Merge pull request #2329 from Trusted-AI/dependabot/pip/torch-2.1.1

Bump torch from 1.13.1 to 2.1.1

.github/actions/deepspeech-v3/Dockerfile

@@ -1,5 +1,5 @@
-# Get base from a pytorch image
-FROM pytorch/pytorch:1.6.0-cuda10.1-cudnn7-runtime
+pod# Get base from a pytorch image
+FROM pytorch/pytorch:2.1.1-cuda12.1-cudnn8-runtime
 
 # Set to install things in non-interactive mode
 ENV DEBIAN_FRONTEND noninteractive
@@ -17,26 +17,19 @@ RUN apt-get update \
         curl \
         libsndfile-dev \
         libsndfile1 \
+        vim \
+        curl \
      && apt-get clean all \
      && rm -r /var/lib/apt/lists/*
 
-RUN /opt/conda/bin/conda install --yes \
-    astropy \
-    matplotlib \
-    pandas \
-    scikit-learn \
-    scikit-image
-
 # Install necessary libraries for deepspeech v3
-RUN pip install torch
-RUN pip install tensorflow
-RUN pip install torchaudio==0.6.0
-RUN pip install --no-build-isolation fairscale
+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 && pip install -r requirements.txt
-RUN cd deepspeech.pytorch && pip install -e .
+RUN 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 pip install numba==0.50.0
-RUN pip install pytest-cov
-RUN pip install pydub==0.25.1
+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"

.github/workflows/ci-deepspeech-v3.yml

@@ -23,9 +23,9 @@ on:
 
 jobs:
   test_deepspeech_v3_torch_1_10:
-    name: PyTorchDeepSpeech v3 / PyTorch 1.10
+    name: PyTorchDeepSpeech v3 / PyTorch 2.1.1
     runs-on: ubuntu-latest
-    container: adversarialrobustnesstoolbox/art_testing_envs:deepspeech_v3_torch_1_10
+    container: adversarialrobustnesstoolbox/art_testing_envs:deepspeech_v3_torch_2_1_1
     steps:
       - name: Checkout Repo
         uses: actions/checkout@v3

.github/workflows/ci-pytorch.yml

@@ -28,24 +28,18 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - name: PyTorch 1.12.1 (Python 3.9)
-            framework: pytorch
-            python: 3.9
-            torch: 1.12.1+cpu
-            torchvision: 0.13.1+cpu
-            torchaudio: 0.12.1
-          - name: PyTorch 1.13.1 (Python 3.9)
-            framework: pytorch
-            python: 3.9
-            torch: 1.13.1+cpu
-            torchvision: 0.14.1+cpu
-            torchaudio: 0.13.1
           - name: PyTorch 1.13.1 (Python 3.10)
             framework: pytorch
             python: '3.10'
             torch: 1.13.1+cpu
             torchvision: 0.14.1+cpu
             torchaudio: 0.13.1
+          - name: PyTorch 2.1.1 (Python 3.10)
+            framework: pytorch
+            python: '3.10'
+            torch: 2.1.1
+            torchvision: 0.16.1+cpu
+            torchaudio: 2.1.1
 
     name: ${{ matrix.name }}
     steps:

.github/workflows/ci-style-checks.yml

@@ -31,15 +31,15 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v4
         with:
-          python-version: 3.8
+          python-version: '3.10'
       - name: Pre-install
         run: |
           sudo apt-get update
           sudo apt-get -y -q install ffmpeg libavcodec-extra
       - name: Install Dependencies
         run: |
           python -m pip install --upgrade pip setuptools wheel
-          pip install -q pylint==2.12.2 mypy==0.931 pycodestyle==2.8.0 black==21.12b0
+          pip install -q pylint==2.12.2 mypy==1.7.1 pycodestyle==2.8.0 black==21.12b0
           pip install -q -r <(sed '/^numpy/d;/^pluggy/d;/^tensorflow/d;/^keras/d' requirements_test.txt)
           pip install numpy==1.22.4
           pip install pluggy==0.13.1

.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' requirements_test.txt)
+          pip install -q -r <(sed '/^pandas/d;/^scipy/d;/^matplotlib/d;/^xgboost/d;/^tensorflow/d;/^keras/d;/^jax/d;/^torch/d' requirements_test.txt)
           pip install pandas==1.3.5
           pip install scipy==1.7.2
           pip install matplotlib==3.5.3
@@ -57,6 +57,9 @@ jobs:
           pip install tensorflow==${{ matrix.tensorflow }}
           pip install keras==${{ matrix.keras }}
           pip install numpy==1.20
+          pip install torch==1.13.1
+          pip install torchaudio==0.13.1
+          pip install torchvision==0.14.1+cpu
           pip list
       - name: Run Tests
         run: ./run_tests.sh ${{ matrix.framework }}

art/attacks/evasion/adversarial_patch/adversarial_patch_numpy.py

@@ -251,7 +251,11 @@ def generate(  # type: ignore
         return self.patch, self._get_circular_patch_mask()
 
     def apply_patch(
-        self, x: np.ndarray, scale: float, patch_external: np.ndarray = None, mask: Optional[np.ndarray] = None
+        self,
+        x: np.ndarray,
+        scale: float,
+        patch_external: Optional[np.ndarray] = None,
+        mask: Optional[np.ndarray] = None,
     ) -> np.ndarray:
         """
         A function to apply the learned adversarial patch to images or videos.

art/attacks/evasion/dpatch.py

@@ -264,7 +264,7 @@ def _augment_images_with_patch(
         random_location: bool,
         channels_first: bool,
         mask: Optional[np.ndarray] = None,
-        transforms: List[Dict[str, int]] = None,
+        transforms: Optional[List[Dict[str, int]]] = None,
     ) -> Tuple[np.ndarray, List[Dict[str, int]]]:
         """
         Augment images with patch.

art/attacks/evasion/imperceptible_asr/imperceptible_asr.py

@@ -540,14 +540,17 @@ def _approximate_power_spectral_density_torch(
 
         # compute short-time Fourier transform (STFT)
         # pylint: disable=W0212
-        stft_matrix = torch.stft(
-            perturbation,
-            n_fft=self._window_size,
-            hop_length=self._hop_size,
-            win_length=self._window_size,
-            center=False,
-            window=torch.hann_window(self._window_size).to(self.estimator._device),
-        ).to(self.estimator._device)
+        stft_matrix = torch.view_as_real(
+            torch.stft(
+                perturbation,
+                n_fft=self._window_size,
+                hop_length=self._hop_size,
+                win_length=self._window_size,
+                center=False,
+                window=torch.hann_window(self._window_size).to(self.estimator._device),
+                return_complex=True,
+            ).to(self.estimator._device)
+        )
 
         # compute power spectral density (PSD)
         # note: fixes implementation of Qin et al. by also considering the square root of gain_factor

art/attacks/evasion/imperceptible_asr/imperceptible_asr_pytorch.py

@@ -399,7 +399,10 @@ class only supports targeted attack.
                 loss.backward()
 
             # Get sign of the gradients
-            self.global_optimal_delta.grad = torch.sign(self.global_optimal_delta.grad)
+            if self.global_optimal_delta.grad is not None:
+                self.global_optimal_delta.grad = torch.sign(self.global_optimal_delta.grad)
+            else:
+                raise ValueError("Received None instead of gradient tensor.")
 
             # Do optimization
             self.optimizer_1.step()
@@ -747,14 +750,17 @@ def _psd_transform(self, delta: "torch.Tensor", original_max_psd: np.ndarray) ->
         window_fn = torch.hann_window  # type: ignore
 
         # Return STFT of delta
-        delta_stft = torch.stft(
-            delta,
-            n_fft=self.n_fft,
-            hop_length=self.hop_length,
-            win_length=self.win_length,
-            center=False,
-            window=window_fn(self.win_length).to(self.estimator.device),
-        ).to(self.estimator.device)
+        delta_stft = torch.view_as_real(
+            torch.stft(
+                delta,
+                n_fft=self.n_fft,
+                hop_length=self.hop_length,
+                win_length=self.win_length,
+                center=False,
+                window=window_fn(self.win_length).to(self.estimator.device),
+                return_complex=True,
+            ).to(self.estimator.device)
+        )
 
         # Take abs of complex STFT results
         transformed_delta = torch.sqrt(torch.sum(torch.square(delta_stft), -1))

art/attacks/evasion/over_the_air_flickering/over_the_air_flickering_pytorch.py

@@ -296,7 +296,10 @@ def _get_loss_gradients(self, x: "torch.Tensor", y: "torch.Tensor", perturbation
             # Compute gradients
             loss.backward()
             grads = eps.grad
-            grads_batch.append(grads[0, ...])
+            if grads is not None:
+                grads_batch.append(grads[0, ...])
+            else:
+                raise ValueError("Received None instead of gradient tensor.")
 
         grads_batch_tensor = torch.stack(grads_batch)
 

art/attacks/inference/membership_inference/shadow_models.py

@@ -164,8 +164,8 @@ def _hill_climbing_synthesis(
         max_iterations: int = 40,
         max_rejections: int = 3,
         min_features_randomized: int = 1,
-        random_record_fn: Callable[[], np.ndarray] = None,
-        randomize_features_fn: Callable[[np.ndarray, int], np.ndarray] = None,
+        random_record_fn: Optional[Callable[[], np.ndarray]] = None,
+        randomize_features_fn: Optional[Callable[[np.ndarray, int], np.ndarray]] = None,
     ) -> np.ndarray:
         """
         This method implements the hill climbing algorithm from R. Shokri et al. (2017)
@@ -247,8 +247,8 @@ def generate_synthetic_shadow_dataset(
         member_ratio: float = 0.5,
         min_confidence: float = 0.4,
         max_retries: int = 6,
-        random_record_fn: Callable[[], np.ndarray] = None,
-        randomize_features_fn: Callable[[np.ndarray, int], np.ndarray] = None,
+        random_record_fn: Optional[Callable[[], np.ndarray]] = None,
+        randomize_features_fn: Optional[Callable[[np.ndarray, int], np.ndarray]] = None,
     ) -> Tuple[Tuple[np.ndarray, np.ndarray, np.ndarray], Tuple[np.ndarray, np.ndarray, np.ndarray]]:
         """
         Generates a shadow dataset (member and nonmember samples and their corresponding model predictions) by training

art/attacks/poisoning/perturbations/audio_perturbations.py

@@ -21,8 +21,10 @@
 because loading the audio trigger from disk (librosa.load()) is very slow
 and should be done only once.
 """
-import numpy as np
+from typing import Optional
+
 import librosa
+import numpy as np
 
 
 class CacheTrigger:
@@ -89,7 +91,7 @@ def __init__(
         self,
         sampling_rate: int = 16000,
         backdoor_path: str = "../../../utils/data/backdoors/cough_trigger.wav",
-        duration: float = None,
+        duration: Optional[float] = None,
         **kwargs,
     ):
         """

art/attacks/poisoning/sleeper_agent_attack.py

@@ -431,7 +431,7 @@ def _select_poison_indices(
             classifier.model.trainable = model_trainable
         else:
             raise NotImplementedError("SleeperAgentAttack is currently implemented only for PyTorch and TensorFlowV2.")
-        indices = sorted(range(len(grad_norms)), key=lambda k: grad_norms[k])
+        indices = sorted(range(len(grad_norms)), key=lambda k: grad_norms[k])  # type: ignore
         indices = indices[-num_poison:]
         return np.array(indices)  # this will get only indices for target class
 

art/defences/trainer/adversarial_trainer_awp_pytorch.py

@@ -89,7 +89,7 @@ def fit(
         validation_data: Optional[Tuple[np.ndarray, np.ndarray]] = None,
         batch_size: int = 128,
         nb_epochs: int = 20,
-        scheduler: "torch.optim.lr_scheduler._LRScheduler" = None,
+        scheduler: Optional["torch.optim.lr_scheduler._LRScheduler"] = None,
         **kwargs,
     ):  # pylint: disable=W0221
         """
@@ -198,7 +198,7 @@ def fit_generator(
         generator: DataGenerator,
         validation_data: Optional[Tuple[np.ndarray, np.ndarray]] = None,
         nb_epochs: int = 20,
-        scheduler: "torch.optim.lr_scheduler._LRScheduler" = None,
+        scheduler: Optional["torch.optim.lr_scheduler._LRScheduler"] = None,
         **kwargs,
     ):  # pylint: disable=W0221
         """