GSK-1191 Added support of scan Transformation Function to the Hub

docs/knowledge/catalogs/transformation-function-catalog/index.rst

@@ -4,6 +4,15 @@ Transformation functions
 - **Text transformation function**
 
   - :func:`~.giskard.functions.transformation.keyboard_typo_transformation`
-  - :func:`~.giskard.functions.transformation.uppercase_transformation`
-  - :func:`~.giskard.functions.transformation.lowercase_transformation`
-  - :func:`~.giskard.functions.transformation.strip_punctuation`
+  - :func:`~.giskard.functions.transformation.text_uppercase`
+  - :func:`~.giskard.functions.transformation.text_lowercase`
+  - :func:`~.giskard.functions.transformation.text_title_case`
+  - :func:`~.giskard.functions.transformation.text_typo`
+  - :func:`~.giskard.functions.transformation.text_typo_from_ocr`
+  - :func:`~.giskard.functions.transformation.text_punctuation_removal`
+  - :func:`~.giskard.functions.transformation.text_accent_removal`
+  - :func:`~.giskard.functions.transformation.text_gender_switch`
+  - :func:`~.giskard.functions.transformation.text_number_to_word`
+  - :func:`~.giskard.functions.transformation.text_religion_switch`
+  - :func:`~.giskard.functions.transformation.text_nationality_switch`
+  - :func:`~.giskard.functions.transformation.text_typo_from_speech`

docs/reference/transformation-functions/index.rst

@@ -15,10 +15,18 @@ Textual transformation functions
 --------------------------------
 
 .. autofunction:: giskard.functions.transformation.keyboard_typo_transformation
-.. autofunction:: giskard.functions.transformation.uppercase_transformation
-.. autofunction:: giskard.functions.transformation.lowercase_transformation
-.. autofunction:: giskard.functions.transformation.strip_punctuation
-.. autofunction:: giskard.functions.transformation.change_writing_style
+.. autofunction:: giskard.functions.transformation.text_uppercase
+.. autofunction:: giskard.functions.transformation.text_lowercase
+.. autofunction:: giskard.functions.transformation.text_title_case
+.. autofunction:: giskard.functions.transformation.text_typo
+.. autofunction:: giskard.functions.transformation.text_typo_from_ocr
+.. autofunction:: giskard.functions.transformation.text_punctuation_removal
+.. autofunction:: giskard.functions.transformation.text_accent_removal
+.. autofunction:: giskard.functions.transformation.text_gender_switch
+.. autofunction:: giskard.functions.transformation.text_number_to_word
+.. autofunction:: giskard.functions.transformation.text_religion_switch
+.. autofunction:: giskard.functions.transformation.text_nationality_switch
+.. autofunction:: giskard.functions.transformation.text_typo_from_speech
 
 
 Special transformations used by the scan

giskard/functions/transformation.py

@@ -1,14 +1,12 @@
 import random
 import re
-import string
 
-import numpy as np
 import pandas as pd
 from scipy.stats import median_abs_deviation
 
-from giskard.registry.transformation_function import transformation_function
-
+from ..datasets import Dataset
 from ..llm import get_default_client
+from ..registry.transformation_function import transformation_function
 
 nearbykeys = {
     "a": ["q", "w", "s", "x", "z"],
@@ -71,39 +69,90 @@ def keyboard_typo_transformation(text: str, rate: float = 0.1) -> str:
     return " ".join(words)
 
 
-@transformation_function(name="To uppercase", tags=["text"], cell_level=True)
-def uppercase_transformation(text: str) -> str:
-    """
-    Transform the text to uppercase
-    """
-    return np.nan if pd.isnull(text) else text.upper()
+@transformation_function(name="Transform to uppercase", row_level=False)
+def text_uppercase(data: pd.DataFrame, column: str):
+    from ..scanner.robustness.text_transformations import TextUppercase
 
+    return TextUppercase(column).execute(data)
 
-@transformation_function(name="To lowercase", tags=["text"], cell_level=True)
-def lowercase_transformation(text: str) -> str:
-    """
-    Transform the text of the column 'column_name' to lowercase
-    """
-    return np.nan if pd.isnull(text) else text.lower()
 
+@transformation_function(name="Transform to lowercase", row_level=False)
+def text_lowercase(data: pd.DataFrame, column: str):
+    from ..scanner.robustness.text_transformations import TextLowercase
 
-@transformation_function(name="Strip punctuation", tags=["text"], cell_level=True)
-def strip_punctuation(text: str) -> str:
-    """
-    Remove all punctuation symbols (e.g., ., !, ?) from the text of the column 'column_name'
-    """
-    if pd.isnull(text):
-        return text
+    return TextLowercase(column).execute(data)
+
+
+@transformation_function(name="Transform to title case", row_level=False)
+def text_title_case(data: pd.DataFrame, column: str):
+    from ..scanner.robustness.text_transformations import TextTitleCase
+
+    return TextTitleCase(column).execute(data)
+
+
+@transformation_function(name="Add typos", row_level=False)
+def text_typo(data: pd.DataFrame, column: str, rate: float = 0.05, min_length: int = 10, rng_seed: int = 1729):
+    from ..scanner.robustness.text_transformations import TextTypoTransformation
+
+    return TextTypoTransformation(column, rate, min_length, rng_seed).execute(data)
+
+
+@transformation_function(name="Add typos from OCR", row_level=False)
+def text_typo_from_ocr(data: pd.DataFrame, column: str, rate: float = 0.05, min_length: int = 10, rng_seed: int = 1729):
+    from ..scanner.robustness.text_transformations import TextFromOCRTypoTransformation
+
+    return TextFromOCRTypoTransformation(column, rate, min_length, rng_seed).execute(data)
+
+
+@transformation_function(name="Punctuation Removal", row_level=False)
+def text_punctuation_removal(data: pd.DataFrame, column: str):
+    from ..scanner.robustness.text_transformations import TextPunctuationRemovalTransformation
+
+    return TextPunctuationRemovalTransformation(column).execute(data)
+
+
+@transformation_function(name="Accent Removal", row_level=False)
+def text_accent_removal(data: pd.DataFrame, column: str, rate: float = 1.0, rng_seed: int = 1729):
+    from ..scanner.robustness.text_transformations import TextAccentRemovalTransformation
+
+    return TextAccentRemovalTransformation(column, rate, rng_seed).execute(data)
+
+
+@transformation_function(name="Switch Gender", row_level=False, needs_dataset=True)
+def text_gender_switch(dataset: Dataset, column: str):
+    from ..scanner.robustness.text_transformations import TextGenderTransformation
+
+    return TextGenderTransformation(column).execute(dataset)
+
+
+@transformation_function(name="Transform numbers to words", row_level=False, needs_dataset=True)
+def text_number_to_word(dataset: Dataset, column: str):
+    from ..scanner.robustness.text_transformations import TextNumberToWordTransformation
+
+    return TextNumberToWordTransformation(column).execute(dataset)
+
+
+@transformation_function(name="Switch Religion", row_level=False, needs_dataset=True)
+def text_religion_switch(dataset: Dataset, column: str):
+    from ..scanner.robustness.text_transformations import TextReligionTransformation
+
+    return TextReligionTransformation(column).execute(dataset)
+
+
+@transformation_function(
+    name="Switch countries from high- to low-income and vice versa", row_level=False, needs_dataset=True
+)
+def text_nationality_switch(dataset: Dataset, column: str):
+    from ..scanner.robustness.text_transformations import TextNationalityTransformation
 
-    split_urls_from_text = gruber.split(text)
+    return TextNationalityTransformation(column).execute(dataset)
 
-    # The non-URLs are always even-numbered entries in the list and the URLs are odd-numbered.
-    for i in range(0, len(split_urls_from_text), 2):
-        split_urls_from_text[i] = split_urls_from_text[i].translate(str.maketrans("", "", string.punctuation))
 
-    stripped_text = "".join(split_urls_from_text)
+@transformation_function(name="Add text from speech typos", row_level=False, needs_dataset=True)
+def text_typo_from_speech(dataset: Dataset, column: str, rng_seed: int = 1729, min_length: int = 10):
+    from ..scanner.robustness.text_transformations import TextFromSpeechTypoTransformation
 
-    return stripped_text
+    return TextFromSpeechTypoTransformation(column, rng_seed, min_length).execute(dataset)
 
 
 @transformation_function(name="Change writing style", row_level=False, tags=["text"])

giskard/registry/transformation_function.py

@@ -32,10 +32,13 @@ class TransformationFunction(RegistryArtifact[DatasetProcessFunctionMeta]):
     def _get_name(cls) -> str:
         return "transformations"
 
-    def __init__(self, func: Optional[TransformationFunctionType], row_level=True, cell_level=False):
+    def __init__(
+        self, func: Optional[TransformationFunctionType], row_level=True, cell_level=False, needs_dataset=False
+    ):
         self.func = func
         self.row_level = row_level
         self.cell_level = cell_level
+        self.needs_dataset = needs_dataset
 
         test_uuid = get_object_uuid(func)
         meta = tests_registry.get_test(test_uuid)
@@ -99,6 +102,7 @@ def transformation_function(
     cell_level=False,
     name=None,
     tags: Optional[List[str]] = None,
+    needs_dataset=False,
 ):
     """
     Decorator that registers a function as a transformation function and returns a TransformationFunction instance.
@@ -130,19 +134,25 @@ def inner(func: Union[TransformationFunctionType, Type[TransformationFunction]])
 
         if inspect.isclass(func) and issubclass(func, TransformationFunction):
             return func
-        return _wrap_transformation_function(func, row_level, cell_level)()
+        return _wrap_transformation_function(func, row_level, cell_level, needs_dataset)()
 
     if callable(_fn):
         return functools.wraps(_fn)(inner(_fn))
     else:
         return inner
 
 
-def _wrap_transformation_function(original: Callable, row_level: bool, cell_level: bool):
-    transformation_fn = functools.wraps(original)(TransformationFunction(original, row_level, cell_level))
+def _wrap_transformation_function(original: Callable, row_level: bool, cell_level: bool, needs_dataset: bool):
+    transformation_fn = functools.wraps(original)(
+        TransformationFunction(original, row_level, cell_level, needs_dataset)
+    )
 
     if not cell_level:
-        validate_arg_type(transformation_fn, 0, pd.Series if row_level else pd.DataFrame)
+        from ..datasets import Dataset
+
+        validate_arg_type(
+            transformation_fn, 0, pd.Series if row_level else (Dataset if needs_dataset else pd.DataFrame)
+        )
 
     drop_arg(transformation_fn, 0)
 

giskard/scanner/robustness/text_transformations.py

@@ -18,8 +18,8 @@
 class TextTransformation(TransformationFunction):
     name: str
 
-    def __init__(self, column):
-        super().__init__(None, row_level=False, cell_level=False)
+    def __init__(self, column, needs_dataset=False):
+        super().__init__(None, row_level=False, cell_level=False, needs_dataset=needs_dataset)
         self.column = column
         self.meta = DatasetProcessFunctionMeta(type="TRANSFORMATION")
         self.meta.uuid = get_object_uuid(self)
@@ -212,10 +212,8 @@ def make_perturbation(self, text):
 
 
 class TextLanguageBasedTransformation(TextTransformation):
-    needs_dataset = True
-
     def __init__(self, column, rng_seed=1729):
-        super().__init__(column)
+        super().__init__(column, needs_dataset=True)
         self._lang_dictionary = dict()
         self._load_dictionaries()
         self.rng = np.random.default_rng(seed=rng_seed)

tests/functions/test_transformation.py

@@ -0,0 +1,32 @@
+import pandas as pd
+
+from giskard import Dataset
+
+
+def _dataset_from_dict(data):
+    return Dataset(pd.DataFrame(data), target=None)
+
+
+def test_gender_transformation():
+    dataset = _dataset_from_dict(
+        {
+            "text": [
+                "We just got this and my daughter loves it. She has played it several times.",
+                "It did not work.",
+                "“They pushed the feature just 1 minute before the user test”",
+                "He hates doing user tests! for his company",
+                "Il déteste faire les tests en langue française",
+            ]
+        }
+    )
+
+    from giskard.functions.transformation import text_gender_switch
+
+    transformed = dataset.transform(text_gender_switch(column="text"))
+
+    transformed_text = transformed.df.text.str.lower().values
+    assert transformed_text[0] == "We just got this and my son loves it. He has played it several times.".lower()
+    assert transformed_text[1] == "It did not work.".lower()
+    assert transformed_text[2] == "“They pushed the feature just 1 minute before the user test”".lower()
+    assert transformed_text[3] == "She hates doing user tests! for her company".lower()
+    assert transformed_text[4] == "Elle déteste faire les tests en langue français".lower()