Compile

pchavanne · Feb 23, 2017 · 9ad0a08 · 9ad0a08
1 parent e5a5b45
commit 9ad0a08
Show file tree

Hide file tree

Showing 2 changed files with 22 additions and 2 deletions.
diff --git a/tests/test_model.py b/tests/test_model.py
@@ -94,6 +94,8 @@ def test_no_data_found(self, model_no_data, network_unsupervised):
 
     def test_no_network(self, model):
         from yadll.exceptions import NoNetworkFoundException
+        with pytest.raises(NoNetworkFoundException):
+            model.compile(compile_arg='train')
         with pytest.raises(NoNetworkFoundException):
             model.pretrain()
         with pytest.raises(NoNetworkFoundException):
@@ -130,6 +132,8 @@ def test_model(self, model, network, network_unsupervised):
         model.network = network_unsupervised
         model.pretrain()
         model.train()
+        model.hp.patience = 10  # test early stop
+        model.train()
 
     def test_predict(self, data, model, network):
         model.network = network

diff --git a/tests/test_objectives.py b/tests/test_objectives.py
@@ -40,17 +40,27 @@ def test_mean_absolute_error():
 
 def test_binary_hinge_error():
     x_val = np.asarray(np.random.uniform(size=(10, 1)), dtype=yadll.utils.floatX)
-    y_val = np.asarray(np.random.binomial(n=1,p=0.5, size=(10, 1)), dtype=yadll.utils.floatX)
+    y_val = np.asarray(np.random.binomial(n=1, p=0.5, size=(10, 1)), dtype=yadll.utils.floatX)
     y_val = 2 * y_val - 1
     f = theano.function([x, y], yadll.objectives.binary_hinge_error(x, y))
     actual = f(x_val, y_val)
     desired = np.maximum(1. - x_val * y_val, 0.).flatten()
     assert_allclose(actual, desired, rtol=eps)
 
 
+def test_categorical_hinge_error():
+    x_val = np.asarray(np.random.uniform(size=(10, 1)), dtype=yadll.utils.floatX)
+    y_val = np.asarray(np.random.binomial(n=1, p=0.5, size=(10, 1)), dtype=yadll.utils.floatX)
+    y_val = 2 * y_val - 1
+    f = theano.function([x, y], yadll.objectives.categorical_hinge_error(x, y))
+    actual = f(x_val, y_val)
+    desired = np.maximum(1. - x_val * y_val, 0.).flatten()
+    assert_allclose(actual, desired, rtol=eps)
+
+
 def test_binary_crossentropy_error():
     x_val = np.asarray(np.random.uniform(size=(10, 1)), dtype=yadll.utils.floatX)
-    y_val = np.asarray(np.random.binomial(n=1,p=0.5, size=(10, 1)), dtype=yadll.utils.floatX)
+    y_val = np.asarray(np.random.binomial(n=1, p=0.5, size=(10, 1)), dtype=yadll.utils.floatX)
     f = theano.function([x, y], yadll.objectives.binary_crossentropy_error(x, y))
     actual = f(x_val, y_val)
     desired = np.mean(-(y_val * np.log(x_val) + (1 - y_val) * np.log(1 - x_val)), axis=-1)
@@ -63,3 +73,9 @@ def test_categorical_crossentropy_error():
     desired = np.mean(-np.sum(y_val * np.log(x_val), axis=-1))
     assert_allclose(actual, desired, rtol=eps)
 
+
+def test_kullback_leibler_divergence():
+    f = theano.function([x, y], yadll.objectives.kullback_leibler_divergence(x, y))
+    actual = f(x_val, y_val)
+    desired = np.sum(y_val * np.log(y_val/x_val), axis=-1)
+    assert_allclose(actual, desired, rtol=eps)