diff --git a/mhcflurry/testing_utils.py b/mhcflurry/testing_utils.py
index 157a11af9f62891be7e61d64f1fae93b5463e830..64e4965b25c6f169ac69b0bdeb5e4ff89cee2473 100644
--- a/mhcflurry/testing_utils.py
+++ b/mhcflurry/testing_utils.py
@@ -2,6 +2,14 @@
 Utilities used in MHCflurry unit tests.
 """
 from . import Class1NeuralNetwork
+from .common import set_keras_backend
+
+
+def startup():
+    """
+    Configure Keras backend for running unit tests.
+    """
+    set_keras_backend("tensorflow-cpu", num_threads=2)
 
 
 def cleanup():
diff --git a/test/test_calibrate_percentile_ranks_command.py b/test/test_calibrate_percentile_ranks_command.py
index 2dc44a0b018a74113b5314bff7ecc79486115fb3..a5656ac35e5480fb2771a514d5a89dfb19e1aad3 100644
--- a/test/test_calibrate_percentile_ranks_command.py
+++ b/test/test_calibrate_percentile_ranks_command.py
@@ -14,9 +14,9 @@ from mhcflurry.downloads import get_path
 
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
-
+setup = startup
 
 def run_and_check(n_jobs=0, delete=True, additional_args=[]):
     source_models_dir = get_path("models_class1_pan", "models.with_mass_spec")
diff --git a/test/test_changing_allele_representations.py b/test/test_changing_allele_representations.py
index 38b9e12c339fd8bb2c6d40624c84bab92b4df636..271cb3d976fa1f53eb4d818e8806044a54f192fa 100644
--- a/test/test_changing_allele_representations.py
+++ b/test/test_changing_allele_representations.py
@@ -8,8 +8,9 @@ from mhcflurry.downloads import get_path
 
 from numpy.testing import assert_equal
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 ALLELE_TO_SEQUENCE = pandas.read_csv(
     get_path(
diff --git a/test/test_class1_affinity_predictor.py b/test/test_class1_affinity_predictor.py
index 8f2b4aa057bf1e42abc99752330ac45d28dded72..aa70a25da2f657578cf9caea84505847a6158a8d 100644
--- a/test/test_class1_affinity_predictor.py
+++ b/test/test_class1_affinity_predictor.py
@@ -15,13 +15,14 @@ from nose.tools import eq_, assert_raises
 from numpy import testing
 
 from mhcflurry.downloads import get_path
-import mhcflurry.testing_utils
+from mhcflurry.testing_utils import cleanup, startup
 
 DOWNLOADED_PREDICTOR = Class1AffinityPredictor.load()
 
 
 def setup():
     global DOWNLOADED_PREDICTOR
+    startup()
     DOWNLOADED_PREDICTOR = Class1AffinityPredictor.load()
     logging.basicConfig(level=logging.DEBUG)
 
@@ -29,7 +30,7 @@ def setup():
 def teardown():
     global DOWNLOADED_PREDICTOR
     DOWNLOADED_PREDICTOR = None
-    mhcflurry.testing_utils.cleanup()
+    cleanup()
 
 
 # To hunt down a weird warning we were seeing in pandas.
diff --git a/test/test_class1_neural_network.py b/test/test_class1_neural_network.py
index 4620334a38e96b917f41ec9db515ac33384711ce..47be495cb94c0f4c9133df458bc31ca97ed861df 100644
--- a/test/test_class1_neural_network.py
+++ b/test/test_class1_neural_network.py
@@ -13,8 +13,9 @@ from mhcflurry.class1_neural_network import Class1NeuralNetwork
 from mhcflurry.downloads import get_path
 from mhcflurry.common import random_peptides
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 
 def test_class1_neural_network_a0205_training_accuracy():
diff --git a/test/test_class1_pan.py b/test/test_class1_pan.py
index 9ad87d69e1ced68490c29fca4a657ab2473da2a3..0528ef60da5b6f82d4381452030cbe51c2ce83ef 100644
--- a/test/test_class1_pan.py
+++ b/test/test_class1_pan.py
@@ -11,8 +11,9 @@ from mhcflurry import Class1AffinityPredictor,Class1NeuralNetwork
 from mhcflurry.allele_encoding import AlleleEncoding
 from mhcflurry.downloads import get_path
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 
 HYPERPARAMETERS = {
diff --git a/test/test_custom_loss.py b/test/test_custom_loss.py
index 2426f48db03242e4427b1bf7fd59a527551401aa..98ee4ab52495b6163f15189c06969455c78062a5 100644
--- a/test/test_custom_loss.py
+++ b/test/test_custom_loss.py
@@ -11,8 +11,9 @@ import keras.backend as K
 
 from mhcflurry.custom_loss import CUSTOM_LOSSES
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 
 def evaluate_loss(loss, y_true, y_pred):
diff --git a/test/test_download_models_class1.py b/test/test_download_models_class1.py
index ffbf8b3cf6ac172cef426a2cadd9d0efa9ecd101..29a4d47a3dc12ba553e26f6fc1af93a338f543ea 100644
--- a/test/test_download_models_class1.py
+++ b/test/test_download_models_class1.py
@@ -5,14 +5,14 @@ from numpy.testing import assert_equal
 
 from mhcflurry import Class1AffinityPredictor, Class1NeuralNetwork
 
-from mhcflurry.testing_utils import cleanup
-
+from mhcflurry.testing_utils import cleanup, startup
 
 DOWNLOADED_PREDICTOR = None
 
 
 def setup():
     global DOWNLOADED_PREDICTOR
+    startup()
     DOWNLOADED_PREDICTOR = Class1AffinityPredictor.load()
 
 
diff --git a/test/test_multi_output.py b/test/test_multi_output.py
index b4254b56d7b1e7282c3127b00d15cf6593b98795..452f7a767c09a653be5f37113d20b0f117f44fd7 100644
--- a/test/test_multi_output.py
+++ b/test/test_multi_output.py
@@ -12,8 +12,9 @@ logging.getLogger('tensorflow').disabled = True
 from mhcflurry.class1_neural_network import Class1NeuralNetwork
 from mhcflurry.common import random_peptides
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 
 def test_multi_output():
diff --git a/test/test_network_merging.py b/test/test_network_merging.py
index 56683a52dbd8bfe63ad60b06ddaa76111d5d186f..69eab37dea033739fdb2f112b35ff3e6655cae8f 100644
--- a/test/test_network_merging.py
+++ b/test/test_network_merging.py
@@ -5,8 +5,7 @@ from mhcflurry import Class1AffinityPredictor, Class1NeuralNetwork
 from mhcflurry.common import random_peptides
 from mhcflurry.downloads import get_path
 
-from mhcflurry.testing_utils import cleanup
-
+from mhcflurry.testing_utils import cleanup, startup
 logging.getLogger('tensorflow').disabled = True
 
 PAN_ALLELE_PREDICTOR = None
@@ -14,6 +13,7 @@ PAN_ALLELE_PREDICTOR = None
 
 def setup():
     global PAN_ALLELE_PREDICTOR
+    startup()
     PAN_ALLELE_PREDICTOR = Class1AffinityPredictor.load(
         get_path("models_class1_pan", "models.with_mass_spec"),
         max_models=4,
diff --git a/test/test_predict_command.py b/test/test_predict_command.py
index 507d7c4a6c62939636ccb7315707957850a99c0a..c3f0a5c1aa22a634b06c862056c9216452849a35 100644
--- a/test/test_predict_command.py
+++ b/test/test_predict_command.py
@@ -6,8 +6,9 @@ from numpy.testing import assert_equal
 
 from mhcflurry import predict_command
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 TEST_CSV = '''
 Allele,Peptide,Experiment
diff --git a/test/test_released_predictors_on_hpv_dataset.py b/test/test_released_predictors_on_hpv_dataset.py
index 3369257455157cc4047aedde6818139dee54b52f..26bf0227a74160187a3e385d1f633573b0400926 100644
--- a/test/test_released_predictors_on_hpv_dataset.py
+++ b/test/test_released_predictors_on_hpv_dataset.py
@@ -13,8 +13,7 @@ from nose.tools import eq_, assert_less, assert_greater, assert_almost_equal
 from mhcflurry import Class1AffinityPredictor
 from mhcflurry.downloads import get_path
 
-from mhcflurry.testing_utils import cleanup
-
+from mhcflurry.testing_utils import cleanup, startup
 
 def data_path(name):
     '''
@@ -30,6 +29,7 @@ PREDICTORS = None
 
 def setup():
     global PREDICTORS
+    startup()
     PREDICTORS = {
         'allele-specific': Class1AffinityPredictor.load(
             get_path("models_class1", "models")),
diff --git a/test/test_released_predictors_well_correlated.py b/test/test_released_predictors_well_correlated.py
index 82071d653e3db2c6147e73d97e8d1f146b785d84..5040843ffa76338b5dcc0f242be7a02784d1c5b2 100644
--- a/test/test_released_predictors_well_correlated.py
+++ b/test/test_released_predictors_well_correlated.py
@@ -15,13 +15,14 @@ from mhcflurry.encodable_sequences import EncodableSequences
 from mhcflurry.downloads import get_path
 from mhcflurry.common import random_peptides
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 
 PREDICTORS = None
 
 
 def setup():
     global PREDICTORS
+    startup()
     PREDICTORS = {
         'allele-specific': Class1AffinityPredictor.load(
             get_path("models_class1", "models")),
@@ -38,7 +39,7 @@ def teardown():
 
 def test_correlation(
         alleles=None,
-        num_peptides_per_length=500,
+        num_peptides_per_length=100,
         lengths=[8, 9, 10],
         debug=False):
     peptides = []
diff --git a/test/test_speed.py b/test/test_speed.py
index eb604a5087458df6c517b7dde5db189d7ebe11e0..037ae61c637a34f70d1e33eb5f382587107537c0 100644
--- a/test/test_speed.py
+++ b/test/test_speed.py
@@ -18,46 +18,41 @@ from mhcflurry.encodable_sequences import EncodableSequences
 from mhcflurry.common import random_peptides
 from mhcflurry.downloads import get_path
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 
-ALLELE_SPECIFIC_PREDICTOR = Class1AffinityPredictor.load(
-    get_path("models_class1", "models"))
 
-PAN_ALLELE_PREDICTOR = Class1AffinityPredictor.load(
-    get_path("models_class1_pan", "models.with_mass_spec"))
-
-
-PREDICTORS = None
+ALLELE_SPECIFIC_PREDICTOR = None
+PAN_ALLELE_PREDICTOR = None
 
 
 def setup():
-    global PREDICTORS
-    PREDICTORS = {
-        'allele-specific': Class1AffinityPredictor.load(
-            get_path("models_class1", "models")),
-        'pan-allele': Class1AffinityPredictor.load(
-            get_path("models_class1_pan", "models.with_mass_spec"))
-    }
+    global ALLELE_SPECIFIC_PREDICTOR, PAN_ALLELE_PREDICTOR
+    startup()
+    ALLELE_SPECIFIC_PREDICTOR = Class1AffinityPredictor.load(
+        get_path("models_class1", "models"))
+
+    PAN_ALLELE_PREDICTOR = Class1AffinityPredictor.load(
+        get_path("models_class1_pan", "models.with_mass_spec"))
 
 
 def teardown():
-    global PREDICTORS
-    PREDICTORS = None
+    global ALLELE_SPECIFIC_PREDICTOR, PAN_ALLELE_PREDICTOR
+    ALLELE_SPECIFIC_PREDICTOR = None
+    PAN_ALLELE_PREDICTOR = None
     cleanup()
 
 
 DEFAULT_NUM_PREDICTIONS = 10000
 
 
-def test_speed_allele_specific(
-        profile=False,
-        predictor=ALLELE_SPECIFIC_PREDICTOR,
-        num=DEFAULT_NUM_PREDICTIONS):
-
+def test_speed_allele_specific(profile=False, num=DEFAULT_NUM_PREDICTIONS):
+    global ALLELE_SPECIFIC_PREDICTOR
     starts = collections.OrderedDict()
     timings = collections.OrderedDict()
     profilers = collections.OrderedDict()
 
+    predictor = ALLELE_SPECIFIC_PREDICTOR
+
     def start(name):
         starts[name] = time.time()
         if profile:
@@ -101,15 +96,14 @@ def test_speed_allele_specific(
         (key, pstats.Stats(value)) for (key, value) in profilers.items())
 
 
-def test_speed_pan_allele(
-        profile=False,
-        predictor=PAN_ALLELE_PREDICTOR,
-        num=DEFAULT_NUM_PREDICTIONS):
-
+def test_speed_pan_allele(profile=False, num=DEFAULT_NUM_PREDICTIONS):
+    global PAN_ALLELE_PREDICTOR
     starts = collections.OrderedDict()
     timings = collections.OrderedDict()
     profilers = collections.OrderedDict()
 
+    predictor = PAN_ALLELE_PREDICTOR
+
     def start(name):
         starts[name] = time.time()
         if profile:
diff --git a/test/test_train_and_related_commands.py b/test/test_train_and_related_commands.py
index 7b893f7be0914669115e0477284aa70fc300a88d..92efbdbda6bf56fc580adf8c70b037c1d5a04086 100644
--- a/test/test_train_and_related_commands.py
+++ b/test/test_train_and_related_commands.py
@@ -14,8 +14,9 @@ from numpy.testing import assert_array_less, assert_equal
 from mhcflurry import Class1AffinityPredictor
 from mhcflurry.downloads import get_path
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
diff --git a/test/test_train_pan_allele_models_command.py b/test/test_train_pan_allele_models_command.py
index 51c7c3ffb5bc16ba4e92504702e06fe8e4ee8d58..53c5bb254de227e455aa598fd35182e5e19901eb 100644
--- a/test/test_train_pan_allele_models_command.py
+++ b/test/test_train_pan_allele_models_command.py
@@ -15,8 +15,9 @@ from numpy.testing import assert_equal, assert_array_less
 from mhcflurry import Class1AffinityPredictor,Class1NeuralNetwork
 from mhcflurry.downloads import get_path
 
-from mhcflurry.testing_utils import cleanup
+from mhcflurry.testing_utils import cleanup, startup
 teardown = cleanup
+setup = startup
 
 os.environ["CUDA_VISIBLE_DEVICES"] = ""