FEAT Function to visualize skops files

docs/changes.rst

@@ -25,6 +25,9 @@ v0.6
 - ``add_*`` methods on :class:`.Card` now have default section names (but
   ``None`` is no longer valid) and no longer add descriptions by default.
   :pr:`321` by `Benjamin Bossan`_.
+- Add possibility to visualize a skops object and show untrusted types by using
+  :func:`skops.io.visualize`. For colored output, install `rich`: `pip install
+  rich`. :pr:`317` by `Benjamin Bossan`_.
 
 v0.5
 ----

docs/persistence.rst

@@ -134,6 +134,64 @@ For example, to convert all ``.pkl`` flies in the current directory:
 Further help for the different supported options can be found by calling
 ``skops convert --help`` in a terminal.
 
+Visualization
+#############
+
+Skops files can be visualized using :func:`skops.io.visualize`. If you have
+a skops file called ``my-model.skops``, you can visualize it like this:
+
+.. code:: python
+
+    import skops.io as sio
+    sio.visualize("my-model.skops")
+
+The output could look like this:
+
+.. code::
+
+    root: sklearn.preprocessing._data.MinMaxScaler
+    └── attrs: builtins.dict
+        ├── feature_range: builtins.tuple
+        │   ├── content: json-type(-555)
+        │   └── content: json-type(123)
+        ├── copy: unsafe_lib.UnsafeType [UNSAFE]
+        ├── clip: json-type(false)
+        └── _sklearn_version: json-type("1.2.0")
+
+``unsafe_lib.UnsafeType`` was recognized as untrusted and marked.
+
+It's also possible to visualize the object dumped as bytes:
+
+    import skops.io as sio
+    my_model = ...
+    sio.visualize(sio.dumps(my_model))
+
+There are various options to customize the output. By default, the security of
+nodes is color coded if `rich <https://github.com/Textualize/rich>`_ is
+installed, otherwise they all have the same color. To install ``rich``, run:
+
+.. code::
+
+    python -m pip install rich
+
+or, when installing skops, install it like this:
+
+    python -m pip install skops[rich]
+
+To disable colors, even if ``rich`` is installed, pass ``use_colors=False`` to
+:func:`skops.io.visualize`.
+
+It's also possible to change what colors are being used, e.g. by passing
+``visualize(..., color_safe="cyan")`` to change the color for trusted nodes from
+green to cyan. The ``rich`` docs list the `supported standard colors
+<https://rich.readthedocs.io/en/stable/appendix/colors.html>`_.
+
+Note that the visualization feature is intended to help understand the structure
+of the object, e.g. what attributes are identified as untrusted. It is not a
+replacement for a proper security check. In particular, just because an object's
+visualization looks innocent does *not* mean you can just call `sio.load(<file>,
+trusted=True)` on this object -- only pass the types you really trust to the
+``trusted`` argument.
 
 Supported libraries
 -------------------

setup.py

@@ -77,6 +77,7 @@ def setup_package():
         extras_require={
             "docs": min_deps.tag_to_packages["docs"],
             "tests": min_deps.tag_to_packages["tests"],
+            "rich": min_deps.tag_to_packages["rich"],
         },
         include_package_data=True,
     )

skops/_min_dependencies.py

@@ -6,7 +6,8 @@
 # 'build' and 'install' is included to have structured metadata for CI.
 # It will NOT be included in setup's extras_require
 # The values are (version_spec, comma separated tags, condition)
-# tags can be: 'build', 'install', 'docs', 'examples', 'tests', 'benchmark'
+# tags can be: 'build', 'install', 'docs', 'examples', 'tests', 'benchmark',
+# 'rich'
 # example:
 #     "tomli": ("1.1.0", "install", "python_full_version < '3.11.0a7'"),
 dependent_packages = {
@@ -34,13 +35,14 @@
     # TODO: remove condition when catboost supports python 3.11
     "catboost": ("1.0", "tests", "python_version < '3.11'"),
     "fairlearn": ("0.7.0", "docs, tests", None),
+    "rich": ("12", "tests, rich", None),
 }
 
 
 # create inverse mapping for setuptools
 tag_to_packages: dict = {
     extra: []
-    for extra in ["build", "install", "docs", "examples", "tests", "benchmark"]
+    for extra in ["build", "install", "docs", "examples", "tests", "benchmark", "rich"]
 }
 for package, (min_version, extras, condition) in dependent_packages.items():
     for extra in extras.split(", "):

skops/conftest.py

@@ -43,3 +43,16 @@ def mock_import(name, *args, **kwargs):
         yield
 
     import matplotlib  # noqa
+
+
+@pytest.fixture
+def rich_not_installed():
+    orig_import = builtins.__import__
+
+    def mock_import(name, *args, **kwargs):
+        if name == "rich":
+            raise ImportError
+        return orig_import(name, *args, **kwargs)
+
+    with patch("builtins.__import__", side_effect=mock_import):
+        yield

skops/io/__init__.py

@@ -1,3 +1,4 @@
 from ._persist import dump, dumps, get_untrusted_types, load, loads
+from ._visualize import visualize
 
-__all__ = ["dumps", "load", "loads", "dump", "get_untrusted_types"]
+__all__ = ["dumps", "load", "loads", "dump", "get_untrusted_types", "visualize"]

skops/io/_audit.py

@@ -287,6 +287,10 @@ def get_unsafe_set(self) -> set[str]:
 
         return res
 
+    def format(self) -> str:
+        """Representation of the node's content."""
+        return f"{self.module_name}.{self.class_name}"
+
 
 class CachedNode(Node):
     def __init__(

skops/io/_general.py

@@ -482,6 +482,14 @@ def get_unsafe_set(self) -> set[str]:
     def _construct(self):
         return json.loads(self.content)
 
+    def format(self) -> str:
+        """Representation of the node's content.
+
+        Since no module is used, just show the content.
+
+        """
+        return f"json-type({self.content})"
+
 
 def bytes_get_state(obj: Any, save_context: SaveContext) -> dict[str, Any]:
     f_name = f"{uuid.uuid4()}.bin"

skops/io/_scipy.py

@@ -40,9 +40,9 @@ def __init__(
         trusted: bool | Sequence[str] = False,
     ) -> None:
         super().__init__(state, load_context, trusted)
-        type = state["type"]
+        self.type = state["type"]
         self.trusted = self._get_trusted(trusted, [spmatrix])
-        if type != "scipy":
+        if self.type != "scipy":
             raise TypeError(
                 f"Cannot load object of type {self.module_name}.{self.class_name}"
             )