[NNVM] Fix grads for sum and expand_like (#1455)

nnvm/python/nnvm/top/transform.py

@@ -15,6 +15,10 @@ reg.register_schedule("expand_dims", _fschedule_broadcast)
 @reg.register_compute("expand_like")
 def compute_expand_like(attrs, inputs, _):
     """Compute definition of expand_like"""
+    if len(inputs[0].shape) == len(inputs[1].shape):
+        # If the number of dimensions is not changed then it is just a broadcasting
+        return topi.broadcast_to(inputs[0], inputs[1].shape)
+
     exclude = attrs.get_bool("exclude")
     axis = attrs.get_int_tuple("axis")
     if exclude:

nnvm/src/top/tensor/reduce.cc

@@ -170,12 +170,18 @@ Example::
   "FGradient", [](const NodePtr& n,
                   const std::vector<NodeEntry>& ograds){
     const ReduceParam& param = nnvm::get<ReduceParam>(n->attrs.parsed);
-    std::ostringstream axis; axis << param.axis;
+    bool exclude = param.exclude;
+    TShape p_axis = param.axis;
+    if (!param.exclude && param.axis.ndim() == 0) {
+      exclude = true;
+      p_axis = TShape();
+    }
+    std::ostringstream axis; axis << p_axis;
     return std::vector<NodeEntry>{
       MakeNode("expand_like", n->attrs.name + "_grad",
                {ograds[0], n->inputs[0]},
                {{"axis", axis.str()},
-                {"exclude", std::to_string(param.exclude)}})
+                {"exclude", std::to_string(exclude)}})
   };
 });
 

nnvm/src/top/tensor/transform.cc

@@ -251,7 +251,8 @@ will return a new array with shape ``(2,1,1,1,1,1,3,4)``.
 
 NNVM_REGISTER_OP(expand_like)
   .describe(R"code(Expand an input array with the shape of second array.
-This operation can always be composed of unsqueezing and expanding dims.
+This operation can be thought of as a composition of expand_dims and broadcast_to.
+If the dimensions are already expanded then it just broadcasts.
 Examples::
   input = [ 12.  19.  27.]
   input.shape = (3,)
@@ -282,11 +283,23 @@ Examples::
     std::ostringstream axis;
     axis << param.axis;
 
+    if (param.axis.ndim() == 0 && !param.exclude) {
+      // Special case needed because sum interprets axis=[] differently
       return std::vector<NodeEntry>{
-      MakeNode("sum", n->attrs.name + "_grad",
+        ograds[0],
+        MakeNode("zeros_like", n->attrs.name + "_zero_grad", {n->inputs[1]})
+      };
+    }
+
+    auto sum_node =
+      MakeNode("sum", n->attrs.name + "_sum_grad",
                {ograds[0]},
                {{"axis", axis.str()},
-                {"exclude", std::to_string(param.exclude)}}),
+                {"exclude", std::to_string(param.exclude)}});
+
+    return std::vector<NodeEntry>{
+      MakeNode("reshape_like", n->attrs.name + "_grad",
+               {sum_node, n->inputs[0]}),
       MakeNode("zeros_like", n->attrs.name + "_zero_grad", {n->inputs[1]})
     };
   })

nnvm/tests/python/compiler/test_top_level4.py

@@ -378,6 +378,13 @@ def verify_expand_like(in_shape, out_shape, axis, exclude):
 
     def forward(x, y):
         odim = len(out_shape)
+
+        if len(x.shape) == len(y.shape):
+            return np.broadcast_to(x, y.shape)
+
+        if x.shape == (1,) and len(y.shape) == odim:
+            x = np.reshape(x, ())
+
         real_axis = [i if i >= 0 else i + odim for i in axis]
         real_axis = sorted(real_axis)
         if exclude:
@@ -391,11 +398,17 @@ def verify_expand_like(in_shape, out_shape, axis, exclude):
 
     def backward(head_grads, x, y):
         odim = len(out_shape)
+
+        keepdims = len(x.shape) == len(y.shape)
+
+        if x.shape == (1,) and len(y.shape) == odim:
+            x = np.reshape(x, ())
+
         real_axis = [i if i >= 0 else i + odim for i in axis]
         real_axis = sorted(real_axis)
         if exclude:
             real_axis = list(set(range(odim)) - set(real_axis))
-        return [np.sum(head_grads, axis=tuple(real_axis)),
+        return [np.sum(head_grads, axis=tuple(real_axis), keepdims=keepdims),
                 np.zeros_like(y)]
 
 
@@ -410,6 +423,11 @@ def test_expand_like():
     verify_expand_like((2,), (2, 3), [1], False)
     verify_expand_like((3, 4), (3, 5, 4), [1], False)
     verify_expand_like((5, 7), (5, 6, 7, 8), [0, 2], True)
+    verify_expand_like((2, 3), (2, 3), [], False)
+    verify_expand_like((1,), (2, 3), [0, 1], False)
+    verify_expand_like((1, 1), (2, 3), [0, 1], False)
+    verify_expand_like((2, 1), (2, 3), [1], False)
+    verify_expand_like((1, 3), (2, 3), [0], False)
 
 
 def verify_elemwise_sum(num_args):

topi/python/topi/transform.py

@@ -65,15 +65,15 @@ def expand_like(a, shape_like, axis):
     """
     odim = len(axis) + len(a.shape)
     if odim != len(shape_like.shape):
+        if len(a.shape) == 1 and len(axis) == len(shape_like.shape):
+            # A special case: `a` is a scalar represented as a 1-dim tensor
+            return tvm.compute(shape_like.shape, lambda *idxs: a(0))
         raise ValueError("shape inconsistent when expand_like ({}, {}, {})".format(
             len(axis), len(a.shape), len(shape_like.shape)))
 
     real_axis = topi.reduction._get_real_axis(len(shape_like.shape), axis)
     real_axis = sorted(real_axis)
 
-    if not real_axis:
-        return a
-
     def _compute(*idxs):
         indices = []
         axis_index = 0