Merge pull request #1125 from lingbai-kong/bug-IndexedSlices

fix: inconsistent shape error while training embedding layer

Author: Oceania2018

src/TensorFlowNET.Core/Framework/IndexedSlices.cs

@@ -49,12 +49,25 @@ public IndexedSlices(Tensor values, Tensor indices, Tensor dense_shape = null)
 
         public static implicit operator Tensor(IndexedSlices indexedSlices)
         {
-            return indexedSlices.values;
+            return _indexed_slices_to_tensor(indexedSlices);
         }
 
         public static implicit operator IndexedSlices(Tensor tensor)
         {
             return tensor.Tag as IndexedSlices;
         }
+
+        /// <summary>
+        /// Converts an IndexedSlices object `value` to a Tensor.
+        /// </summary>
+        /// <param name="indexedSlices"></param>
+        /// <param name="dtype"></param>
+        /// <param name="name"></param>
+        /// <param name="as_ref"></param>
+        /// <returns></returns>
+        public static Tensor _indexed_slices_to_tensor(IndexedSlices indexedSlices, TF_DataType dtype = TF_DataType.DtInvalid, String name = "", bool as_ref = false)
+        {
+            return gen_math_ops.unsorted_segment_sum(indexedSlices.values, indexedSlices.indices, indexedSlices.dense_shape.slice(0));
+        }
     }
 }

test/TensorFlowNET.Keras.UnitTest/Layers/LayersTest.cs

@@ -110,6 +110,17 @@ public void Embedding()
             var output_array = model.predict(input_array);
             Assert.AreEqual((32, 10, 64), output_array.shape);
         }
+        [TestMethod]
+        public void EmbeddingGrad()
+        {
+            var inputs = keras.layers.Input(shape: new[] { 32, 10 });
+            var outputs = keras.layers.Embedding(1000, 64, input_length: 10).Apply(inputs);
+            var model = keras.Model(inputs: inputs, outputs: outputs);
+            var input_array = np.random.randint(1000, size: (1, 32, 10));
+            var output_array = np.random.random(size: (1, 32, 10, 64));
+            model.compile("rmsprop", "mse", new[] { "accuracy" });
+            model.fit(input_array, output_array);
+        }
 
         /// <summary>
         /// https://www.tensorflow.org/api_docs/python/tf/keras/layers/Dense