add onnx

README.md

@@ -28,6 +28,7 @@ So far we support:
 - pandas dataframes:
     - csv
     - parquet
+- onnx.ModelProto instances
 
 ## Dependencies
 

dummio/__init__.py

@@ -5,9 +5,16 @@
 
 from dummio import json as json
 from dummio import text as text
+from dummio.constants import ModuleProtocol as ModuleProtocol
 
 try:
     from dummio import yaml as yaml
 except ImportError:
-    # yaml is an optional dependency
+    # this would require an optional yaml dependency such as pyyaml
+    pass
+
+try:
+    from dummio import onnx as onnx
+except ImportError:
+    # this would require the optional dependency onnx
     pass

dummio/constants.py

@@ -1,10 +1,28 @@
 """Constants for dummio."""
 
+import typing
 from pathlib import Path
-from typing import Any, TypeAlias
+from typing import Any, Protocol, TypeAlias
 
 PathType: TypeAlias = str | Path
 AnyDict: TypeAlias = dict[Any, Any]
 
+# pyright expect a type var to be used at least twice within a single method. It's having
+# trouble respecting how it's use *accross* methods of a class.
+T = typing.TypeVar("T")  # pyright: ignore
+
 DEFAULT_ENCODING = "utf-8"
 DEFAULT_WRITE_MODE = "w"
+
+
+@typing.runtime_checkable
+class ModuleProtocol(Protocol):
+    """Protocol for dummio's IO modules."""
+
+    def save(self, data: T, *, filepath: PathType) -> None:  # pyright: ignore[reportInvalidTypeVarUse]
+        """Declares the signature of an IO module save method."""
+        ...
+
+    def load(self, filepath: PathType) -> T:  # pyright: ignore[reportInvalidTypeVarUse]
+        """Declares the signature of an IO module load method."""
+        ...

dummio/onnx.py

@@ -0,0 +1,28 @@
+"""IO methods for sklearn models using ONNX serialization."""
+
+import onnx
+
+from dummio.constants import PathType
+
+
+def save(data: onnx.ModelProto, *, filepath: PathType) -> None:
+    """Saves a sklearn model to a file using ONNX serialization.
+
+    Args:
+        data: Data to save. This needs to be an sklearn model.
+        filepath: Path to save the data.
+    """
+    byte_str = data.SerializeToString()
+    with open(filepath, "wb") as file:
+        file.write(byte_str)
+
+
+def load(filepath: PathType) -> onnx.ModelProto:
+    """Loads a sklearn model from a file using ONNX serialization.
+
+    Args:
+        filepath: Path to read the data.
+    """
+    with open(filepath, "rb") as file:
+        byte_str = file.read()
+    return onnx.load_model_from_string(byte_str)

dummio/pandas/__init__.py

@@ -1,9 +1,2 @@
-try:
-    import pandas
-
-    del pandas
-except ImportError:
-    raise ImportError("Please install pandas to use dummio.pandas")
-
 from dummio.pandas import df_csv as df_csv
 from dummio.pandas import df_parquet as df_parquet

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "dummio"
-version = "1.0.0"
+version = "1.1.0"
 description = "Easiest-possible IO for basic file types."
 authors = [{ name = "Zach Kurtz", email = "[email protected]" }]
 readme = "README.md"
@@ -12,11 +12,14 @@ dev = [
   "pyright >=1.1.378",
   "ruff >=0.7.4",
   "pytest >=8.3.2",
+  "scikit-learn >=1.0.2",
+  "skl2onnx >=1.10.1",
 ]
 extras = [
   "fastparquet>=2024.11.0",
   "pandas>=1.5.0",
   "pyyaml>=6.0.2",
+  "onnx>=1.10.1",
 ]
 
 [project.urls]

tests/test_assert_module_protocol.py

@@ -0,0 +1,20 @@
+"""Assert that every IO module implements the ModuleProtocol."""
+
+import importlib
+
+from dummio import ModuleProtocol
+
+IO_MODULES = [
+    "dummio.json",
+    "dummio.onnx",
+    "dummio.text",
+    "dummio.yaml",
+    "dummio.pandas.df_csv",
+    "dummio.pandas.df_parquet",
+]
+
+
+def test_assert_module_protocol() -> None:
+    for module_name in IO_MODULES:
+        module = importlib.import_module(module_name)
+        assert isinstance(module, ModuleProtocol)

tests/test_onnx.py

@@ -0,0 +1,28 @@
+import onnx
+from skl2onnx import convert_sklearn
+from skl2onnx.common.data_types import FloatTensorType
+from sklearn.datasets import load_iris
+from sklearn.linear_model import LogisticRegression
+
+from dummio import onnx as onnx_io
+
+
+def test_onnx_io_cycle(tmp_path) -> None:
+    # Empty model:
+    model = onnx.ModelProto()
+    filepath = tmp_path / "model.onnx"
+    onnx_io.save(model, filepath=filepath)
+    loaded_model = onnx_io.load(filepath=filepath)
+    assert model.SerializeToString() == loaded_model.SerializeToString()
+
+    # Minimal sklearn model:
+    iris = load_iris()
+    clr = LogisticRegression(solver="saga", max_iter=10000)
+    clr.fit(iris.data, iris.target)  # pyright: ignore[reportAttributeAccessIssue]
+    initial_type = [("float_input", FloatTensorType([None, 4]))]
+    onx = convert_sklearn(clr, initial_types=initial_type)
+    assert isinstance(onx, onnx.ModelProto)  # pyright wasn't sure
+    filepath = tmp_path / "model.onnx"
+    onnx_io.save(onx, filepath=filepath)
+    loaded_onx = onnx_io.load(filepath=filepath)
+    assert onx.SerializeToString() == loaded_onx.SerializeToString()

tests/test_pandas.py

@@ -1,5 +1,3 @@
-"""Test IO methods for tabulare data types."""
-
 from pathlib import Path
 from types import ModuleType