Fix travis tests (#34)

* Fix num_batches for py27

* Skip MNIST tests when no GPU is available

* Update travis

.travis.yml

@@ -38,19 +38,12 @@ install:
       pip install https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.8.0-cp35-cp35m-linux_x86_64.whl;
     fi
 
-  - time pip install -q -e "bird-or-bicycle"
-  - time pip install -q -e "unrestricted-advex[test]"
-
-  # workaround for version incompatibility between the scipy version in conda
-  # and the system-provided /usr/lib/x86_64-linux-gnu/libstdc++.so.6
-  # by installing a conda-provided libgcc and adding it to the load path
-  - conda install libgcc
-  - export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/travis/miniconda/envs/test-environment/lib
-
 # command to run tests
 script:
   # exit on first error
   - set -e
+  - time pip install -q -e "bird-or-bicycle"
+  - time pip install -q -e "unrestricted-advex[test]"
 
   # Accuracy tests are slow so we run them after finishing all the other tests
   - pytest -s -k "not two_class_mnist_accuracy"

unrestricted-advex/unrestricted_advex/eval_kit.py

@@ -96,7 +96,7 @@ def evaluate_two_class_unambiguous_model(
 
     if attack._stop_after_n_datapoints is not None:
       batch_size = len(data_iter[0][0])
-      n_batches = math.ceil(float(attack._stop_after_n_datapoints) / batch_size)
+      n_batches = int(math.ceil(float(attack._stop_after_n_datapoints) / batch_size))
       attack_data_iter = data_iter[:n_batches]
     else:
       attack_data_iter = data_iter

unrestricted-advex/unrestricted_advex/mnist_baselines/mnist_utils.py

@@ -62,7 +62,7 @@ def two_class_iter(images, labels, num_datapoints, batch_size,
   which = (labels == class1) | (labels == class2)
   images_2class = images[which].astype(np.float32)
   labels_2class = labels[which]
-  num_batches = math.ceil(num_datapoints / batch_size)
+  num_batches = int(math.ceil(num_datapoints / batch_size))
 
   idxs = range(int(num_batches))
   if cycle:

unrestricted-advex/unrestricted_advex/tests/test_two_class_mnist.py

@@ -1,5 +1,6 @@
+import pytest
+import tensorflow as tf
 from unrestricted_advex import eval_kit, attacks
-
 from unrestricted_advex.mnist_baselines import mnist_utils
 
 model_dir = '/tmp/two-class-mnist/test'
@@ -36,6 +37,8 @@ def train_overfit_classifier(num_batches, batch_size):
   return mnist_utils.np_two_class_mnist_model(model_dir)
 
 
+@pytest.mark.skipif(not tf.test.is_gpu_available(),
+                    reason="Running attacks on MNIST currently requires a GPU :( ")
 def test_two_class_mnist_accuracy():
   """ Train an mnist model on a subset of mnist and then evaluate it
   on small attacks *on the training set*.