Merge branch 'master' into feature/add-ci-linting

.github/workflows/ci.yml

@@ -84,9 +84,12 @@ jobs:
  # are removed from tests, this can be removed, allowing all tests to use random seeds.
 
  - name: Upload coverage to Codecov
- if: ${{ success() }}
- run: |
- codecov -F ${{ matrix.os }}-${{ matrix.python-version }}
+ uses: codecov/codecov-action@v3
+ with:
+ directory: .
+ env_vars: ${{matrix.os}}, ${{matrix.python-version}}
+ fail_ci_if_error: false
+ verbose: true
 
  - name: Build Python package
  run: |

.pre-commit-config.yaml

@@ -1,13 +1,13 @@
 repos:
- - repo: https://github.com/pre-commit/pre-commit-hooks
- rev: v2.4.0
+ - repo: https://github.com/pycqa/flake8
+ rev: 6.0.0
  hooks:
  - id: flake8
  - repo: https://github.com/pre-commit/mirrors-mypy
  rev: v1.0.1
  hooks:
  - id: mypy
  additional_dependencies: [
- types-requests~=2.25,
+ types-requests~=2.28,
  types-toml~=0.10
  ]

CHANGELOG.md

@@ -1,7 +1,27 @@
 # Change Log
 
-## v0.12.0dev
-[Full Changelog](https://github.com/SeldonIO/alibi-detect/compare/v0.11.1...master)
+## [v0.11.2](https://github.com/SeldonIO/alibi-detect/tree/v0.11.2) (2023-04-28)
+[Full Changelog](https://github.com/SeldonIO/alibi-detect/compare/v0.11.1...v0.11.2)
+
+### Fixed
+- Failure of `plot_feature_outlier_image` utility function when no outliers are detected ([#774](https://github.com/SeldonIO/alibi-detect/pull/774) - thanks [@signupatgmx](https://github.com/signupatgmx) !).
+
+### Changed
+ - Refactored methods that use `tensorflow` optimizers to work with the new optimizers introduced in `2.11` ([#739](https://github.com/SeldonIO/alibi-detect/pull/739)).
+ - Maximum supported version of `tensorflow` bumped to `2.12.x` ([#764](https://github.com/SeldonIO/alibi-detect/pull/764)).
+ - Maximum supported version of `tensorflow-probability` version to `0.19.x` ([#687](https://github.com/SeldonIO/alibi-detect/pull/687)).
+ - Supported version of `pandas` bumped to `>1.0.0, <3.0.0` ([#765](https://github.com/SeldonIO/alibi-detect/pull/765)).
+ - Maximum supported version of `scikit-image` bumped to `0.20.x` ([#751](https://github.com/SeldonIO/alibi-detect/pull/751)).
+
+### Development
+- Migrate `codecov` to use Github Actions and don't fail CI on coverage report upload failure due to rate limiting ([#768](https://github.com/SeldonIO/alibi-detect/pull/768), [#776](https://github.com/SeldonIO/alibi-detect/pull/776)).
+ - Bump `mypy` version to `>=1.0, <2.0` ([#754](https://github.com/SeldonIO/alibi-detect/pull/754)). 
+ - Bump `sphinx` version to `6.x` ([#709](https://github.com/SeldonIO/alibi-detect/pull/709)).
+ - Bump `sphinx-design` version to `0.4.1` ([#769](https://github.com/SeldonIO/alibi-detect/pull/769)).
+ - Bump `nbsphinx` version to `0.9.x` ([#757](https://github.com/SeldonIO/alibi-detect/pull/757)).
+ - Bump `myst-parser` version to `>=1.0, <2.0` ([#756](https://github.com/SeldonIO/alibi-detect/pull/756)).
+ - Bump `twine` version to `4.x` ([#511](https://github.com/SeldonIO/alibi-detect/pull/511)).
+ - Bump `pre-commit` version to `3.x` and update the config ([#731](https://github.com/SeldonIO/alibi-detect/pull/731)).
 
 ## v0.11.1
 [Full Changelog](https://github.com/SeldonIO/alibi-detect/compare/v0.11.0...v0.11.1)

CITATION.cff

@@ -19,6 +19,6 @@ authors:
 - family-names: "Athorne"
  given-names: "Alex"
 title: "Alibi Detect: Algorithms for outlier, adversarial and drift detection"
-version: 0.11.1
-date-released: 2023-03-03
+version: 0.11.2
+date-released: 2023-04-28
 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 = {0.11.1},
- date = {2023-03-03},
+ version = {0.11.2},
+ date = {2023-04-28},
  year = {2019}
 }
 ```

alibi_detect/ad/adversarialae.py

@@ -10,6 +10,7 @@
 from alibi_detect.models.tensorflow.losses import loss_adv_ae
 from alibi_detect.models.tensorflow.trainer import trainer
 from alibi_detect.utils.tensorflow.prediction import predict_batch
+from alibi_detect.utils._types import OptimizerTF
 from tensorflow.keras.layers import Dense, Flatten
 from tensorflow.keras.losses import kld
 from tensorflow.keras.models import Model
@@ -139,7 +140,7 @@ def fit(self,
  loss_fn: tf.keras.losses = loss_adv_ae,
  w_model: float = 1.,
  w_recon: float = 0.,
- optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam(learning_rate=1e-3),
+ optimizer: OptimizerTF = tf.keras.optimizers.Adam,
  epochs: int = 20,
  batch_size: int = 128,
  verbose: bool = True,
@@ -177,6 +178,7 @@ def fit(self,
  """
  # train arguments
  args = [self.ae, loss_fn, X]
+ optimizer = optimizer() if isinstance(optimizer, type) else optimizer
  kwargs = {
  'optimizer': optimizer,
  'epochs': epochs,

alibi_detect/ad/model_distillation.py

@@ -8,6 +8,7 @@
 from alibi_detect.models.tensorflow.losses import loss_distillation
 from alibi_detect.models.tensorflow.trainer import trainer
 from alibi_detect.utils.tensorflow.prediction import predict_batch
+from alibi_detect.utils._types import OptimizerTF
 from tensorflow.keras.losses import categorical_crossentropy, kld
 
 logger = logging.getLogger(__name__)
@@ -67,7 +68,7 @@ def __init__(self,
  def fit(self,
  X: np.ndarray,
  loss_fn: tf.keras.losses = loss_distillation,
- optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam(learning_rate=1e-3),
+ optimizer: OptimizerTF = tf.keras.optimizers.Adam,
  epochs: int = 20,
  batch_size: int = 128,
  verbose: bool = True,
@@ -101,6 +102,7 @@ def fit(self,
  """
  # train arguments
  args = [self.distilled_model, loss_fn, X]
+ optimizer = optimizer() if isinstance(optimizer, type) else optimizer
  kwargs = {
  'optimizer': optimizer,
  'epochs': epochs,

alibi_detect/cd/tensorflow/classifier.py

@@ -11,6 +11,7 @@
 from alibi_detect.utils.tensorflow.prediction import predict_batch
 from alibi_detect.utils.warnings import deprecated_alias
 from alibi_detect.utils.frameworks import Framework
+from alibi_detect.utils._types import OptimizerTF
 
 
 class ClassifierDriftTF(BaseClassifierDrift):
@@ -31,7 +32,7 @@ def __init__(
  n_folds: Optional[int] = None,
  retrain_from_scratch: bool = True,
  seed: int = 0,
- optimizer: tf.keras.optimizers.Optimizer = tf.keras.optimizers.Adam,
+ optimizer: OptimizerTF = tf.keras.optimizers.Adam,
  learning_rate: float = 1e-3,
  batch_size: int = 32,
  preprocess_batch_fn: Optional[Callable] = None,
@@ -176,8 +177,12 @@ def score(self, x: np.ndarray) -> Tuple[float, float, np.ndarray, np.ndarray, #
  else:
  raise TypeError(f'x needs to be of type np.ndarray or list and not {type(x)}.')
  ds_tr = self.dataset(x_tr, y_tr)
- self.model = clone_model(self.original_model) if self.retrain_from_scratch \
- else self.model
+ if self.retrain_from_scratch:
+ # clone model to re-initialise
+ self.model = clone_model(self.original_model)
+ # Clone optimizer to prevent error due to cloned model (with new tf>=2.11 optimizers)
+ optimizer = self.train_kwargs['optimizer']
+ self.train_kwargs['optimizer'] = optimizer.__class__.from_config(optimizer.get_config())
  train_args = [self.model, self.loss_fn, None]
  self.train_kwargs.update({'dataset': ds_tr})
  trainer(*train_args, **self.train_kwargs) # type: ignore

alibi_detect/cd/tensorflow/learned_kernel.py

@@ -8,6 +8,7 @@
 from alibi_detect.utils.tensorflow.distance import mmd2_from_kernel_matrix, batch_compute_kernel_matrix
 from alibi_detect.utils.warnings import deprecated_alias
 from alibi_detect.utils.frameworks import Framework
+from alibi_detect.utils._types import OptimizerTF
 
 
 class LearnedKernelDriftTF(BaseLearnedKernelDrift):
@@ -26,7 +27,7 @@ def __init__(
  reg_loss_fn: Callable = (lambda kernel: 0),
  train_size: Optional[float] = .75,
  retrain_from_scratch: bool = True,
- optimizer: tf.keras.optimizers.Optimizer = tf.keras.optimizers.Adam,
+ optimizer: OptimizerTF = tf.keras.optimizers.Adam,
  learning_rate: float = 1e-3,
  batch_size: int = 32,
  batch_size_predict: int = 32,
@@ -204,7 +205,7 @@ def score(self, x: Union[np.ndarray, list]) -> Tuple[float, float, float]:
  def trainer(
  j_hat: JHat,
  datasets: Tuple[tf.keras.utils.Sequence, tf.keras.utils.Sequence],
- optimizer: tf.keras.optimizers.Optimizer = tf.keras.optimizers.Adam,
+ optimizer: OptimizerTF = tf.keras.optimizers.Adam,
  learning_rate: float = 1e-3,
  preprocess_fn: Callable = None,
  epochs: int = 20,

alibi_detect/cd/tensorflow/mmd_online.py

@@ -166,7 +166,7 @@ def _configure_thresholds(self):
  2 * tf.reduce_sum(k_xy_col_sums[w:w + w_size]))
  for k_xx_sum, y_inds_w, k_xy_col_sums in zip(k_xx_sums_all, y_inds_all_w, k_xy_col_sums_all)
  ]
- mmds = tf.concat(mmds, axis=0) # an mmd for each bootstrap sample
+ mmds = tf.stack(mmds, axis=0) # an mmd for each bootstrap sample
 
  # Now we discard all bootstrap samples for which mmd is in top (1/ert)% and record the thresholds
  thresholds.append(quantile(mmds, 1 - self.fpr))

alibi_detect/models/pytorch/gmm.py

@@ -0,0 +1,42 @@
+from torch import nn
+import torch
+
+
+class GMMModel(nn.Module):
+ def __init__(self, n_components: int, dim: int) -> None:
+ """Gaussian Mixture Model (GMM).
+
+ Parameters
+ ----------
+ n_components
+ The number of mixture components.
+ dim
+ The dimensionality of the data.
+ """
+ super().__init__()
+ self.weight_logits = nn.Parameter(torch.zeros(n_components))
+ self.means = nn.Parameter(torch.randn(n_components, dim))
+ self.inv_cov_factor = nn.Parameter(torch.randn(n_components, dim, dim)/10)
+
+ @property
+ def _inv_cov(self) -> torch.Tensor:
+ return torch.bmm(self.inv_cov_factor, self.inv_cov_factor.transpose(1, 2))
+
+ @property
+ def _weights(self) -> torch.Tensor:
+ return nn.functional.softmax(self.weight_logits, dim=0)
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ """Compute the log-likelihood of the data.
+
+ Parameters
+ ----------
+ x
+ Data to score.
+ """
+ det = torch.linalg.det(self._inv_cov) # Note det(A^-1)=1/det(A)
+ to_means = x[:, None, :] - self.means[None, :, :]
+ likelihood = ((-0.5 * (
+ torch.einsum('bke,bke->bk', (torch.einsum('bkd,kde->bke', to_means, self._inv_cov), to_means))
+ )).exp()*det[None, :]*self._weights[None, :]).sum(-1)
+ return -likelihood.log()

alibi_detect/models/tensorflow/trainer.py

@@ -10,7 +10,7 @@ def trainer(
  x_train: np.ndarray,
  y_train: np.ndarray = None,
  dataset: tf.keras.utils.Sequence = None,
- optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam(learning_rate=1e-3),
+ optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam,
  loss_fn_kwargs: dict = None,
  preprocess_fn: Callable = None,
  epochs: int = 20,
@@ -57,6 +57,7 @@ def trainer(
  callbacks
  Callbacks used during training.
  """
+ optimizer = optimizer() if isinstance(optimizer, type) else optimizer
  return_xy = False if not isinstance(dataset, tf.keras.utils.Sequence) and y_train is None else True
  if not isinstance(dataset, tf.keras.utils.Sequence): # create dataset
  train_data = x_train if y_train is None else (x_train, y_train)