reminisce commented on a change in pull request #17530: Add deferred compute
support
URL: https://github.com/apache/incubator-mxnet/pull/17530#discussion_r378670188
##########
File path: src/imperative/imperative.cc
##########
@@ -273,10 +288,132 @@ void Imperative::RecordOp(
info.outputs.back().dtype_ = outputs[i]->dtype();
info.outputs.back().storage_type_ = outputs[i]->storage_type();
}
- outputs[i]->entry_ = nnvm::NodeEntry{node, i, 0};
+ outputs[i]->autograd_entry_ = nnvm::NodeEntry{node, i, 0};
+ }
+}
+
+void Imperative::RecordDeferredCompute(nnvm::NodeAttrs &&attrs,
+ const std::vector<NDArray *> &inputs,
+ const std::vector<NDArray *> &outputs) {
+ CHECK(!is_recording())
+ << "Autograd recording is not supported during deferred compute mode.";
+
+ for (const NDArray *output : outputs) {
+ CHECK(DCInfo::IsNone(*output))
+ << "Inplace operations (+=, -=, x[:]=, etc) are not supported when "
+ << "recording in deferred compute mode.";
+ // However, an inplace operation on a non-deferred compute array inside
+ // deferred compute scope will work. For example:
+ // a = mx.nd.arange(10)
+ // with dc.context():
+ // a[:5] = 0
+ }
+ DispatchMode dispatch_mode = DispatchMode::kUndefined;
+ Context default_ctx = Context::CPU();
+ Context ctx = imperative::GetContext(attrs, inputs, outputs, default_ctx);
+ imperative::SetShapeType(ctx, attrs, inputs, outputs, &dispatch_mode);
+
+ nnvm::ObjectPtr node = nnvm::Node::Create();
+ node->inputs.reserve(inputs.size());
+ // Get NodeEntries for inputs
+ for (const NDArray *array : inputs) {
+ // For non-deferred compute arrays, array->deferredcompute_entry_ will be
+ // nullptr. We handle this in in GetDeferredComputeSymbol
+ node->inputs.emplace_back(array->deferredcompute_entry_);
+ }
+ node->attrs = std::move(attrs);
+ // Need to support NameManager in imperative API to better name
node->attrs.name
+ node->attrs.name = "node_" + std::to_string(node_count_++);
+ DCInfo::Create(node, inputs, outputs);
+
+ for (uint32_t i = 0; i < outputs.size(); ++i) {
+ outputs[i]->deferredcompute_entry_ = nnvm::NodeEntry{node, i, 0};
}
}
+nnvm::Symbol *Imperative::GetDeferredComputeSymbol(
+ const std::vector<std::pair<NDArray *, std::string>> &inputs,
+ const std::vector<NDArray *> &outputs
+ ) {
+ Symbol s;
+ s.outputs.reserve(outputs.size());
+ for (NDArray * ndoutput : outputs) {
+ CHECK(!Imperative::DCInfo::IsNone(*ndoutput))
+ << "ValueError: output_arrays for GetDeferredComputeSymbol "
+ << "must have a deferred compute history associated with them.";
+ s.outputs.emplace_back(ndoutput->deferredcompute_entry_);
+ }
+ std::unordered_map<NDArray *, nnvm::ObjectPtr> ndinput_to_variable;
+ std::unordered_set<const NDArray *> missing_inputs;
+ auto add_symbol_variables = [&inputs, &ndinput_to_variable,
+ &missing_inputs](const nnvm::ObjectPtr &node) {
+ if (node == nullptr) {
+ // This (nonexistant) "Node" belongs to an array created outside of
deferred compute scope.
+ return;
+ }
+
+ // Check if node has any non-deferred compute inputs
+ for (uint32_t i = 0; i < node->inputs.size(); i++) {
+ nnvm::NodeEntry &node_entry = node->inputs[i];
+ if (node_entry.node == nullptr || node_entry.node->is_variable()) {
+ // Node has non-deferred compute input (nullptr). Find the
corresponding
+ // NDArray and create a variable for it. If GetDeferredComputeSymbol
has
+ // been called before, a variable already exists and only the name
needs
+ // to be updated.
+ Imperative::DCInfo &dcinfo = Imperative::DCInfo::Get(node);
+ const NDArray *array = dcinfo.input_handles_.at(i);
+
+ // Make sure this array is part of GetDeferredComputeSymbol inputs
+ auto is_equal = [array](const std::pair<NDArray *, std::string>
&input) {
+ return array == std::get<0>(input);
+ };
+
+ // std::vector<std::pair<NDArray *, std::string>>::iterator
input_search =
+ auto input_search = std::find_if(inputs.begin(), inputs.end(),
is_equal);
+ // Create symbol variable
+ if (input_search != inputs.end()) {
+ NDArray *ndinput;
+ std::string input_name;
+ std::tie(ndinput, input_name) = *input_search;
+
+ nnvm::ObjectPtr input_variable;
+
+ auto variable_search = ndinput_to_variable.find(ndinput);
+ if (variable_search == ndinput_to_variable.end()) {
+ // No variable for this ndarray yet
+ input_variable = nnvm::CreateVariableNode(input_name);
+ ndinput_to_variable.insert({ndinput, input_variable});
+ } else {
+ input_variable = variable_search->second;
+ }
+
+ node_entry.node = input_variable;
+ } else {
+ missing_inputs.insert(array);
+ }
+ }
+ }
+ };
+ nnvm::DFSVisit(s.outputs, add_symbol_variables);
+
+ CHECK_EQ(ndinput_to_variable.size(), inputs.size())
+ << "ValueError: Invalid input to GetDeferredComputeSymbol. "
+ << ndinput_to_variable.size() << " inputs are required, but "
+ << inputs.size() << " were specified.";
+
+ CHECK_EQ(missing_inputs.size(), 0)
+ << "ValueError: Invalid input to GetDeferredComputeSymbol. "
+ << missing_inputs.size() << " required inputs unspecified.";
+
+ // Deep copy of symbol as subsequent calls to this function may change the
+ // name of input variables.
+ s = s.Copy();
+ Symbol *out = new Symbol();
Review comment:
Related to the above comment, it's better to move this to the caller and
simply do `return s.Copy()` here to prevent potential memory leak when an
exception is thrown in the following line.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
