torch/csrc/jit/passes/onnx.cpp

#include "torch/csrc/utils/pybind.h"
#include "torch/csrc/jit/passes/onnx.h"
#include "torch/csrc/jit/passes/dead_code_elimination.h"
#include "torch/csrc/autograd/function.h"
#include "torch/csrc/autograd/symbolic.h"
#include "torch/csrc/jit/assertions.h"
#include "torch/csrc/utils/functional.h"
#include <unordered_map>
#include <sstream>

namespace torch { namespace jit {

// Transform PythonOps into Nodes that match ONNX semantics.
std::shared_ptr<Graph> ToONNX(std::shared_ptr<Graph>& graph, ::torch::onnx::OperatorExportTypes operator_export_type) {
  auto new_graph = std::make_shared<Graph>(graph->current_scope());
  std::unordered_map<Value*, Value*> env;
  BlockToONNX(graph->block(), new_graph->block(), operator_export_type, env);
  return new_graph;
}

void BlockToONNX(Block* old_block, Block* new_block, ::torch::onnx::OperatorExportTypes operator_export_type, std::unordered_map<Value*, Value*> env) {
  torch::autograd::SymbolicContext ctx;
  ctx.block = new_block;

  py::object onnx = py::module::import("torch.onnx");
  py::object onnx_symbolic = py::module::import("torch.onnx.symbolic");

  // Returns a node that n maps to in the new graph
  auto envFn = [&env](Value * n) -> Value* {
    auto it = env.find(n);
    JIT_ASSERTM(it != env.end(), "Dangling node reference");
    JIT_ASSERTM(it->second, "Unused node was subsequently used");
    return it->second;
  };

  // Initialize context and environment
  for (auto input : old_block->inputs()) {
    auto n = ctx.block->addInput()->copyMetadata(input);
    env[input] = n;
  }
  // Put the new outputs in our environment map, and copy the type from the
  // input graph if they were not set by the symbolic. This is called only
  // with results of symbolic call (not for nodes that are just cloned).
  auto setOutputs = [&](const std::string& op_name, Node * node, const value_list & outputs) {
    auto old_outputs = node->outputs();
    // Count all outputs, excluding Handles
    auto num_old_outputs = old_outputs.size();
    if (outputs.size() != num_old_outputs) {
      std::ostringstream ss;
      ss << "symbolic for " << op_name << " produced an incorrect number of outputs (expected ";
      ss << num_old_outputs << ", but got " << outputs.size() << ")";
      throw std::runtime_error(ss.str());
    }
    for (size_t i = 0; i < num_old_outputs; ++i) {
      auto old = old_outputs[i];
      if (outputs[i]) {
        // Allow symbolic() to skip specifying the type of the return node.
        // Unfortunately, they are on the hook for all internal nodes
        // (though in practice, the types are not computed.)
        outputs[i]->setType(old->type());
        // Copy over source location and scope information to all nodes
        // created by the symbolic
        outputs[i]->node()->setSourceLocation(node->getSourceLocation());
        outputs[i]->node()->setScope(node->scope());
        env[old] = outputs[i];
      } else {
        // Null output means that the ONNX op doesn't have outputs corresponding
        // to certain PyTorch outputs
        env[old] = nullptr;
        if (!old->uses().empty()) {
          std::ostringstream ss;
          ss << "symbolic for " << op_name << " returned None for the output " << i;
          ss << " (indicating conversion for that particular output is not supported), ";
          ss << "but the network uses this output later";
          // TODO: Say what actually used it
          throw std::runtime_error(ss.str());
        }
      }
    }
  };

  // Clone the node and add it to the new graph
  auto cloneNode = [&](Node * node) {
    auto n_ = ctx.block->appendNode(ctx.block->owningGraph()->createClone(node, envFn));
    for(size_t i = 0; i < node->outputs().size(); i++) {
      // n_->outputs()[i]->setType(node->outputs()[i]->type());
      env[node->outputs()[i]] = n_->outputs()[i];
    }
  };

  // Cast output of symbolic() python implementation
  auto processSymbolicOutput = [&](const std::string& op_name, Node* n, const py::object& raw_output) {
    if (raw_output.ptr() == Py_None) {
      cloneNode(n);
      return;
    }
    // Cast the outputs back to C++ and put them in the new graph
    std::vector<Value*> outputs;
    try {
      if (py::isinstance<Value>(raw_output)) {
        outputs = value_list{py::cast<Value*>(raw_output)};
      } else {
        outputs = py::cast<std::vector<Value*>>(raw_output);
      }
    } catch (const std::exception& ex) {
      std::ostringstream ss;
      ss << "Error casting results of symbolic for " << op_name
         << ": expected to return list of op nodes, instead received type ''"
         << py::str(raw_output.get_type()) << "': " << py::str(raw_output);
      throw std::runtime_error(ss.str());
    }

    setOutputs(op_name, n, outputs);
  };

  auto callPySymbolicFunction = [&](Node* n) {
    // The idea is delegate as much of the actual argument massaging to
    // Python as possible

    py::tuple py_inputs(n->inputs().size());
    Py_ssize_t input_nr = 0;
    for (auto* input : n->inputs()) {
        py_inputs[input_nr++] = py::cast(envFn(input));
    }

    WithInsertPoint insert_point_guard(ctx.block);
    WithCurrentScope scope_guard(*ctx.block->owningGraph(), n->scope());
    py::object raw_output = onnx.attr("_run_symbolic_function")(ctx.block->owningGraph(), n, py_inputs, env, operator_export_type);

    // TODO: Assert it's an ATen identifier???
    // (Sometimes it's not...)
    processSymbolicOutput(n->kind().toUnqualString(), n, raw_output);
  };

  auto callPySymbolicMethod = [&](PythonOp* op) {

    // Test if there is a symbolic function; bail if there is not
    auto pyobj = py::handle(op->pyobj.get());
    auto func = op->autogradFunction();
    if(func) {
      pyobj = func->get();
    }

    if(!py::hasattr(pyobj, "symbolic")) {
      cloneNode(op);
      return;
    }

    // Prepare args for Python. First one is the graph, and is followed
    // by regular args, with Variables replaced by corresponding nodes.
    Py_ssize_t input_nr = 0;
    py::tuple py_symbolic_args(1 + op->cconv.size());
    py_symbolic_args[input_nr++] = py::cast(ctx.block->owningGraph());
    auto inputs = op->inputs();
    auto node_it = inputs.begin();
    auto scalar_it = op->scalar_args.begin();
    for (auto arg_type : op->cconv) {
      py::object obj;
      if (arg_type == 'c') {
        JIT_ASSERTM(scalar_it != op->scalar_args.end(), "expected too many scalar args");
        obj = py::reinterpret_borrow<py::object>(py::handle((scalar_it++)->get()));
      } else if (arg_type == 'd') {
        JIT_ASSERTM(node_it != inputs.end(), "expected too many inputs");
        obj = py::cast(envFn(*node_it++));
      } else {
        throw std::runtime_error("unexpected calling convention");
      }
      py_symbolic_args[input_nr++] = obj;
    }

    WithInsertPoint insert_point_guard(ctx.block);
    WithCurrentScope scope_guard(*ctx.block->owningGraph(), op->scope());
    // Call the symbolic function
    // Use a little trampoline function so we can give good error messages
    // upon argument mismatch
    py::object raw_output = onnx.attr("_run_symbolic_method")(op->name(), pyobj.attr("symbolic"), py_symbolic_args);

    processSymbolicOutput(op->name(), op, raw_output);
  };

  // Finally, visit all nodes in the graph
  for (auto node : old_block->nodes()) {
    IR_IFM(node, PythonOp)
      callPySymbolicMethod(value);
    IR_ELSE()
      callPySymbolicFunction(node);
    IR_END()
  }
  for (auto output : old_block->outputs()) {
    ctx.block->registerOutput(env.at(output));
    env.at(output)->setType(output->type());
  }

  EliminateDeadCode(ctx.block);
}

}}