Merge patch/v0.10.5 into master

CHANGELOG.md

@@ -1,7 +1,7 @@
 # Change Log
 
 ## v0.11.0dev
-[Full Changelog](https://github.com/SeldonIO/alibi-detect/compare/v0.10.3...master)
+[Full Changelog](https://github.com/SeldonIO/alibi-detect/compare/v0.10.5...master)
 
 ### Added
 - **New feature** MMD drift detector has been extended with a [KeOps](https://www.kernel-operations.io/keops/index.html) backend to scale and speed up the detector. 
@@ -26,6 +26,15 @@ See the [documentation](https://docs.seldon.io/projects/alibi-detect/en/latest/c
 - UTF-8 decoding is enforced when `README.md` is opened by `setup.py`. This is to prevent pip install errors on systems with `PYTHONIOENCODING` set to use other encoders ([#605](https://github.com/SeldonIO/alibi-detect/pull/605)).
 - Skip specific save/load tests that require downloading remote artefacts if the relevant URI(s) is/are down ([#607](https://github.com/SeldonIO/alibi-detect/pull/607)).
 
+## v0.10.5
+## [v0.10.5](https://github.com/SeldonIO/alibi-detect/tree/v0.10.5) (2023-01-26)
+[Full Changelog](https://github.com/SeldonIO/alibi-detect/compare/v0.10.4...v0.10.5)
+
+### Fixed
+- Fixed two bugs preventing backward compatibility when loading detectors saved with `<v0.10.0`
+([#729](https://github.com/SeldonIO/alibi-detect/pull/729) and [#732](https://github.com/SeldonIO/alibi-detect/pull/732)). This bug also meant that detectors
+saved with `save_detector(..., legacy=True)` in `>=v0.10.0` did not properly obey the legacy file format. The `config.toml` file format used by default in `>=v0.10.0` is unaffected. 
+
 ## v0.10.4
 ## [v0.10.4](https://github.com/SeldonIO/alibi-detect/tree/v0.10.4) (2022-10-21)
 [Full Changelog](https://github.com/SeldonIO/alibi-detect/compare/v0.10.3...v0.10.4)

CITATION.cff

@@ -19,6 +19,6 @@ authors:
 - family-names: "Athorne"
   given-names: "Alex"
 title: "Alibi Detect: Algorithms for outlier, adversarial and drift detection"
-version: 0.10.4
-date-released: 2022-10-21
+version: 0.10.5
+date-released: 2023-01-26
 url: "https://github.com/SeldonIO/alibi-detect"

README.md

@@ -407,8 +407,8 @@ BibTeX entry:
   title = {Alibi Detect: Algorithms for outlier, adversarial and drift detection},
   author = {Van Looveren, Arnaud and Klaise, Janis and Vacanti, Giovanni and Cobb, Oliver and Scillitoe, Ashley and Samoilescu, Robert and Athorne, Alex},
   url = {https://github.com/SeldonIO/alibi-detect},
-  version = {0.10.4},
-  date = {2022-10-21},
+  version = {0.10.5},
+  date = {2023-01-26},
   year = {2019}
 }
 ```

alibi_detect/saving/_tensorflow/loading.py

@@ -36,7 +36,7 @@
 
 
 def load_model(filepath: Union[str, os.PathLike],
-               load_dir: str = 'model',
+               filename: str = 'model',
                custom_objects: dict = None,
                layer: Optional[int] = None,
                ) -> tf.keras.Model:
@@ -47,8 +47,8 @@ def load_model(filepath: Union[str, os.PathLike],
     ----------
     filepath
         Saved model directory.
-    load_dir
-        Name of saved model folder within the filepath directory.
+    filename
+        Name of saved model within the filepath directory.
     custom_objects
         Optional custom objects when loading the TensorFlow model.
     layer
@@ -60,11 +60,12 @@ def load_model(filepath: Union[str, os.PathLike],
     Loaded model.
     """
     # TODO - update this to accept tf format - later PR.
-    model_dir = Path(filepath).joinpath(load_dir)
+    model_dir = Path(filepath)
+    model_name = filename + '.h5'
     # Check if model exists
-    if 'model.h5' not in [f.name for f in model_dir.glob('[!.]*.h5')]:
-        raise FileNotFoundError(f'No .h5 file found in {model_dir}.')
-    model = tf.keras.models.load_model(model_dir.joinpath('model.h5'), custom_objects=custom_objects)
+    if model_name not in [f.name for f in model_dir.glob('[!.]*.h5')]:
+        raise FileNotFoundError(f'{model_name} not found in {model_dir.resolve()}.')
+    model = tf.keras.models.load_model(model_dir.joinpath(model_name), custom_objects=custom_objects)
     # Optionally extract hidden layer
     if isinstance(layer, int):
         model = HiddenOutput(model, layer=layer)
@@ -233,7 +234,17 @@ def load_detector_legacy(filepath: Union[str, os.PathLike], suffix: str, **kwarg
     # load outlier detector specific parameters
     state_dict = dill.load(open(filepath.joinpath(detector_name + suffix), 'rb'))
 
+    # Update the drift detector preprocess kwargs if state_dict is from an old alibi-detect version (<v0.10).
+    # See https://github.com/SeldonIO/alibi-detect/pull/732
+    if 'kwargs' in state_dict and 'other' in state_dict:  # A drift detector if both of these exist
+        if 'x_ref_preprocessed' not in state_dict['kwargs']:  # if already exists then must have been saved w/ >=v0.10
+            # Set x_ref_preprocessed to True
+            state_dict['kwargs']['x_ref_preprocessed'] = True
+            # Move `preprocess_x_ref` from `other` to `kwargs`
+            state_dict['kwargs']['preprocess_x_ref'] = state_dict['other']['preprocess_x_ref']
+
     # initialize detector
+    model_dir = filepath.joinpath('model')
     detector: Optional[Detector] = None  # to avoid mypy errors
     if detector_name == 'OutlierAE':
         ae = load_tf_ae(filepath)
@@ -254,13 +265,13 @@ def load_detector_legacy(filepath: Union[str, os.PathLike], suffix: str, **kwarg
     elif detector_name == 'AdversarialAE':
         ae = load_tf_ae(filepath)
         custom_objects = kwargs['custom_objects'] if 'custom_objects' in k else None
-        model = load_model(filepath, custom_objects=custom_objects)
+        model = load_model(model_dir, custom_objects=custom_objects)
         model_hl = load_tf_hl(filepath, model, state_dict)
         detector = init_ad_ae(state_dict, ae, model, model_hl)
     elif detector_name == 'ModelDistillation':
-        md = load_model(filepath, load_dir='distilled_model')
+        md = load_model(model_dir, filename='distilled_model')
         custom_objects = kwargs['custom_objects'] if 'custom_objects' in k else None
-        model = load_model(filepath, custom_objects=custom_objects)
+        model = load_model(model_dir, custom_objects=custom_objects)
         detector = init_ad_md(state_dict, md, model)
     elif detector_name == 'OutlierProphet':
         detector = init_od_prophet(state_dict)  # type: ignore[assignment]
@@ -274,8 +285,9 @@ def load_detector_legacy(filepath: Union[str, os.PathLike], suffix: str, **kwarg
         if state_dict['other']['load_text_embedding']:
             emb, tokenizer = load_text_embed(filepath)
         try:  # legacy load_model behaviour was to return None if not found. Now it raises error, hence need try-except.
-            model = load_model(filepath, load_dir='encoder')
+            model = load_model(model_dir, filename='encoder')
         except FileNotFoundError:
+            logger.warning('No model found in {}, setting `model` to `None`.'.format(model_dir))
             model = None
         if detector_name == 'KSDrift':
             load_fn = init_cd_ksdrift  # type: ignore[assignment]
@@ -287,7 +299,7 @@ def load_detector_legacy(filepath: Union[str, os.PathLike], suffix: str, **kwarg
             load_fn = init_cd_tabulardrift  # type: ignore[assignment]
         elif detector_name == 'ClassifierDriftTF':
             # Don't need try-except here since model is not optional for ClassifierDrift
-            clf_drift = load_model(filepath, load_dir='clf_drift')
+            clf_drift = load_model(model_dir, filename='clf_drift')
             load_fn = partial(init_cd_classifierdrift, clf_drift)  # type: ignore[assignment]
         else:
             raise NotImplementedError

alibi_detect/saving/_tensorflow/saving.py

@@ -81,7 +81,7 @@ def save_model_config(model: Callable,
 
     if model is not None:
         filepath = base_path.joinpath(local_path)
-        save_model(model, filepath=filepath, save_dir='model')
+        save_model(model, filepath=filepath.joinpath('model'))
         cfg_model = {
             'flavour': Framework.TENSORFLOW.value,
             'src': local_path.joinpath('model')
@@ -91,7 +91,7 @@ def save_model_config(model: Callable,
 
 def save_model(model: tf.keras.Model,
                filepath: Union[str, os.PathLike],
-               save_dir: Union[str, os.PathLike] = 'model',
+               filename: str = 'model',
                save_format: Literal['tf', 'h5'] = 'h5') -> None:  # TODO - change to tf, later PR
     """
     Save TensorFlow model.
@@ -102,20 +102,20 @@ def save_model(model: tf.keras.Model,
         The tf.keras.Model to save.
     filepath
         Save directory.
-    save_dir
-        Name of folder to save to within the filepath directory.
+    filename
+        Name of file to save to within the filepath directory.
     save_format
         The format to save to. 'tf' to save to the newer SavedModel format, 'h5' to save to the lighter-weight
         legacy hdf5 format.
     """
     # create folder to save model in
-    model_path = Path(filepath).joinpath(save_dir)
+    model_path = Path(filepath)
     if not model_path.is_dir():
         logger.warning('Directory {} does not exist and is now created.'.format(model_path))
         model_path.mkdir(parents=True, exist_ok=True)
 
     # save model
-    model_path = model_path.joinpath('model.h5') if save_format == 'h5' else model_path
+    model_path = model_path.joinpath(filename + '.h5') if save_format == 'h5' else model_path
 
     if isinstance(model, tf.keras.Model):
         model.save(model_path, save_format=save_format)
@@ -254,30 +254,31 @@ def save_detector_legacy(detector, filepath):
         dill.dump(state_dict, f)
 
     # save detector specific TensorFlow models
+    model_dir = filepath.joinpath('model')
     if isinstance(detector, OutlierAE):
         save_tf_ae(detector, filepath)
     elif isinstance(detector, OutlierVAE):
         save_tf_vae(detector, filepath)
     elif isinstance(detector, (ChiSquareDrift, ClassifierDrift, KSDrift, MMDDrift, TabularDrift)):
         if model is not None:
-            save_model(model, filepath, save_dir='encoder')
+            save_model(model, model_dir, filename='encoder')
         if embed is not None:
             save_embedding_legacy(embed, embed_args, filepath)
         if tokenizer is not None:
             tokenizer.save_pretrained(filepath.joinpath('model'))
         if detector_name == 'ClassifierDriftTF':
-            save_model(clf_drift, filepath, save_dir='clf_drift')
+            save_model(clf_drift, model_dir, filename='clf_drift')
     elif isinstance(detector, OutlierAEGMM):
         save_tf_aegmm(detector, filepath)
     elif isinstance(detector, OutlierVAEGMM):
         save_tf_vaegmm(detector, filepath)
     elif isinstance(detector, AdversarialAE):
         save_tf_ae(detector, filepath)
-        save_model(detector.model, filepath)
+        save_model(detector.model, model_dir)
         save_tf_hl(detector.model_hl, filepath)
     elif isinstance(detector, ModelDistillation):
-        save_model(detector.distilled_model, filepath, save_dir='distilled_model')
-        save_model(detector.model, filepath, save_dir='model')
+        save_model(detector.distilled_model, model_dir, filename='distilled_model')
+        save_model(detector.model, model_dir, filename='model')
     elif isinstance(detector, OutlierSeq2Seq):
         save_tf_s2s(detector, filepath)
     elif isinstance(detector, LLR):

alibi_detect/saving/loading.py

@@ -281,7 +281,7 @@ def _load_model_config(cfg: dict) -> Callable:
                                 "a compatible model.")
 
     if flavour == Framework.TENSORFLOW:
-        model = load_model_tf(src, load_dir='.', custom_objects=custom_obj, layer=layer)
+        model = load_model_tf(src, custom_objects=custom_obj, layer=layer)
     elif flavour == Framework.PYTORCH:
         model = load_model_pt(src, layer=layer)
     elif flavour == Framework.SKLEARN:

alibi_detect/utils/tests/test_saving_legacy.py

@@ -1,6 +1,6 @@
 """
 Tests for saving/loading of detectors with legacy .dill state_dict. As legacy save/load functionality becomes
-deprecated, these tests will be removed, and more tests will be added to test_savin.py.
+deprecated, these tests will be removed, and more tests will be added to test_saving.py.
 """
 from alibi_detect.utils.missing_optional_dependency import MissingDependency
 from functools import partial