Fix up regularizer progress-assert utility.

test/regularizers_test.py

@@ -3,21 +3,27 @@
 import util
 
 
-class TestNetwork(util.Base):
+class Mixin:
+    def assert_progress(self, **kwargs):
+        start = best = None
+        for _, val in self.exp.itertrain(
+                [self.INPUTS, self.OUTPUTS],
+                algorithm='sgd',
+                patience=2,
+                min_improvement=0.01,
+                batch_size=self.NUM_EXAMPLES,
+                **kwargs):
+            if start is None:
+                start = best = val['loss']
+            if val['loss'] < best:
+                best = val['loss']
+        assert best < start   # should have made progress!
+
+
+class TestNetwork(Mixin, util.Base):
     def setUp(self):
         self.exp = theanets.Regressor([self.NUM_INPUTS, 10, self.NUM_OUTPUTS])
 
-    def assert_progress(self, **kwargs):
-        train0, valid0 = next(self.exp.itertrain([self.INPUTS, self.OUTPUTS]))
-        trainN, validN = self.exp.train(
-            [self.INPUTS, self.OUTPUTS],
-            algorithm='sgd',
-            patience=2,
-            min_improvement=0.01,
-            batch_size=self.NUM_EXAMPLES,
-            **kwargs)
-        assert trainN['loss'] < valid0['loss']   # should have made progress!
-
     def test_input_noise(self):
         self.assert_progress(input_noise=0.001)
 
@@ -49,21 +55,10 @@ def test_contractive(self):
         self.assert_progress(contractive=0.001)
 
 
-class TestRecurrent(util.RecurrentBase):
+class TestRecurrent(Mixin, util.RecurrentBase):
     def setUp(self):
         self.exp = theanets.recurrent.Regressor([
-            self.NUM_INPUTS, (10, 'rnn'), self.NUM_OUTPUTS])
+            self.NUM_INPUTS, (20, 'rnn'), self.NUM_OUTPUTS])
 
     def test_recurrent_l2(self):
         self.assert_progress(recurrent_l2=0.001)
-
-    def assert_progress(self, **kwargs):
-        train0, valid0 = next(self.exp.itertrain([self.INPUTS, self.OUTPUTS]))
-        trainN, validN = self.exp.train(
-            [self.INPUTS, self.OUTPUTS],
-            algorithm='sgd',
-            patience=2,
-            min_improvement=0.01,
-            batch_size=self.NUM_EXAMPLES,
-            **kwargs)
-        assert trainN['loss'] < valid0['loss']   # should have made progress!