Use Session instead of InteractiveSession

Avoids weird side effects after session closes unexpectedly.

CHANGES.rst

@@ -33,6 +33,8 @@ Release History
 - Fix bug in uneven step_blocks rounding
 - Fix bug in Simulator.print_params
 - Fix bug related to merging of learning rule with different dimensionality
+- Use tf.Session instead of tf.InteractiveSession, to avoid strange side
+  effects if the simulator isn't closed properly
 
 
 0.3.0 (April 25, 2017)

nengo_dl/simulator.py

@@ -181,8 +181,7 @@ def reset(self, seed=None):
             log_device_placement=False,
         )
 
-        self.sess = tf.InteractiveSession(graph=self.tensor_graph.graph,
-                                          config=config)
+        self.sess = tf.Session(graph=self.tensor_graph.graph, config=config)
         self.closed = False
 
         # initialize variables
@@ -679,7 +678,8 @@ def save_params(self, path):
             raise SimulationError("Simulation has been closed, cannot save "
                                   "parameters")
 
-        path = tf.train.Saver().save(self.sess, path)
+        with self.tensor_graph.graph.as_default():
+            path = tf.train.Saver().save(self.sess, path)
         logger.info("Model parameters saved to %s", path)
 
     def load_params(self, path):
@@ -694,7 +694,8 @@ def load_params(self, path):
             raise SimulationError("Simulation has been closed, cannot load "
                                   "parameters")
 
-        tf.train.Saver().restore(self.sess, path)
+        with self.tensor_graph.graph.as_default():
+            tf.train.Saver().restore(self.sess, path)
 
     def print_params(self, msg=None):
         """Print current values of trainable network parameters.
@@ -861,9 +862,10 @@ def check_gradients(self, outputs=None, atol=1e-5, rtol=1e-3):
                 dx, dy = gradient_checker._compute_dx_and_dy(
                     inp, out, out_shape)
 
-                analytic = gradient_checker._compute_theoretical_jacobian(
-                    inp, inp_shape, np.zeros(inp_shape), dy, out_shape, dx,
-                    extra_feed_dict=feed)
+                with self.sess.as_default():
+                    analytic = gradient_checker._compute_theoretical_jacobian(
+                        inp, inp_shape, np.zeros(inp_shape), dy, out_shape, dx,
+                        extra_feed_dict=feed)
 
                 if np.any(np.isnan(analytic)) or np.any(np.isnan(numeric)):
                     raise SimulationError("NaNs detected in gradient")

nengo_dl/tensor_graph.py

@@ -464,9 +464,9 @@ def build_loss(self, objective, targets):
                 else:
                     raise NotImplementedError
 
-        # average loss across probes (note: this will also average across
-        # the output of `objective` if it doesn't return a scalar)
-        loss = tf.reduce_mean(loss)
+            # average loss across probes (note: this will also average across
+            # the output of `objective` if it doesn't return a scalar)
+            loss = tf.reduce_mean(loss)
 
         self.losses[(objective, targets)] = loss
 

nengo_dl/tests/test_tensorflow_patch.py

@@ -38,13 +38,13 @@ def test_dynamic_stitch():
 
 
 def test_state_grads():
-    v = tf.Variable([0., 0., 0.])
-    x = tf.ones((3,))
+    with tf.Session() as sess:
+        v = tf.Variable([0., 0., 0.])
+        x = tf.ones((3,))
 
-    y0 = tf.assign(v, x)
-    y1 = tf.assign_add(v, x)
+        y0 = tf.assign(v, x)
+        y1 = tf.assign_add(v, x)
 
-    with tf.Session() as sess:
         # TODO: the ._ref() is necessary due to something in tensorflow 1.0.0,
         # can remove if we upgrade requirements
         grad0 = tf.gradients(y0, [v._ref(), x])
@@ -57,11 +57,12 @@ def test_state_grads():
         assert np.allclose(grad_vals[1][0], 1)
         assert np.allclose(grad_vals[1][1], 1)
 
-    x = tf.ones((1,))
-    y0 = tf.scatter_update(v, [0], x)
-    y1 = tf.scatter_add(v, [0], x)
-
     with tf.Session() as sess:
+        v = tf.Variable([0., 0., 0.])
+        x = tf.ones((1,))
+        y0 = tf.scatter_update(v, [0], x)
+        y1 = tf.scatter_add(v, [0], x)
+
         grad0 = tf.gradients(y0, [v._ref(), x])
         grad1 = tf.gradients(y1, [v._ref(), x])
         grad_vals = sess.run((grad0, grad1))

nengo_dl/tests/test_utils.py

@@ -102,11 +102,11 @@ def test_print_and_flush(capsys):
 
 
 def test_print_op(capsys):
-    x = tf.constant(0)
-    y = utils.print_op(x, "hello")
-    z = y + 0
-
     with tf.Session() as sess:
+        x = tf.constant(0)
+        y = utils.print_op(x, "hello")
+        z = y + 0
+
         sess.run(z)
 
     out, _ = capsys.readouterr()