Fix loss computation bug in Model training/eval methods with eager execution enabled.

fchollet · fchollet · commit a722cdf7a62a · 2018-04-20T16:54:04.000-07:00
Fixes tensorflow#18642. PiperOrigin-RevId: 193423288
diff --git a/tensorflow/python/keras/_impl/keras/engine/training_eager.py b/tensorflow/python/keras/_impl/keras/engine/training_eager.py
@@ -150,7 +150,7 @@ def _model_loss(model, inputs, targets, sample_weights=None, training=False):
       weighted_masked_fn = training_utils.weighted_masked_objective(loss_fn)
       with backend.name_scope(model.output_names[i] + '_loss'):
         output_loss = weighted_masked_fn(
-            outs[i], targets[i], weights, mask=mask)
+            targets[i], outs[i], weights, mask=mask)
       loss_metrics.append(backend.mean(output_loss))
 
       loss_weight = model.loss_weights_list[i]
diff --git a/tensorflow/python/keras/_impl/keras/engine/training_eager_test.py b/tensorflow/python/keras/_impl/keras/engine/training_eager_test.py
@@ -21,6 +21,7 @@
 import numpy as np
 
 from tensorflow.python.framework import ops
+from tensorflow.python.framework import test_util as tf_test_util
 from tensorflow.python.keras._impl import keras
 from tensorflow.python.keras._impl.keras import testing_utils
 from tensorflow.python.platform import test
@@ -625,6 +626,30 @@ def test_class_weight_invalid_use_case(self):
       model.fit(x_np, [y_np, y_np], epochs=1, sample_weight={'1': bad_w_np})
 
 
+class CorrectnessTest(test.TestCase):
+
+  @tf_test_util.run_in_graph_and_eager_modes()
+  def test_loss_correctness(self):
+    # Test that training loss is the same in eager and graph
+    # (by comparing it to a reference value in a deterministic case)
+    model = keras.Sequential()
+    model.add(keras.layers.Dense(3,
+                                 activation='relu',
+                                 input_dim=4,
+                                 kernel_initializer='ones'))
+    model.add(keras.layers.Dense(2,
+                                 activation='softmax',
+                                 kernel_initializer='ones'))
+    model.compile(loss='sparse_categorical_crossentropy',
+                  optimizer=RMSPropOptimizer(learning_rate=0.001))
+    x = np.ones((100, 4))
+    np.random.seed(123)
+    y = np.random.randint(0, 1, size=(100, 1))
+    history = model.fit(x, y, epochs=1, batch_size=10)
+    self.assertEqual(
+        np.around(history.history['loss'][-1], decimals=4), 0.6173)
+
+
 if __name__ == '__main__':
   ops.enable_eager_execution()
   test.main()
