[Infer] More robust inference, support backward inference (apache#54)

Author: tqchen

nnvm/src/pass/infer_shape_type.cc

@@ -11,15 +11,16 @@ namespace nnvm {
 namespace pass {
 namespace {
 
-template<typename AttrType, typename IsNone>
+template<typename AttrType, typename IsNone, typename FDefault>
 Graph InferAttr(Graph &&ret,
-                const AttrType default_val,
+                const AttrType empty_val,
                 const char* infer_name,
                 const char* input_name,
                 const char* attr_key_name,
                 const char* attr_name,
                 const char* unknown_name,
-                IsNone fis_none) {
+                IsNone fis_none,
+                FDefault fdefault) {
   using AttrVector = std::vector<AttrType>;
   const IndexedGraph& idx = ret.indexed_graph();
   static auto& finfer_shape =
@@ -31,7 +32,7 @@ Graph InferAttr(Graph &&ret,
   if (ret.attrs.count(attr_name) != 0) {
     rshape = ret.MoveCopyAttr<AttrVector>(attr_name);
   } else {
-    rshape.resize(idx.num_node_entries(), default_val);
+    rshape.resize(idx.num_node_entries(), empty_val);
   }
 
   if (ret.attrs.count(input_name) != 0) {
@@ -51,12 +52,12 @@ Graph InferAttr(Graph &&ret,
     // erase the provided arguments
     ret.attrs.erase(attr_key_name);
   }
-
   // Temp space for shape inference.
   std::vector<AttrType> ishape, oshape;
-  // number of completed nodes
-  size_t num_unknown = 0;
-  for (uint32_t nid = 0; nid < idx.num_nodes(); ++nid) {
+  size_t num_unknown;
+
+  // inference step function for nid
+  auto infer_step = [&](uint32_t nid) {
     const auto& inode = idx[nid];
     const uint32_t num_inputs = inode.inputs.size();
     const uint32_t num_outputs = inode.source->num_outputs();
@@ -72,27 +73,6 @@ Graph InferAttr(Graph &&ret,
           CHECK(is >> rshape[out_ent_id]) << "Invalid attribute";
         }
       }
-    } else if (finfer_shape.count(inode.source->op())) {
-      // Forward operator inference.
-      ishape.resize(num_inputs, default_val);
-      for (uint32_t i = 0; i < ishape.size(); ++i) {
-        ishape[i] = rshape[idx.entry_id(inode.inputs[i])];
-      }
-      oshape.resize(num_outputs, default_val);
-      for (uint32_t i = 0; i < oshape.size(); ++i) {
-        oshape[i] = rshape[idx.entry_id(nid, i)];
-      }
-      // Call inference function of the operator.
-      bool forward_known = finfer_shape[inode.source->op()](
-          inode.source->attrs, &ishape, &oshape);
-      num_unknown += !forward_known;
-      // Save to the result map.
-      for (uint32_t i = 0; i < num_inputs; ++i) {
-        rshape[idx.entry_id(inode.inputs[i])] = ishape[i];
-      }
-      for (uint32_t i = 0; i < num_outputs; ++i) {
-        rshape[idx.entry_id(nid, i)] = oshape[i];
-      }
     } else if (backward_map.count(inode.source->op())) {
       // Backward operator inference.
       CHECK_GE(inode.control_deps.size(), 1)
@@ -111,6 +91,47 @@ Graph InferAttr(Graph &&ret,
         if (fis_none(rshape[idx.entry_id(nid, i)])) known = false;
       }
       num_unknown += !known;
+    } else {
+      bool forward_known = true;
+      // Forward operator inference.
+      ishape.resize(num_inputs, empty_val);
+      for (uint32_t i = 0; i < ishape.size(); ++i) {
+        ishape[i] = rshape[idx.entry_id(inode.inputs[i])];
+        if (fis_none(ishape[i])) forward_known = false;
+      }
+      oshape.resize(num_outputs, empty_val);
+      for (uint32_t i = 0; i < oshape.size(); ++i) {
+        oshape[i] = rshape[idx.entry_id(nid, i)];
+        if (fis_none(oshape[i])) forward_known = false;
+      }
+      if (!forward_known) {
+        auto finfer = finfer_shape.get(inode.source->op(), fdefault);
+        CHECK(finfer != nullptr)
+          << "Attribute " << infer_name
+          << " is not registed by op " << inode.source->op()->name;
+        // Call inference function of the operator.
+        forward_known = finfer(inode.source->attrs, &ishape, &oshape);
+      }
+      num_unknown += !forward_known;
+      // Save to the result map.
+      for (uint32_t i = 0; i < num_inputs; ++i) {
+        rshape[idx.entry_id(inode.inputs[i])] = ishape[i];
+      }
+      for (uint32_t i = 0; i < num_outputs; ++i) {
+        rshape[idx.entry_id(nid, i)] = oshape[i];
+      }
+    }
+  };
+
+  num_unknown = 0;
+  for (uint32_t nid = 0; nid < idx.num_nodes(); ++nid) {
+    infer_step(nid);
+  }
+  if (num_unknown != 0) {
+    num_unknown = 0;
+    // backward inference
+    for (uint32_t i = idx.num_nodes(); i != 0; --i) {
+      infer_step(i - 1);
     }
   }
   // set the shapes
@@ -127,19 +148,48 @@ NNVM_REGISTER_PASS(InferShape)
         std::move(ret), TShape(),
         "FInferShape", "shape_inputs", "shape_attr_key",
         "shape", "shape_num_unknown_nodes",
-        [](const TShape& s) { return s.ndim() == 0; });
+        [](const TShape& s) { return s.ndim() == 0; },
+        nullptr);
   })
 .set_change_graph(false)
 .provide_graph_attr("shape");
 
+// inference fucntion for same type
+inline bool SameType(const NodeAttrs& attrs,
+                     std::vector<int> *iattr,
+                     std::vector<int> *oattr) {
+  int def_v = -1;
+  for (int v : *oattr) {
+    if (v != -1) {
+      def_v = v; break;
+    }
+  }
+  if (def_v == -1) {
+    for (int v : *iattr) {
+      if (v != -1) {
+        def_v = v; break;
+      }
+    }
+  }
+  if (def_v == -1) return false;
+  for (int& v : *oattr) {
+    v = def_v;
+  }
+  for (int& v : *iattr) {
+    v = def_v;
+  }
+  return true;
+}
+
 NNVM_REGISTER_PASS(InferType)
 .describe("Infer the dtype of each node entries.")
 .set_body([](Graph ret) {
     return InferAttr<int>(
-        std::move(ret), 0,
+        std::move(ret), -1,
         "FInferType", "dtype_inputs", "dtype_attr_key",
         "dtype", "dtype_num_unknown_nodes",
-        [](const int t) { return t == -1; });
+        [](const int t) { return t == -1; },
+        SameType);
   })
 .set_change_graph(false)
 .provide_graph_attr("dtype");