Fix invalid-beta error when using BatchNormalization layer inside Tim…

…eDistribution

* Add test case for BatchNormalization as inner layer of TimeDistributed

* Fix rank of slices passed to inner layer of TimeDistribution

include/fdeep/layers/time_distributed_layer.hpp

@@ -67,6 +67,11 @@ class time_distributed_layer : public layer
         else
             raise_error("invalid input dim for TimeDistributed");
 
+        for (auto& slice: slices)
+        {
+            slice.shrink_rank();
+        }
+
         if (td_output_len_ == 2)
             concat_axis = 2;
         else if (td_output_len_ == 3)

include/fdeep/tensor.hpp

@@ -123,6 +123,10 @@ class tensor
     {
         return shape_;
     }
+    void shrink_rank()
+    {
+        shape_.shrink_rank();
+    }
     std::size_t depth() const
     {
         return shape().depth_;

include/fdeep/tensor_shape.hpp

@@ -122,6 +122,24 @@ class tensor_shape
         return rank_;
     }
 
+    std::size_t minimal_rank() const
+    {
+        if (size_dim_5_ > 1)
+            return 5;
+        if (size_dim_4_ > 1)
+            return 4;
+        if (height_ > 1)
+            return 3;
+        if (width_ > 1)
+            return 2;
+        return 1;
+    }
+
+    void shrink_rank()
+    {
+        rank_ = minimal_rank();
+    }
+
     std::vector<std::size_t> dimensions() const
     {
         if (rank() == 5)

keras_export/generate_test_models.py

@@ -466,6 +466,7 @@ def get_test_model_recurrent():
 
     outputs.append(TimeDistributed(MaxPooling2D(2, 2))(inputs[3]))
     outputs.append(TimeDistributed(AveragePooling2D(2, 2))(inputs[3]))
+    outputs.append(TimeDistributed(BatchNormalization())(inputs[3]))
 
     model = Model(inputs=inputs, outputs=outputs, name='test_model_recurrent')
     model.compile(loss='mse', optimizer='nadam')