhals/dtypes.py

@@ -64,6 +64,9 @@ class String(DType): ...
 class Boolean(DType): ...
 
 
+class Object(DType): ...
+
+
 class Datetime(TemporalType): ...
 
 
@@ -145,6 +148,8 @@ def to_narwhals_dtype(dtype: Any, *, is_polars: bool) -> DType:
         return String()
     if dtype == pl.Boolean:
         return Boolean()
+    if dtype == pl.Object:
+        return Object()
     if dtype == pl.Categorical:
         return Categorical()
     if dtype == pl.Datetime:

tests/frame/schema_test.py

@@ -1,5 +1,10 @@
 from datetime import datetime
 from datetime import timezone
+from typing import Any
+
+import pandas as pd
+import polars as pl
+import pytest
 
 import narwhals as nw
 
@@ -12,3 +17,25 @@
 def test_schema_comparison() -> None:
     assert {"a": nw.String()} != {"a": nw.Int32()}
     assert {"a": nw.Int32()} == {"a": nw.Int32()}
+
+
+def test_object() -> None:
+    df = pd.DataFrame({"a": [1, 2, 3]}).astype(object)
+    result = nw.from_native(df).schema
+    assert result["a"] == nw.Object
+
+
+def test_string_disguised_as_object() -> None:
+    df = pd.DataFrame({"a": ["foo", "bar"]}).astype(object)
+    result = nw.from_native(df).schema
+    assert result["a"] == nw.String
+
+
+@pytest.mark.parametrize("constructor", [pd.DataFrame, pl.DataFrame])
+def test_actual_object(constructor: Any) -> None:
+    class Foo: ...
+
+    data = {"a": [Foo()]}
+    df = nw.from_native(constructor(data))
+    result = df.schema
+    assert result == {"a": nw.Object}