[GSK-1597] Push typo perturbation stochasticity

python-client/giskard/push/perturbation.py

@@ -12,6 +12,7 @@
 
 import numpy as np
 import pandas as pd
+import sys
 
 from giskard.core.core import SupportedModelTypes
 from giskard.datasets.base import Dataset
@@ -35,7 +36,7 @@
 )
 from ..push import PerturbationPush
 
-text_transfo_list = [
+text_transformation_list = [
     TextLowercase,
     TextUppercase,
     TextTitleCase,
@@ -45,9 +46,7 @@
 ]
 
 
-def create_perturbation_push(
-    model: BaseModel, ds: Dataset, df: pd.DataFrame
-) -> PerturbationPush:
+def create_perturbation_push(model: BaseModel, ds: Dataset, df: pd.DataFrame) -> PerturbationPush:
     """Create a perturbation notification by applying transformations.
 
     Applies supported perturbations to each feature in the dataset
@@ -66,6 +65,7 @@ def create_perturbation_push(
     for feat, coltype in ds.column_types.items():
         coltype = coltype_to_supported_perturbation_type(coltype)
         transformation_info = _apply_perturbation(model, ds, df, feat, coltype)
+        # df contains only one row, which is the sample being looked at in the debugger
         value = df.iloc[0][feat]
         if transformation_info is not None:
             return PerturbationPush(
@@ -108,6 +108,7 @@ def _apply_perturbation(
     transformation_function = list()
     value_perturbed = list()
     transformation_functions_params = list()
+
     passed = False
     # Create a slice of the dataset with only the row to perturb
     ds_slice = Dataset(
@@ -167,9 +168,20 @@ def _text(
 ):
     passed = False
     # Iterate over the possible text transformations
-    for text_transformation in text_transfo_list:
+    for text_transformation in text_transformation_list:
         # Create the transformation
-        t = text_transformation(column=feature)
+        _is_typo_transformation = issubclass(text_transformation, TextTypoTransformation)
+        kwargs = {}
+        if _is_typo_transformation:
+            # TextTypoTransformation generates a random typo for text features. In order to have the same typo per
+            # sample with the push feature in the debugger, we need to generate a unique seed per sample (hashed_seed)
+            # to guarantee the same perturbation per sample.
+            hashed_seed = hash(f"{', '.join(map(lambda x: repr(x), ds_slice_copy.df.values))}".encode("utf-8"))
+            # hash could give negative ints, and np.random.seed accepts only positive ints
+            positive_hashed_seed = hashed_seed % ((sys.maxsize + 1) * 2)
+            kwargs = {"rng_seed": positive_hashed_seed}
+
+        t = text_transformation(column=feature, **kwargs)
 
         # Transform the slice
         transformed = ds_slice_copy.transform(t)
@@ -206,6 +218,7 @@ def _numeric(
             np.unique(np.linspace(-2 * mad, 0, num=10).round().astype(int)),
             np.unique(np.linspace(2 * mad, 0, num=10).round().astype(int)),
         ]
+    # df contains only one row, which is the sample being looked at in the debugger
     value_to_perturb = ds.df[feature].iloc[0]
     for values_added in values_added_list:
         for value in values_added:
@@ -222,17 +235,13 @@ def _numeric(
                 if perturbed:
                     value_perturbed.append(transformed.df[feature].values.item(0))
                     transformation_function.append(t)
-                    transformation_functions_params.append(
-                        dict(column_name=feature, value_added=float(value))
-                    )
+                    transformation_functions_params.append(dict(column_name=feature, value_added=float(value)))
                 else:
                     break
     return len(transformation_function) > 0
 
 
-def _check_after_perturbation(
-    model: BaseModel, ref_row: Dataset, row_perturbed: Dataset
-) -> bool:
+def _check_after_perturbation(model: BaseModel, ref_row: Dataset, row_perturbed: Dataset) -> bool:
     """
     Check if perturbation changed the model's prediction.