Merge pull request #62 from lscheinkman/DEVOPS-398

DEVOPS-398: Fix tensorflow tests

.circleci/config.yml

@@ -48,9 +48,12 @@ jobs:
       - restore_cache:
           key: pip-cache-v1-{{ arch }}-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
       - run: pip install --user flake8-formatter-junit-xml
-      - run: pip install --user -r requirements-dev.txt
       - run: pip install --user ray
-      - run: python setup.py install --user
+      - run: pip install --user git+https://github.com/numenta/nupic.tensorflow.git#egg=nupic.tensorflow
+      - run: pip install --user git+https://github.com/numenta/nupic.torch.git#egg=nupic.torch
+      - run: pip install --user -r requirements.txt
+      - run: pip install --user -r requirements-dev.txt
+      - run: python setup.py develop --user
       - save_cache:
           key: pip-cache-v1-{{ arch }}-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
           paths:

nupic/research/frameworks/tensorflow/utils/pytorch_utils.py

@@ -22,8 +22,11 @@
 """
 
 import numpy as np
+import tensorflow as tf
 import tensorflow.keras.backend as K
 
+TF_LOGGER = tf.get_logger()
+
 
 def _reflatten_linear_weight(x):
     """
@@ -129,5 +132,7 @@ def load_pytorch_weights(model_tf, model_pt, weights_map=None):
             if transform is not None:
                 value = transform(value)
             batch_values.append((var, value))
+        else:
+            TF_LOGGER.warn("Unknown variable: %s", var.name)
 
     K.batch_set_value(batch_values)

requirements.txt

@@ -19,6 +19,8 @@
 #  http://numenta.org/licenses/
 #
 # ------------------------------------------------------------------------------
+nupic.torch
+nupic.tensorflow
 numpy
 matplotlib
 pandas
@@ -27,9 +29,7 @@ librosa
 scikit-image
 requests
 ray
-nupic.tensorflow
 tqdm
 torchvision
-nupic.torch
 python-dotenv
 elasticsearch

setup.cfg

@@ -25,11 +25,13 @@ license_files = LICENSE
 [bdist_wheel]
 universal=0
 
-#[aliases]
-#test=pytest
+[aliases]
+test=pytest
 
 [tool:pytest]
 testpaths=tests
+filterwarnings =
+    ignore:.*:DeprecationWarning
 
 [flake8]
 ; Select all

tests/unit/frameworks/tensorflow/k_winner_test.py

@@ -22,6 +22,7 @@
 
 import numpy as np
 import tensorflow as tf
+from tensorflow.python.framework import test_util
 from tensorflow.python.keras.layers import Input
 from tensorflow.python.keras.models import Model
 
@@ -54,6 +55,7 @@ def setUp(self):
         )
         self.duty_cycle = tf.constant(1.0 / 3.0, shape=(2, 7))
 
+    @test_util.run_all_in_graph_and_eager_modes
     def test_inference(self):
         """boost factor 0, k=3, batch size 2"""
         expected = np.zeros(self.x.shape)
@@ -107,6 +109,7 @@ def setUp(self):
         x[1, 2, 1, 1] = 1.7
         self.x2 = tf.constant(x)
 
+    @test_util.run_all_in_graph_and_eager_modes
     def test_one(self):
         """Equal duty cycle, boost factor 0, k=4, batch size 1."""
         x = self.x
@@ -126,6 +129,7 @@ def test_one(self):
             y = k_winners(x, training=True)
             self.assertAllClose(y, expected)
 
+    @test_util.run_all_in_graph_and_eager_modes
     def test_two(self):
         """Equal duty cycle, boost factor 0, k=3."""
         x = self.x
@@ -144,6 +148,7 @@ def test_two(self):
             y = k_winners(x, training=True)
             self.assertAllClose(y, expected)
 
+    @test_util.run_all_in_graph_and_eager_modes
     def test_three(self):
         """Equal duty cycle, boost factor=0, k=4, batch size=2."""
         x = self.x2
@@ -167,6 +172,7 @@ def test_three(self):
             y = k_winners(x, training=True)
             self.assertAllClose(y, expected)
 
+    @test_util.run_all_in_graph_and_eager_modes
     def test_four(self):
         """Equal duty cycle, boost factor=0, k=3, batch size=2."""
         x = self.x2
@@ -188,6 +194,7 @@ def test_four(self):
             y = k_winners(x, training=True)
             self.assertAllClose(y, expected)
 
+    @test_util.run_all_in_graph_and_eager_modes
     def test_five(self):
         x = self.x2
         expected = np.zeros_like(x)

tests/unit/frameworks/tensorflow/load_pytorch_weights_test.py

@@ -21,82 +21,68 @@
 import numpy as np
 import tensorflow as tf
 import torch
-from tensorflow.python.framework import test_util
+from tensorflow.python.keras import keras_parameterized
+from tensorflow.python.platform import test
 
 from nupic.research.frameworks.tensorflow.utils import load_pytorch_weights
 from nupic.tensorflow.models import GSCSparseCNN, GSCSuperSparseCNN
 
-SEED = 18
 
-# Tensorflow configuration.
-# Make sure to use one thread in order to keep the results deterministic
-CONFIG = tf.ConfigProto(
-    intra_op_parallelism_threads=1,
-    inter_op_parallelism_threads=1,
-    device_count={
-        "CPU": 1},
-)
+@keras_parameterized.run_all_keras_modes
+class LoadPytorchWeightsTest(keras_parameterized.TestCase):
 
+    def setUp(self):
+        super(LoadPytorchWeightsTest, self).setUp()
+        self.data = np.random.rand(10, 1, 32, 32)
 
-class LoadPytorchWeightsTest(tf.test.TestCase):
-
-    @classmethod
-    def setUpClass(cls) -> None:
-        super().setUpClass()
-        cls.data = np.random.rand(10, 1, 32, 32)
-
-    @test_util.run_all_in_graph_and_eager_modes
     def test_gsc_sparse_cnn(self):
         """
         Test loading gsc_sparse_cnn from pytorch
         """
-        with self.session(config=CONFIG):
-            model_pt = torch.hub.load(github="numenta/nupic.torch",
-                                      model="gsc_sparse_cnn",
-                                      pretrained=True)
-            model_pt.eval()
-
-            model_tf = GSCSparseCNN(pre_trained=False)
-            load_pytorch_weights(model_tf, model_pt)
-            model_tf.compile(optimizer="sgd",
-                             loss="sparse_categorical_crossentropy",
-                             metrics=["accuracy"])
-            data_pt = torch.tensor(self.data, dtype=torch.float32)
-            data_tf = tf.convert_to_tensor(self.data.transpose(0, 2, 3, 1),
-                                           dtype=tf.float32)
-
-            out_pt = model_pt(data_pt)
-            out_tf = model_tf(data_tf, training=False)
-            out_tf = tf.log(out_tf)
-
-            self.assertAllClose(out_pt.detach().numpy(), out_tf)
+        model_pt = torch.hub.load(github="numenta/nupic.torch",
+                                  model="gsc_sparse_cnn",
+                                  pretrained=True)
+        model_pt.eval()
+
+        model_tf = GSCSparseCNN(pre_trained=False)
+        load_pytorch_weights(model_tf, model_pt)
+        model_tf.compile(optimizer="sgd",
+                         loss="sparse_categorical_crossentropy",
+                         metrics=["accuracy"])
+        data_pt = torch.tensor(self.data, dtype=torch.float32)
+        data_tf = tf.convert_to_tensor(self.data.transpose(0, 2, 3, 1),
+                                       dtype=tf.float32)
+
+        out_pt = model_pt(data_pt)
+        out_tf = model_tf(data_tf, training=False)
+        out_tf = tf.log(out_tf)
+
+        self.assertAllClose(out_pt.detach().numpy(), out_tf)
 
-    @test_util.run_all_in_graph_and_eager_modes
     def test_gsc_super_sparse_cnn(self):
         """
         Test loading gsc_sparse_cnn from pytorch
         """
-        with self.session(config=CONFIG):
-            model_pt = torch.hub.load(github="numenta/nupic.torch",
-                                      model="gsc_super_sparse_cnn",
-                                      pretrained=True)
-            model_pt.eval()
+        model_pt = torch.hub.load(github="numenta/nupic.torch",
+                                  model="gsc_super_sparse_cnn",
+                                  pretrained=True)
+        model_pt.eval()
 
-            model_tf = GSCSuperSparseCNN(pre_trained=False)
-            load_pytorch_weights(model_tf, model_pt)
-            model_tf.compile(optimizer="sgd",
-                             loss="sparse_categorical_crossentropy",
-                             metrics=["accuracy"])
-            data_pt = torch.tensor(self.data, dtype=torch.float32)
-            data_tf = tf.convert_to_tensor(self.data.transpose(0, 2, 3, 1),
-                                           dtype=tf.float32)
+        model_tf = GSCSuperSparseCNN(pre_trained=False)
+        load_pytorch_weights(model_tf, model_pt)
+        model_tf.compile(optimizer="sgd",
+                         loss="sparse_categorical_crossentropy",
+                         metrics=["accuracy"])
+        data_pt = torch.tensor(self.data, dtype=torch.float32)
+        data_tf = tf.convert_to_tensor(self.data.transpose(0, 2, 3, 1),
+                                       dtype=tf.float32)
 
-            out_pt = model_pt(data_pt)
-            out_tf = model_tf(data_tf, training=False)
-            out_tf = tf.log(out_tf)
+        out_pt = model_pt(data_pt)
+        out_tf = model_tf(data_tf, training=False)
+        out_tf = tf.log(out_tf)
 
-            self.assertAllClose(out_pt.detach().numpy(), out_tf)
+        self.assertAllClose(out_pt.detach().numpy(), out_tf)
 
 
 if __name__ == "__main__":
-    tf.test.main()
+    test.main()