README.md

tensorpandas

Provides extension types to store arbitrary ndarrays (tensors) in a pandas column:

See also [Examples.ipynb]

n = 100
shape = (2, 3)
arrays = [np.random.rand(*shape) for _ in range(n)]

df = pd.DataFrame({
    "x": np.arange(n),
    "tensor": TensorArray(arrays)
})

df["same_tensor"] = pd.Series(arrays).astype("Tensor")  # also works

df["tensor"].tensor.values  # shape (100, 2, 3) for direct data access

# we can manipulate underlying ND array directly via the `tensor` accessor
df2["tensor"].tensor.values[:, -1] *= 2

df.to_parquet("test.parquet". engine="pyarrow")  # store each tensor as fixed-size binary blob

Features

• store n-dimensional arrays in a pandas Series or DataFrame
• support for pyarrow: df.to_parquet/pd.read_parquet support the Tensor extension type
• data is accessible as an (n+1)-dimensional array for efficient slicing or in-place manipulation