Fix bug in sparse_categorical_accuracy

keras/metrics.py

@@ -34,10 +34,13 @@ def categorical_accuracy(y_true, y_pred):
 
 
 def sparse_categorical_accuracy(y_true, y_pred):
-    # flatten y_true in case it's in shape (num_samples, 1) instead of (num_samples,)
-    return K.cast(K.equal(K.flatten(y_true),
-                          K.cast(K.argmax(y_pred, axis=-1), K.floatx())),
-                  K.floatx())
+    # reshape in case it's in shape (num_samples, 1) instead of (num_samples,)
+    if K.ndim(y_true) == K.ndim(y_pred):
+        y_true = K.squeeze(y_true, -1)
+    # convert dense predictions to labels
+    y_pred_labels = K.argmax(y_pred, axis=-1)
+    y_pred_labels = K.cast(y_pred_labels, K.floatx())
+    return K.cast(K.equal(y_true, y_pred_labels), K.floatx())
 
 
 def top_k_categorical_accuracy(y_true, y_pred, k=5):

tests/keras/metrics_test.py

@@ -45,14 +45,17 @@ def test_sparse_metrics():
 
 
 def test_sparse_categorical_accuracy_correctness():
-    y_a = K.variable(np.random.randint(0, 7, (6,)), dtype=K.floatx())
-    y_b = K.variable(np.random.random((6, 7)), dtype=K.floatx())
-    # use one_hot embedding to convert sparse labels to equivalent dense labels
-    y_a_dense_labels = K.cast(K.one_hot(K.cast(y_a, dtype='int32'), num_classes=7),
-                              dtype=K.floatx())
-    sparse_categorical_acc = metrics.sparse_categorical_accuracy(y_a, y_b)
-    categorical_acc = metrics.categorical_accuracy(y_a_dense_labels, y_b)
-    assert np.allclose(K.eval(sparse_categorical_acc), K.eval(categorical_acc))
+    input_shapes = [(6,), (6, 3)]
+    for shape in input_shapes:
+        y_a = K.variable(np.random.randint(0, 7, shape), dtype=K.floatx())
+        y_b_shape = shape + (7,)
+        y_b = K.variable(np.random.random(y_b_shape), dtype=K.floatx())
+        # use one_hot embedding to convert sparse labels to equivalent dense labels
+        y_a_dense_labels = K.cast(K.one_hot(K.cast(y_a, dtype='int32'), 7),
+                                  dtype=K.floatx())
+        sparse_categorical_acc = metrics.sparse_categorical_accuracy(y_a, y_b)
+        categorical_acc = metrics.categorical_accuracy(y_a_dense_labels, y_b)
+        assert np.allclose(K.eval(sparse_categorical_acc), K.eval(categorical_acc))
 
 
 def test_serialize():