Upgrade nnvm to use automatic gradient correspondence guessing

[tqchen]

@piiswrong

[piiswrong]

whats this for?

[tqchen]

This is for enforcing shape constraints in the gradient input. This is needed, in case some of the backward node is simply forward path, Say gradient of add returns an identity of ograd, and we need shape of the gradient input variable to start with

[piiswrong]

Remind me why did we need to register isbackward in op instead of marking it in node?

…

On Nov 25, 2016 9:36 AM, "Tianqi Chen" ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In src/executor/graph_executor.cc <#3973>: > @@ -162,8 +171,9 @@ nnvm::Graph GraphExecutor::InitFullGraph( } if (!need_grad) return g; for (size_t i = 0; i < g.outputs.size(); ++i) { - head_grad_entry_.emplace_back(NodeEntry{nnvm::Node::Create(), 0, 0}); - head_grad_map_[head_grad_entry_.back().node.get()] = i; + NodeEntry ngrad{nnvm::Node::Create(), 0, 0}; + head_grad_entry_.emplace_back(AttrHint(ngrad, g.outputs[i])); This is for enforcing shape constraints in the gradient input. This is needed, in case some of the backward node is simply forward path, Say gradient of add returns an identity of ograd, and we need shape of the gradient input variable to start with — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3973>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAiudJVuE9NV7LOXBsnp_38XvpqcuUo_ks5rBxyQgaJpZM4K8DC5> .

[tqchen]

There is no is_backward field in the node. We used to have it, but concluded that is too artificial. Normal operator can also occur in gradient path, in which case is_backward is not registered, and normal shape inference function is used