Fix docs build

Signed-off-by: Beat Buesser <[email protected]>

art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_pytorch.py

@@ -29,7 +29,6 @@
 from typing import Optional, TYPE_CHECKING
 
 import numpy as np
-import torch
 
 from art.config import ART_NUMPY_DTYPE
 from art.attacks.evasion.projected_gradient_descent.projected_gradient_descent_numpy import (
@@ -121,6 +120,8 @@ def generate(self, x: np.ndarray, y: Optional[np.ndarray] = None, **kwargs) -> n
         :type mask: `np.ndarray`
         :return: An array holding the adversarial examples.
         """
+        import torch
+
         # Check whether random eps is enabled
         self._random_eps()
 
@@ -232,6 +233,8 @@ def _compute_perturbation(self, x: "torch.Tensor", y: "torch.Tensor", mask: "tor
                      perturbed.
         :return: Perturbations.
         """
+        import torch
+
         # Pick a small scalar to avoid division by 0
         tol = 10e-8
 
@@ -266,6 +269,8 @@ def _apply_perturbation(self, x: "torch.Tensor", perturbation: "torch.Tensor", e
         :param eps_step: Attack step size (input variation) at each iteration.
         :return: Adversarial examples.
         """
+        import torch
+
         x = x + eps_step * perturbation
 
         if self.estimator.clip_values is not None:
@@ -302,6 +307,8 @@ def _compute_torch(
                             original input.
         :return: Adversarial examples.
         """
+        import torch
+
         if random_init:
             n = x.shape[0]
             m = np.prod(x.shape[1:])
@@ -344,6 +351,8 @@ def _projection(self, values: "torch.Tensor", eps: float, norm_p: int) -> "torch
         :param norm_p: L_p norm to use for clipping supporting 1, 2 and `np.Inf`.
         :return: Values of `values` after projection.
         """
+        import torch
+
         # Pick a small scalar to avoid division by 0
         tol = 10e-8
         values_tmp = values.reshape(values.shape[0], -1)

art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_tensorflow_v2.py

@@ -29,7 +29,6 @@
 from typing import Optional, TYPE_CHECKING
 
 import numpy as np
-import tensorflow as tf
 
 from art.config import ART_NUMPY_DTYPE
 from art.attacks.evasion.projected_gradient_descent.projected_gradient_descent_numpy import (
@@ -120,6 +119,8 @@ def generate(self, x: np.ndarray, y: Optional[np.ndarray] = None, **kwargs) -> n
         :type mask: `np.ndarray`
         :return: An array holding the adversarial examples.
         """
+        import tensorflow as tf
+
         # Check whether random eps is enabled
         self._random_eps()
 
@@ -222,6 +223,8 @@ def _compute_perturbation(self, x: "tf.Tensor", y: "tf.Tensor", mask: "tf.Tensor
                      perturbed.
         :return: Perturbations.
         """
+        import tensorflow as tf
+
         # Pick a small scalar to avoid division by 0
         tol = 10e-8
 
@@ -256,6 +259,8 @@ def _apply_perturbation(self, x: "tf.Tensor", perturbation: "tf.Tensor", eps_ste
         :param eps_step: Attack step size (input variation) at each iteration.
         :return: Adversarial examples.
         """
+        import tensorflow as tf
+
         x = x + eps_step * perturbation
 
         if self.estimator.clip_values is not None:
@@ -292,6 +297,8 @@ def _compute_tf(
                             original input.
         :return: Adversarial examples.
         """
+        import tensorflow as tf
+
         if random_init:
             n = x.shape[0]
             m = np.prod(x.shape[1:])
@@ -334,6 +341,8 @@ def _projection(values: "tf.Tensor", eps: float, norm_p: int) -> "tf.Tensor":
         :param norm_p: L_p norm to use for clipping supporting 1, 2 and `np.Inf`.
         :return: Values of `values` after projection.
         """
+        import tensorflow as tf
+
         # Pick a small scalar to avoid division by 0
         tol = 10e-8
         values_tmp = tf.reshape(values, (values.shape[0], -1))

art/defences/trainer/adversarial_trainer_FBF_Pytorch.py

@@ -26,8 +26,6 @@
 import time
 
 import numpy as np
-import torch
-import torch.nn as nn
 
 from art.config import ART_NUMPY_DTYPE
 from art.defences.trainer.adversarial_trainer_FBF import AdversarialTrainerFBF
@@ -200,6 +198,7 @@ def _batch_process(self, x_batch, y_batch, lr):
         :type lr: `float`
         :return: `(float, float, float)`
         """
+        import torch
 
         n = x_batch.shape[0]
         m = np.prod(x_batch.shape[1:])
@@ -239,7 +238,7 @@ def _batch_process(self, x_batch, y_batch, lr):
             loss.backward()
 
         # clip the gradients
-        nn.utils.clip_grad_norm_(self._classifier._model.parameters(), 0.5)
+        torch.nn.utils.clip_grad_norm_(self._classifier._model.parameters(), 0.5)
         self._classifier._optimizer.step()
 
         train_loss = loss.item() * o_batch.size(0)

setup.py

@@ -28,6 +28,12 @@
     "sphinx_rtd_theme",
     "sphinx-autodoc-annotation",
     "sphinx-autodoc-typehints",
+    "matplotlib",
+    "numpy",
+    "scipy",
+    "six==1.13.0",
+    "scikit-learn==0.22.1",
+    "Pillow>=6.0.0",
 ]
 
 

tests/defences/test_adversarial_trainer_FBF.py

@@ -73,3 +73,7 @@ def test_adversarial_trainer_FBF_Pytorch_fit_and_predict(get_adv_trainer, fix_ge
 
     np.testing.assert_array_almost_equal(accuracy, 0.32, decimal=0.001)
     np.testing.assert_array_almost_equal(accuracy_new, 0.14, decimal=0.001)
+
+
+if __name__ == "__main__":
+    pytest.cmdline.main("-q -s {} --mlFramework=pytorch --durations=0".format(__file__).split(" "))