pytorch/torch/csrc/jit/python_ir.cpp

#include "torch/csrc/python_headers.h"

#include "torch/csrc/jit/ir.h"
#include "torch/csrc/jit/import.h"
#include "torch/csrc/jit/pybind.h"
#include "torch/csrc/jit/python_tracer.h"
#include "torch/csrc/utils/pybind.h"
#include "torch/csrc/jit/export.h"
#include "torch/csrc/jit/passes/shape_analysis.h"
#include "torch/csrc/jit/argument_spec.h"
#include "torch/csrc/utils/auto_gil.h"
#include "torch/csrc/utils/python_strings.h"


#include <iostream>
#include <sstream>

namespace torch { namespace jit {

std::string getPythonName(const PyObject* obj_) {
  AutoGIL gil;
  PyObject* obj = const_cast<PyObject*>(obj_);
  auto v = py::getattr(obj, "__name__", py::str("<python_value>"));
  // if this was a autograd.Function recover the name of the class
  return py::str(v);
}

std::ostream& printPyObject(std::ostream & out, const THPObjectPtr& obj) {
  AutoGIL gil;
  auto pyobj = py::handle(const_cast<PyObject*>(obj.get()));
  if (py::isinstance<py::tuple>(pyobj)) {
    // This special-case for printing tuples handles a problem where
    // str((2L, 3L)) outputs "(2L, 3L)" in Python 2 but "(2, 3)"
    // in Python 3.  In order to suppress the L-suffix, we must
    // manually print the string ourselves, calling str() on the
    // sub-elements.
    //
    // This is a fairly fragile fix (What if you have nested tuples
    // in tuples? What if you have dictionaries?) but it seems to hit
    // the cases that are triggered in practice in onnx-pytorch.  Revisit
    // this code if this is not the case.
    //
    // By the way, one non-solution for this problem is to monkeypatch
    // tuple.__str__; this doesn't work because Python doesn't allow
    // monkeypatching methods of built-in types.
    auto pytuple = pyobj.cast<py::tuple>();
    out << "(";
    size_t i = 0;
    for (auto& o : pytuple) {
      if (i > 0) {
        out << ", ";
      }
      THPObjectPtr str(py::str(o).release().ptr());
      out << THPUtils_unpackString(str.get());
      i++;
    }
    if (i == 1) {
      out << ",";
    }
    out << ")";
    return out;
  } else {
    return out << THPUtils_unpackString(py::str(pyobj).ptr());
  }
}

// execute a Python function, used for Ops we can't optimize but that we want to optimize around
struct ConcretePythonOp : public PythonOp {
 ConcretePythonOp(Graph * graph)
 : PythonOp(graph) {}
 virtual std::string name() const override {
   AutoGIL gil;
   if(auto autograd = autogradFunction()) {
     return getPythonName(autograd->get());
   } else {
     return getPythonName(pyobj.get());
   }
 }
 virtual void cloneFrom(Node * other_) override {
   Node::cloneFrom(other_);
   auto other = other_->cast<PythonOp>();
   this->cconv = other->cconv;
   Py_INCREF(other->pyobj.get());
   this->pyobj = THPObjectPtr(other->pyobj.get());
   for(auto & sa : other->scalar_args) {
     Py_INCREF(sa.get());
     this->scalar_args.emplace_back(sa.get());
   }
 }
 virtual Node * allocNewInstance(Graph * g) override {
   return new ConcretePythonOp(g);
 }
 // recover the autograd.Function instance, if this PythonOp's function
 // was originally SomeFunction.apply
 // used in ONNX for discovering symbolics
 virtual at::optional<THPObjectPtr> autogradFunction() const override {
   AutoGIL gil;
   py::handle obj = const_cast<PyObject*>(pyobj.get());

   auto r = py::getattr(obj, "__self__", py::none());
   if(r.is_none())
     return at::nullopt;

   auto apply = py::getattr(r, "apply", py::none());
   if(apply.is_none())
     return at::nullopt;

   auto c = PyObject_RichCompareBool(apply.ptr(), obj.ptr(), Py_NE);
   if(PyErr_Occurred())
     throw py::error_already_set();
   if(c)
     return at::nullopt;

   return THPObjectPtr(r.release().ptr());
 }

 virtual void writeScalars(std::ostream& out) const override {
   out << "(";
   int i = 0;
   for (auto& scalar : scalar_args) {
     if (i++ > 0)
       out << ", ";
     printPyObject(out, scalar);
   }
   out << ")";
 }

};

PythonOp* pythonAllocPythonOp(Graph* g) {
  return new ConcretePythonOp(g);
}

void initPythonIRBindings(PyObject * module_) {
  setAllocPythonOp(pythonAllocPythonOp);

  auto m = py::handle(module_).cast<py::module>();
  #define GS(name) \
    def(#name,&Graph :: name)
  py::class_<Graph,std::shared_ptr<Graph>>(m,"Graph")
    .def(py::init<>())
    .def("__repr__",[](Graph & g) {
      std::stringstream ss;
      ss << g;
      return ss.str();
    })
    .def("propagate_shapes", [](Graph& g, std::vector<at::Tensor> inputs, bool with_grad) {
      PropagateInputShapes(g, ArgumentSpec(with_grad, variable_tensor_list(std::move(inputs))));
    })
    .def("export", [](const std::shared_ptr<Graph> g, const std::vector<at::Tensor>& initializers,
                      int64_t onnx_opset_version, bool defer_weight_export,
                      ::torch::onnx::OperatorExportTypes operator_export_type) {
      std::string graph;
      RawDataExportMap export_map;
      std::tie(graph, export_map) = ExportGraph(
        g, initializers, onnx_opset_version, defer_weight_export, operator_export_type);
      std::unordered_map<std::string, py::bytes> python_serialized_export_map;
      for (auto& kv : export_map) {
        auto t = kv.second;
        size_t copy_bytes = t.type().elementSizeInBytes() * t.numel();
        // TODO: this is an unecessary copy. In theory we can directly return
        // the map from identifier to Tensor, but we need some API in Python
        // to get raw `bytes` containing the raw tensor data.
        python_serialized_export_map[kv.first] = py::bytes(static_cast<const char*>(t.data_ptr()), copy_bytes);
      }
      return std::make_tuple(py::bytes(graph), python_serialized_export_map);
    }, py::arg("initializers"),
       py::arg("onnx_opset_version")=0,
       py::arg("defer_weight_export")=false,
       py::arg("operator_export_type")=::torch::onnx::OperatorExportTypes::ONNX)
    .def("prettyPrintExport", [](const std::shared_ptr<Graph> g, const std::vector<at::Tensor>& initializers,
                      int64_t onnx_opset_version, bool defer_weight_export,
                      ::torch::onnx::OperatorExportTypes operator_export_type) {
      return PrettyPrintExportedGraph(
        g, initializers, onnx_opset_version, defer_weight_export, operator_export_type);
    }, py::arg("initializers"),
       py::arg("onnx_opset_version")=0,
       py::arg("defer_weight_export")=false,
       py::arg("operator_export_type")=::torch::onnx::OperatorExportTypes::ONNX)
    .def("wrapPyFuncWithSymbolic", [](Graph &g, py::function func, std::vector<Value*> inputs, size_t n_outputs, py::function symbolic) {
      // This function should be used for situations where we have a Python function
      // that should have different behavior when exporting for JIT interpreter
      // execution v.s. for ONNX export. For example, nn.utils.rnn.pack_padded_sequence
      // emits a placeholder under ONNX export, but we want to keep the ability to
      // run this in the interpreter, thus we emit a PythonOp for that use case.

      // Concretely, this function emits a PythonOp wrapping the passed-in
      // parameter `func`, while storing the function `symbolic` for use by the
      // ONNX export
      std::string cconv(inputs.size(), 't');
      func.attr("symbolic") = symbolic;
      Node* new_node = g.insertNode(g.createPythonOp(
        THPObjectPtr(func.release().ptr()), cconv, {}));
      for (auto i : inputs)
        new_node->addInput(i);
      std::vector<Value*> outputs;
      for (size_t i = 0; i < n_outputs; ++i)
        new_node->addOutput();
      auto sl = std::make_shared<StringSourceLocation>(tracer::getPythonInterpreterStackTrace());
      new_node->setSourceLocation(sl);
      return py::make_iterator(new_node->outputs().begin(), new_node->outputs().end());
    }, py::return_value_policy::reference_internal)
    .def("inputs",[](Graph &g) {
      return py::make_iterator(g.inputs().begin(), g.inputs().end());
    })
    .def("outputs",[](Graph &g) {
      return py::make_iterator(g.outputs().begin(), g.outputs().end());
    })
    // TODO: Iterator invalidation might make this hazardous
    .def("nodes",[](Graph &g) {
      return py::make_iterator(g.nodes().begin(), g.nodes().end());
    })
    .def("addInput",[](Graph &g) { return g.addInput(); })
    .def("copy",[](Graph &g) {
      return g.copy();
    })
    .GS(advanceStage)
    .GS(stage)
    .GS(eraseInput)
    .GS(registerOutput)
    .def("create",[](Graph & g, const char * str) {
      return g.create(Symbol::fromQualString(str));
    })
    .def("create",[](Graph & g, const char * str, size_t noutputs) {
      return g.create(Symbol::fromQualString(str), noutputs);
    })
    .def("create",[](Graph & g, const char * str, const std::vector<Value*> & inputs) {
      return g.create(Symbol::fromQualString(str),inputs);
    })
    .def("create",[](Graph & g, const char * str, const std::vector<Value*> & inputs, size_t noutputs) {
      return g.create(Symbol::fromQualString(str),inputs, noutputs);
    })
    .def("param_node", [](Graph &g) {
      return g.block()->param_node();
    })
    .def("return_node", [](Graph &g) {
      return g.block()->return_node();
    })
    .GS(createFusionGroup)
    .def("createClone",[](Graph & g, Node * n, py::object fn) {
      return g.createClone(n, [&](Value * e) {
        return fn(e).cast<Value*>();
      });
    })
    .GS(appendNode)
    .GS(prependNode)
    .GS(lint)
    .GS(insertNode)
    ;
    #undef GS

  #define VS(name) \
    def(#name,&Value :: name)
  py::class_<Value,std::unique_ptr<Value, py::nodelete>>(m,"Value")
    .def("__repr__",[](Value & n) {
      std::stringstream ss;
      ss << n.uniqueName() << " defined in (" << *n.node() << ")";
      return ss.str();
    })
    .VS(type)
    .VS(setType)
    .VS(inferTypeFrom)
    // skip owningGraph because it returns a raw pointer to a otherwise
    // std::shared_ptr stored graph object, and would cause a double free
    .VS(unique)
    .VS(uniqueName)
    .VS(setUniqueName)
    .VS(setStage)
    .VS(stage)
    .VS(offset)
    .VS(uses)
    .VS(replaceAllUsesWith)
    .def("node",[](Value &v) { return v.node(); })
    .def("setTypeAs", [](Value * node, Value * other) {
      node->setType(other->type());
      return node;
    })
    .VS(copyMetadata)
    .VS(isTensor)
    ;

  #undef VS

  py::class_<Block, std::unique_ptr<Block, py::nodelete>>(m, "Block");

  #define NS(name) \
    def(#name,&Node :: name)
  py::class_<Node,std::unique_ptr<Node, py::nodelete>>(m,"Node")
    .def("__repr__",[](Node & n) {
      std::stringstream ss;
      ss << n;
      return ss.str();
    })
    .def("hasMultipleOutputs",[](Node&n) {
      return n.outputs().size() > 1;
    })
    .def("outputsSize",[](Node &n) {
      return n.outputs().size();
    })
    .NS(kind)
    .NS(stage)
    .NS(setStage)
    .def("inputs",[](Node &n) {
      return py::make_iterator(n.inputs().begin(), n.inputs().end());
    })
    .def("outputs",[](Node &n) {
      return py::make_iterator(n.outputs().begin(), n.outputs().end());
    })
    .NS(output)
    .NS(addInput)
    .NS(replaceInput)
    .NS(replaceInputWith)
    .NS(replaceAllUsesWith)
    .NS(insertBefore)
    .NS(insertAfter)
    .NS(moveAfter)
    .NS(moveBefore)
    .NS(removeInput)
    .NS(removeAllInputs)
    .NS(destroy)
    .NS(hasUses)
    .NS(eraseOutput)
    .NS(addOutput)
    .NS(scopeName)
    .def("blocks", [](Node& n) {
      return py::make_iterator(n.blocks().begin(), n.blocks().end());
    })
    .NS(addBlock)

#define AS(name) def(#name,&Attributes<Node> :: name)
    // methods from Attributes
    .AS(copyAttributes)
    .AS(hasAttributes)
#undef AS
#define AS(name) def(#name,&Attributes<Node> :: name ## S)
    // The default method names take Symbol, but the string conversion for
    // Symbol you to qualify with attr::. This is not very user friendly
    // for attributes, so expose the string variants instead.
    .AS(hasAttribute)
    .AS(kindOf)
    .AS(removeAttribute)
    .AS(attributeNames)
#undef AS
#define CREATE_ACCESSOR(Kind,method) \
    def(#method "_",[](Node & n, const char * name, Kind##Attr::ValueType v) { \
      return n . method ## _(Symbol::attr(name), std::move(v)); \
    }) \
    .def(#method, [](Node & n, const char * name) { \
      return n.method(Symbol::attr(name)); \
    })
    .CREATE_ACCESSOR(Float,f)
    .CREATE_ACCESSOR(Floats,fs)
    .CREATE_ACCESSOR(String,s)
    .CREATE_ACCESSOR(Strings,ss)
    .CREATE_ACCESSOR(Int,i)
    .CREATE_ACCESSOR(Ints,is)
    .CREATE_ACCESSOR(Graph,g)
    .CREATE_ACCESSOR(Graphs,gs)
#undef CREATE_ACCESSOR
    // Tensor (t_) -- manually written to unwrap the variable into a tensor.
    .def("t_",[](Node & n, const char * name, torch::autograd::Variable v) {
      return n.t_(Symbol::attr(name), std::move(v.data()));
    })
    .def("t", [](Node & n, const char * name) {
      return torch::autograd::make_variable(n.t(Symbol::attr(name)), /*requires_grad=*/false);
    })
    // Tensors (ts_) -- manually written to unwrap variables into tensors.
    .def("ts_",[](Node & n, const char * name, std::vector<torch::autograd::Variable> vs) {
      std::vector<at::Tensor> tensors;
      tensors.reserve(vs.size());
      for (auto& variable : vs) {
        tensors.push_back(std::move(variable.data()));
      }
      return n.ts_(Symbol::attr(name), std::move(tensors));
    })
    .def("ts", [](Node & n, const char * name) {
      auto tensors = n.ts(Symbol::attr(name));
      std::vector<torch::autograd::Variable> variables;
      variables.reserve(tensors.size());
      for (auto& tensor : tensors) {
        variables.push_back(torch::autograd::make_variable(
            std::move(tensor), /*requires_grad=*/false));
      }
      return variables;
    })
    .def("z_",[](Node & n, const char * name, at::Tensor v) {
        return n.t_(Symbol::attr(name), autograd::Variable(v.view({})).data());
    })
    .def("z",[](Node & n, const char * name) {
        return n.t(Symbol::attr(name));
    })
    .def("zs_",[](Node & n, const char * name, TensorsAttr::ValueType v) {
        for (size_t i = 0; i < v.size(); ++ i) {
            v[i] = autograd::Variable(v[i].view({})).data();
        }
        return n.ts_(Symbol::attr(name), std::move(v));
    })
    .def("zs",[](Node & n, const char * name) {
        return n.ts(Symbol::attr(name));
    })
    .def("pyobj",[](Node & n) {
      return py::handle(n.expect<PythonOp>()->pyobj.get()).cast<py::object>();
    })
    .def("cconv",[](Node & n) {
      return n.expect<PythonOp>()->cconv;
    })
    .def("pyname",[](Node & n) {
      return n.expect<PythonOp>()->name();
    })
    .def("scalar_args",[](Node & n) {
      auto op = n.expect<PythonOp>();
      auto scalars = py::list();
      auto append = scalars.attr("append");
      for(auto & arg : op->scalar_args) {
        append(py::handle(arg.get()));
      }
      return scalars;
    })
    ;

  py::class_<Type,std::shared_ptr<Type>>(m,"Type")
    .def("__repr__",[](Type & t) {
      return t.str();
    })
    .def("kind",[](Type& t_) {
      Type * t = &t_;
      switch(t->kind()) {
        case TypeKind::DynamicType:
          return "DynamicType";
        case TypeKind::TensorType:
          return "TensorType";
        case TypeKind::TupleType:
          return "TupleType";
        default:
          torch::barf("unknown type kind");
          return "";
        }
    })
    .def("sizes",[](Type& t) {
      return t.expect<TensorType>()->sizes();
    })
    .def("strides",[](Type& t) {
      return t.expect<TensorType>()->strides();
    })
    .def("contiguous",[](Type& t) {
      return std::static_pointer_cast<Type>(t.expect<TensorType>()->contiguous());
    })
    .def("scalarType",[](Type& t) {
      return at::toString(t.expect<TensorType>()->scalarType());
    })
    ;

  py::class_<DynamicType, Type, std::shared_ptr<DynamicType>>(m, "DynamicType")
    .def(py::init<>());
  py::class_<TupleType, Type, std::shared_ptr<TupleType>>(m, "TupleType")
    .def(py::init<std::vector<TypePtr>>());

  py::class_<Use>(m,"Use")
  .def_readonly("user",&Use::user)
  .def_readonly("offset",&Use::offset);

  m.def("_jit_import_graph", [](const std::string& serialized_graph) {
    std::vector<at::Tensor> initializers;
    auto graph = ImportIRGraph(serialized_graph, initializers);
    std::vector<torch::autograd::Variable> variables;
    variables.reserve(initializers.size());
    for (auto& tensor : initializers) {
      variables.push_back(torch::autograd::make_variable(
          std::move(tensor), /*requires_grad=*/false));
    }
    return std::make_tuple(graph, variables);
  });
}
}}