Fix the keras.sparse_categorical_crossentropy. (#985)

Author: AsakusaRinne

src/TensorFlowNET.Keras/Losses/SparseCategoricalCrossentropy.cs

@@ -14,6 +14,13 @@ public override Tensor Apply(Tensor target, Tensor output, bool from_logits = fa
         {
             target = tf.cast(target, dtype: TF_DataType.TF_INT64);
 
+            if (!from_logits)
+            {
+                var epsilon = tf.constant(KerasApi.keras.backend.epsilon(), output.dtype);
+                output = tf.clip_by_value(output, epsilon, 1 - epsilon);
+                output = tf.log(output);
+            }
+
             // Try to adjust the shape so that rank of labels = rank of logits - 1.
             var output_shape = array_ops.shape_v2(output);
             var output_rank = output.shape.ndim;