PERF: DataFrame(pytorch_tensor)

asv_bench/benchmarks/frame_ctor.py

@@ -182,4 +182,21 @@ def time_frame_from_arrays_sparse(self):
         )
 
 
+class From3rdParty:
+    # GH#44616
+
+    def setup_cache(self):
+        try:
+            import torch
+        except ImportError:
+            raise NotImplementedError
+
+        row = 700000
+        col = 64
+        self.val_tensor = torch.randn(row, col)
+
+    def time_from_torch(self):
+        DataFrame(self.val_tensor)
+
+
 from .pandas_vb_common import setup  # noqa: F401 isort:skip

pandas/core/frame.py

@@ -705,11 +705,16 @@ def __init__(
         # For data is list-like, or Iterable (will consume into list)
         elif is_list_like(data):
             if not isinstance(data, (abc.Sequence, ExtensionArray)):
-                data = list(data)
+                if hasattr(data, "__array__"):
+                    # GH#44616 big perf improvement for e.g. pytorch tensor
+                    data = np.asarray(data)
+                else:
+                    data = list(data)
             if len(data) > 0:
                 if is_dataclass(data[0]):
                     data = dataclasses_to_dicts(data)
-                if treat_as_nested(data):
+                if not isinstance(data, np.ndarray) and treat_as_nested(data):
+                    # exclude ndarray as we may have cast it a few lines above
                     if columns is not None:
                         # error: Argument 1 to "ensure_index" has incompatible type
                         # "Collection[Any]"; expected "Union[Union[Union[ExtensionArray,

pandas/tests/test_downstream.py

@@ -5,7 +5,7 @@
 import subprocess
 import sys
 
-import numpy as np  # noqa:F401 needed in namespace for statsmodels
+import numpy as np
 import pytest
 
 import pandas.util._test_decorators as td
@@ -176,6 +176,20 @@ def test_pyarrow(df):
     tm.assert_frame_equal(result, df)
 
 
+def test_torch_frame_construction(using_array_manager):
+    # GH#44616
+    torch = import_module("torch")
+    val_tensor = torch.randn(700, 64)
+
+    df = DataFrame(val_tensor)
+
+    if not using_array_manager:
+        assert np.shares_memory(df, val_tensor)
+
+    ser = pd.Series(val_tensor[0])
+    assert np.shares_memory(ser, val_tensor)
+
+
 def test_yaml_dump(df):
     # GH#42748
     yaml = import_module("yaml")