Fix scalar batch handling in arithmetic ops (#1449)

Adjust test for checking if the Tensor should be considered a scalar - uniform shape of 1-dim scalar elements. Allow Scalar batch to be broadcasted in the operation properly - no offset in tiles for scalar-like data, same code-path as constants. Return shape is always a batch - fixed for Constant/Scalar inputs. Add C++ unit tests for Scalar Batch. Test if Pipeline behaves properly when switching between scalar and non-scalar inputs. Python test was extended to cover Scalar inputs, size of some of the tests was reduced. Signed-off-by: Krzysztof Lecki <[email protected]>
NVIDIA · Nov 8, 2019 · b33c5ac · b33c5ac
1 parent a5c275a
commit b33c5ac
Show file tree

Hide file tree

Showing 6 changed files with 400 additions and 128 deletions.
diff --git a/dali/operators/expressions/arithmetic.h b/dali/operators/expressions/arithmetic.h
@@ -145,7 +145,8 @@ inline std::vector<ExprImplTask> CreateExecutionTasks(const ExprNode &expr, Expr
   return result;
 }
 
-inline TensorListShape<> ShapePromotion(std::string op, span<const TensorListShape<> *> shapes) {
+inline TensorListShape<> ShapePromotion(std::string op, span<const TensorListShape<> *> shapes,
+                                        int batch_size) {
   const TensorListShape<> *out_shape = nullptr;
   for (int i = 0; i < shapes.size(); i++) {
     if (IsScalarLike(*shapes[i]))
@@ -159,14 +160,15 @@ inline TensorListShape<> ShapePromotion(std::string op, span<const TensorListSha
                               *out_shape, ", ", *shapes[i], ")."));
     }
   }
-  return out_shape ? *out_shape : TensorListShape<>{{1}};
+  return out_shape ? *out_shape : uniform_list_shape(batch_size, {1});
 }
 
 template <typename Backend>
 DLL_PUBLIC inline const TensorListShape<> &PropagateShapes(ExprNode &expr,
-                                                    const workspace_t<Backend> &ws) {
+                                                           const workspace_t<Backend> &ws,
+                                                           int batch_size) {
   if (expr.GetNodeType() == NodeType::Constant) {
-    expr.SetShape(TensorListShape<>{{1}});
+    expr.SetShape(uniform_list_shape(batch_size, {1}));
     return expr.GetShape();
   }
   if (expr.GetNodeType() == NodeType::Tensor) {
@@ -182,9 +184,9 @@ DLL_PUBLIC inline const TensorListShape<> &PropagateShapes(ExprNode &expr,
   SmallVector<const TensorListShape<> *, kMaxArity> shapes;
   shapes.resize(subexpression_count);
   for (int i = 0; i < subexpression_count; i++) {
-    shapes[i] = &PropagateShapes<Backend>(func[i], ws);
+    shapes[i] = &PropagateShapes<Backend>(func[i], ws, batch_size);
   }
-  func.SetShape(ShapePromotion(func.GetFuncName(), make_span(shapes)));
+  func.SetShape(ShapePromotion(func.GetFuncName(), make_span(shapes), batch_size));
   return func.GetShape();
 }
 
@@ -242,7 +244,7 @@ class ArithmeticGenericOp : public Operator<Backend> {
       types_layout_inferred_ = true;
     }
 
-    result_shape_ = PropagateShapes<Backend>(*expr_, ws);
+    result_shape_ = PropagateShapes<Backend>(*expr_, ws, batch_size_);
     AllocateIntermediateNodes();
     exec_order_ = CreateExecutionTasks<Backend>(*expr_, cache_, ws.has_stream() ? ws.stream() : 0);
 

diff --git a/dali/operators/expressions/arithmetic_meta.h b/dali/operators/expressions/arithmetic_meta.h
@@ -429,7 +429,7 @@ inline ArithmeticOp NameToOp(const std::string &op_name) {
 }
 
 inline bool IsScalarLike(const TensorListShape<> &shape) {
-  return shape.num_samples() == 1 && shape.num_elements() == 1;
+  return is_uniform(shape) && shape.sample_dim() == 1 && shape.tensor_shape_span(0)[0] == 1;
 }
 
 }  // namespace dali