Merge branch 'dev_1.18.0' into dev_1.18.0_scaling

.github/actions/deepspeech-v3/Dockerfile

@@ -1,5 +1,5 @@
-pod# Get base from a pytorch image
-FROM pytorch/pytorch:2.1.1-cuda12.1-cudnn8-runtime
+# Get base from a pytorch image
+FROM python:3.10.13-slim-bullseye
 
 # Set to install things in non-interactive mode
 ENV DEBIAN_FRONTEND noninteractive
@@ -26,10 +26,10 @@ RUN apt-get update \
 RUN pip install --ignore-installed PyYAML torch==2.1.1 tensorflow==2.14.1 torchaudio==2.1.1 pytorch-lightning==2.1.2 scikit-learn==1.3.2
 RUN pip install --no-build-isolation fairscale==0.4.13
 
-RUN git clone https://github.com/SeanNaren/deepspeech.pytorch.git
-RUN cd deepspeech.pytorch && sed -i '/^sklearn/d' requirements.txt && pip install -r requirements.txt && pip install -e .
+RUN git clone https://github.com/SeanNaren/deepspeech.pytorch.git && cd deepspeech.pytorch && sed -i '/^sklearn/d' requirements.txt && pip install -r requirements.txt && pip install -e .
 
 RUN pip install numba==0.56.4 pytest-cov==4.1.0 pydub==0.25.1
-RUN pip list
 
 RUN mkdir -p /root/.art/data && cd /root/.art/data && curl -LJO "https://github.com/SeanNaren/deepspeech.pytorch/releases/download/V3.0/librispeech_pretrained_v3.ckpt"
+
+CMD ["/bin/bash"]

.github/workflows/ci-lingvo.yml

@@ -50,7 +50,7 @@ jobs:
           sudo apt-get update
           sudo apt-get -y -q install ffmpeg libavcodec-extra
           python -m pip install --upgrade pip setuptools wheel
-          pip install -q -r <(sed '/^scipy/d;/^matplotlib/d;/^pandas/d;/^statsmodels/d;/^numba/d;/^jax/d;/^h5py/d;/^Pillow/d;/^pytest/d;/^pytest-mock/d;/^torch/d;/^torchaudio/d;/^torchvision/d;/^xgboost/d;/^requests/d;/^tensorflow/d;/^keras/d;/^kornia/d;/^librosa/d;/^tqdm/d;/^timm/d;/^catboost/d' requirements_test.txt)
+          pip install -q -r <(sed '/^scipy/d;/^matplotlib/d;/^pandas/d;/^statsmodels/d;/^numba/d;/^jax/d;/^h5py/d;/^Pillow/d;/^pytest/d;/^pytest-mock/d;/^torch/d;/^torchaudio/d;/^torchvision/d;/^xgboost/d;/^requests/d;/^tensorflow/d;/^keras/d;/^kornia/d;/^librosa/d;/^tqdm/d;/^timm/d;/^catboost/d;/^scikit-learn/d' requirements_test.txt)
           pip install scipy==1.5.4
           pip install matplotlib==3.3.4
           pip install pandas==1.1.5
@@ -75,6 +75,7 @@ jobs:
           pip install librosa==0.9.2
           pip install tqdm==4.64.1
           pip install catboost==1.1.1
+          pip install scikit-learn==0.24.2
           pip list
       - name: Run ${{ matrix.name }} Tests
         run: pytest --cov-report=xml --cov=art --cov-append -q -vv tests/estimators/speech_recognition/test_tensorflow_lingvo.py --framework=${{ matrix.framework }} --durations=0

.github/workflows/ci-scikit-learn.yml

@@ -28,18 +28,14 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - name: scikit-learn 1.1.3 (Python 3.9)
-            framework: scikitlearn
-            scikit-learn: 1.1.3
-            python: 3.9
-          - name: scikit-learn 1.2.2 (Python 3.10)
-            framework: scikitlearn
-            scikit-learn: 1.2.2
-            python: '3.10'
           - name: scikit-learn 1.3.2 (Python 3.10)
             framework: scikitlearn
             scikit-learn: 1.3.2
             python: '3.10'
+          - name: scikit-learn 1.4.0 (Python 3.10)
+            framework: scikitlearn
+            scikit-learn: 1.4.0
+            python: '3.10'
 
     name: ${{ matrix.name }}
     steps:

.github/workflows/ci-tensorflow-v1.yml

@@ -48,7 +48,7 @@ jobs:
           sudo apt-get update
           sudo apt-get -y -q install ffmpeg libavcodec-extra
           python -m pip install --upgrade pip setuptools wheel
-          pip install -q -r <(sed '/^pandas/d;/^scipy/d;/^matplotlib/d;/^xgboost/d;/^tensorflow/d;/^keras/d;/^jax/d;/^torch/d;/^Pillow/d;/^h5py/d' requirements_test.txt)
+          pip install -q -r <(sed '/^pandas/d;/^scipy/d;/^matplotlib/d;/^xgboost/d;/^tensorflow/d;/^keras/d;/^jax/d;/^torch/d;/^Pillow/d;/^h5py/d;/^scikit-learn/d' requirements_test.txt)
           pip install pandas==1.3.5
           pip install scipy==1.7.2
           pip install matplotlib==3.5.3
@@ -62,6 +62,7 @@ jobs:
           pip install torchvision==0.14.1+cpu
           pip install Pillow==9.5.0
           pip install h5py==3.8.0
+          pip install scikit-learn==1.0.2
           pip list
       - name: Run Tests
         run: ./run_tests.sh ${{ matrix.framework }}

.github/workflows/dockerhub.yml

@@ -31,11 +31,11 @@ jobs:
 
       - name: Extract metadata (tags, labels) for Docker
         id: meta
-        uses: docker/metadata-action@9dc751fe249ad99385a2583ee0d084c400eee04e
+        uses: docker/metadata-action@dbef88086f6cef02e264edb7dbf63250c17cef6c
         with:
           images: adversarialrobustnesstoolbox/releases
           tags: |
-            type=raw,value={{branch}}-1.17.0-{{sha}}
+            type=raw,value={{branch}}-1.17.1-{{sha}}
             type=semver,pattern={{version}}
 
       - name: Build and push Docker image

art/__init__.py

@@ -12,7 +12,7 @@
 from art import preprocessing
 
 # Semantic Version
-__version__ = "1.17.0"
+__version__ = "1.17.1"
 
 # pylint: disable=C0103
 

art/estimators/certification/object_seeker/object_seeker.py

@@ -52,7 +52,7 @@
 from sklearn.cluster import DBSCAN
 from tqdm.auto import tqdm
 
-from art.utils import intersection_over_area, non_maximum_suppression
+from art.utils import intersection_over_area
 
 logger = logging.getLogger(__name__)
 
@@ -94,68 +94,16 @@ def __init__(
         self.epsilon = epsilon
         self.verbose = verbose
 
-    @property
     @abc.abstractmethod
-    def channels_first(self) -> bool:
+    def _image_dimensions(self) -> Tuple[int, int]:
         """
-        :return: Boolean to indicate index of the color channels in the sample `x`.
-        """
-        pass
-
-    @property
-    @abc.abstractmethod
-    def input_shape(self) -> Tuple[int, ...]:
-        """
-        :return: Shape of one input sample.
-        """
-        pass
+        Get the height and width of a sample input image.
 
-    @abc.abstractmethod
-    def _predict_classifier(self, x: np.ndarray, batch_size: int = 128, **kwargs) -> List[Dict[str, np.ndarray]]:
-        """
-        Perform prediction for a batch of inputs.
-
-        :param x: Samples of shape NCHW or NHWC.
-        :param batch_size: Batch size.
-        :return: Predictions of format `List[Dict[str, np.ndarray]]`, one for each input image. The fields of the Dict
-                 are as follows:
-
-                 - boxes [N, 4]: the boxes in [x1, y1, x2, y2] format, with 0 <= x1 < x2 <= W and 0 <= y1 < y2 <= H.
-                 - labels [N]: the labels for each image
-                 - scores [N]: the scores or each prediction.
+        :return: Tuple containing the height and width of a sample input image.
         """
         raise NotImplementedError
 
-    def predict(self, x: np.ndarray, batch_size: int = 128, **kwargs) -> List[Dict[str, np.ndarray]]:
-        """
-        Perform prediction for a batch of inputs.
-
-        :param x: Samples of shape NCHW or NHWC.
-        :param batch_size: Batch size.
-        :return: Predictions of format `List[Dict[str, np.ndarray]]`, one for each input image. The fields of the Dict
-                 are as follows:
-
-                 - boxes [N, 4]: the boxes in [x1, y1, x2, y2] format, with 0 <= x1 < x2 <= W and 0 <= y1 < y2 <= H.
-                 - labels [N]: the labels for each image
-                 - scores [N]: the scores or each prediction.
-        """
-        predictions = []
-
-        for x_i in tqdm(x, desc="ObjectSeeker", disable=not self.verbose):
-            base_preds, masked_preds = self._masked_predictions(x_i, batch_size=batch_size, **kwargs)
-            pruned_preds = self._prune_boxes(masked_preds, base_preds)
-            unionized_preds = self._unionize_clusters(pruned_preds)
-
-            preds = {
-                "boxes": np.concatenate([base_preds["boxes"], unionized_preds["boxes"]]),
-                "labels": np.concatenate([base_preds["labels"], unionized_preds["labels"]]),
-                "scores": np.concatenate([base_preds["scores"], unionized_preds["scores"]]),
-            }
-
-            predictions.append(preds)
-
-        return predictions
-
+    @abc.abstractmethod
     def _masked_predictions(
         self, x_i: np.ndarray, batch_size: int = 128, **kwargs
     ) -> Tuple[Dict[str, np.ndarray], Dict[str, np.ndarray]]:
@@ -167,70 +115,7 @@ def _masked_predictions(
         :batch_size: Batch size.
         :return: Predictions for the base unmasked image and merged predictions for the masked image.
         """
-        x_mask = np.repeat(x_i[np.newaxis], self.num_lines * 4 + 1, axis=0)
-
-        if self.channels_first:
-            height = self.input_shape[1]
-            width = self.input_shape[2]
-        else:
-            height = self.input_shape[0]
-            width = self.input_shape[1]
-            x_mask = np.transpose(x_mask, (0, 3, 1, 2))
-
-        idx = 1
-
-        # Left masks
-        for k in range(1, self.num_lines + 1):
-            boundary = int(width / (self.num_lines + 1) * k)
-            x_mask[idx, :, :, :boundary] = 0
-            idx += 1
-
-        # Right masks
-        for k in range(1, self.num_lines + 1):
-            boundary = width - int(width / (self.num_lines + 1) * k)
-            x_mask[idx, :, :, boundary:] = 0
-            idx += 1
-
-        # Top masks
-        for k in range(1, self.num_lines + 1):
-            boundary = int(height / (self.num_lines + 1) * k)
-            x_mask[idx, :, :boundary, :] = 0
-            idx += 1
-
-        # Bottom masks
-        for k in range(1, self.num_lines + 1):
-            boundary = height - int(height / (self.num_lines + 1) * k)
-            x_mask[idx, :, boundary:, :] = 0
-            idx += 1
-
-        if not self.channels_first:
-            x_mask = np.transpose(x_mask, (0, 2, 3, 1))
-
-        predictions = self._predict_classifier(x=x_mask, batch_size=batch_size, **kwargs)
-        filtered_predictions = [
-            non_maximum_suppression(
-                pred, iou_threshold=self.iou_threshold, confidence_threshold=self.confidence_threshold
-            )
-            for pred in predictions
-        ]
-
-        # Extract base predictions
-        base_predictions = filtered_predictions[0]
-
-        # Extract and merge masked predictions
-        boxes = np.concatenate([pred["boxes"] for pred in filtered_predictions[1:]])
-        labels = np.concatenate([pred["labels"] for pred in filtered_predictions[1:]])
-        scores = np.concatenate([pred["scores"] for pred in filtered_predictions[1:]])
-        merged_predictions = {
-            "boxes": boxes,
-            "labels": labels,
-            "scores": scores,
-        }
-        masked_predictions = non_maximum_suppression(
-            merged_predictions, iou_threshold=self.iou_threshold, confidence_threshold=self.confidence_threshold
-        )
-
-        return base_predictions, masked_predictions
+        raise NotImplementedError
 
     def _prune_boxes(
         self, masked_preds: Dict[str, np.ndarray], base_preds: Dict[str, np.ndarray]
@@ -332,6 +217,36 @@ def _unionize_clusters(self, masked_preds: Dict[str, np.ndarray]) -> Dict[str, n
         }
         return unionized_predictions
 
+    def predict(self, x: np.ndarray, batch_size: int = 128, **kwargs) -> List[Dict[str, np.ndarray]]:
+        """
+        Perform prediction for a batch of inputs.
+
+        :param x: Samples of shape NCHW or NHWC.
+        :param batch_size: Batch size.
+        :return: Predictions of format `List[Dict[str, np.ndarray]]`, one for each input image. The fields of the Dict
+                 are as follows:
+
+                 - boxes [N, 4]: the boxes in [x1, y1, x2, y2] format, with 0 <= x1 < x2 <= W and 0 <= y1 < y2 <= H.
+                 - labels [N]: the labels for each image
+                 - scores [N]: the scores or each prediction.
+        """
+        predictions = []
+
+        for x_i in tqdm(x, desc="ObjectSeeker", disable=not self.verbose):
+            base_preds, masked_preds = self._masked_predictions(x_i, batch_size=batch_size, **kwargs)
+            pruned_preds = self._prune_boxes(masked_preds, base_preds)
+            unionized_preds = self._unionize_clusters(pruned_preds)
+
+            preds = {
+                "boxes": np.concatenate([base_preds["boxes"], unionized_preds["boxes"]]),
+                "labels": np.concatenate([base_preds["labels"], unionized_preds["labels"]]),
+                "scores": np.concatenate([base_preds["scores"], unionized_preds["scores"]]),
+            }
+
+            predictions.append(preds)
+
+        return predictions
+
     def certify(
         self,
         x: np.ndarray,
@@ -348,10 +263,7 @@ def certify(
         :return: A list containing an array of bools for each bounding box per image indicating if the bounding
                  box is certified against the given patch.
         """
-        if self.channels_first:
-            _, height, width = self.input_shape
-        else:
-            height, width, _ = self.input_shape
+        height, width = self._image_dimensions()
 
         patch_size = np.sqrt(height * width * patch_size)
         height_offset = offset * height