[MXNET-978] Higher order gradient for sigmoid

src/common/exec_utils.h

@@ -286,7 +286,6 @@ inline void LogMemoryPlan(const nnvm::Graph& g) {
   const auto &idx = g.indexed_graph();
   const auto& vshape = g.GetAttr<mxnet::ShapeVector>("shape");
   const auto& vtype = g.GetAttr<nnvm::DTypeVector>("dtype");
-  const auto& vstorage = g.GetAttr<nnvm::StorageVector>("storage_id");
   // find node range
   uint32_t node_start = 0, node_end = idx.num_nodes();
   if (g.attrs.count("node_range")) {
@@ -304,13 +303,13 @@ inline void LogMemoryPlan(const nnvm::Graph& g) {
         auto eid = idx.entry_id(e);
         size_t kilo_bytes = vshape[eid].Size() * mshadow::mshadow_sizeof(vtype[eid]) / 1024;
         LOG(INFO) << "\t\tinput " << eid << ": " << vshape[eid] << " ("
-                  << kilo_bytes << " KB) -> " << storage_str(vstorage[eid]);
+                  << kilo_bytes << " KB)";
       }
       for (uint32_t index = 0; index < inode.source->num_outputs(); ++index) {
         uint32_t eid = idx.entry_id(nid, index);
         size_t kilo_bytes = vshape[eid].Size() * mshadow::mshadow_sizeof(vtype[eid]) / 1024;
         LOG(INFO) << "\t\toutput " << eid << ": " << vshape[eid] << " ("
-                  << kilo_bytes << " KB) -> " << storage_str(vstorage[eid]);
+                  << kilo_bytes << " KB)";
       }
     }
   }

src/imperative/imperative.cc

@@ -501,6 +501,10 @@ std::vector<NDArray*> Imperative::Backward(
     }
   }
 
+  if (dmlc::GetEnv("MXNET_MEM_PLAN_VERBOSE_LOGGING", false)) {
+    common::LogMemoryPlan(graph);
+  }
+
   // Execution
 
   bool prev_recording = set_is_recording(create_graph);

src/operator/tensor/elemwise_unary_op_basic.cc

@@ -121,7 +121,30 @@ The storage type of ``sigmoid`` output is always dense
 .set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseOut{"_backward_sigmoid"});
 
 MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU(_backward_sigmoid,
-                                               unary_bwd<mshadow_op::sigmoid_grad>);
+                                               unary_bwd<mshadow_op::sigmoid_grad>)
+.set_attr<nnvm::FGradient>("FGradient",
+    [](const nnvm::NodePtr& n, const std::vector<nnvm::NodeEntry>& ograds) {
+      // n->inputs[0] : y_grad
+      // n->inputs[1] : f(x) = sigmoid(x)
+      // ograds[0] : head_grads
+      // f''(x) = f'(x) * (1 - 2*f(x))
+      auto ones = MakeNode("ones_like", n->attrs.name + "_grad_ones", {n->inputs[1]}, nullptr, &n);
+      const std::unordered_map<std::string, std::string> args = {{"scalar", "2.0"}};
+      auto two_y = MakeNode("_mul_scalar", n->attrs.name + "_mul_two", {n->inputs[1]}, &args, &n);
+      auto one_minus_two_y = MakeNode("elemwise_sub", n->attrs.name + "_grad_sub",
+                                    {nnvm::NodeEntry{ones}, nnvm::NodeEntry{two_y}}, nullptr, &n);
+      auto grad_grad_mid = MakeNode("elemwise_mul", n->attrs.name + "_grad_mul",
+                                    {n->inputs[0], nnvm::NodeEntry{one_minus_two_y}}, nullptr, &n);
+      // when building gradient graph, the backward node of n->inputs[1] will be
+      // added to the graph again, therefore f`(x) will be multiplied
+      std::vector<nnvm::NodeEntry> ret;
+      ret.emplace_back(ograds[0]);  // this output is not passed out if gradient w.r.t x only
+      ret.emplace_back(MakeNode("elemwise_mul", n->attrs.name + "backward_grad_grad_in",
+                                {ograds[0], nnvm::NodeEntry{grad_grad_mid}}, nullptr, &n));
+      return ret;
+    });
+
+
 
 DMLC_REGISTER_PARAMETER(HardSigmoidParam);
 MXNET_OPERATOR_REGISTER_UNARY(hard_sigmoid)

tests/python/unittest/test_higher_order_grad.py

@@ -106,6 +106,23 @@ def grad_grad_op(x):
         check_second_order_unary(array, log10, grad_grad_op)
 
 
+@with_seed()
+def test_sigmoid():
+    def sigmoid(x):
+        return nd.sigmoid(x)
+
+    def grad_op(x):
+        return sigmoid(x) * (1 - sigmoid(x))
+
+    def grad_grad_op(x):
+        return grad_op(x) * (1 - 2 * sigmoid(x))
+
+    for dim in range(1, 5):
+        shape = rand_shape_nd(dim)
+        array = random_arrays(shape)
+        check_second_order_unary(array, sigmoid, grad_grad_op)
+
+
 def check_second_order_unary(x, op, grad_grad_op):
     x = nd.array(x)
     grad_grad_x = grad_grad_op(x)