Support tuples in ScriptModule inputs/outputs (#20784)

Summary:
- [x] Add tests after #20256 is merged

- Support exporting ScriptModule with inputs/outputs of arbitrarily constructed tuples.

- Moved the assigning of output shapes to after graph conversion to ONNX is completed. By then all tuples in the IR has already been lowered by the pass ```_jit_pass_lower_all_tuples```. If assigning output shapes is required to happen before that, we'll need to hand parse the tuple structures in the graph, and repeat the same logic in ```_jit_pass_lower_all_tuples```. Handling inputs is easier because all tuple information is encoded within the input tensor type.

- Swap the order of ```_jit_pass_lower_all_tuples``` and ```_jit_pass_erase_number_types```. Ops like ```prim::TupleIndex``` relies on index being a scalar. ```_jit_pass_erase_number_types``` will convert these kind of scalars to tensors.
Pull Request resolved: #20784

Reviewed By: zrphercule

Differential Revision: D15484171

Pulled By: houseroad

fbshipit-source-id: 4767a84038244c929f5662758047af6cb92228d3

test/onnx/test_pytorch_onnx_caffe2.py

@@ -1651,6 +1651,29 @@ def forward(self, lstm_in):
 
         self.run_model_test(MyModel(), train=False, input=lstm_in, batch_size=3, use_gpu=False)
 
+    def test_tuple_input_output(self):
+        class TupleModel(torch.jit.ScriptModule):
+            @torch.jit.script_method
+            def forward(self, a):
+                # type: (Tuple[Tensor, Tensor]) -> Tuple[Tensor, Tensor]
+                return a
+
+        x = (torch.randn(3, 4), torch.randn(4, 3))
+        self.run_model_test(TupleModel(), train=False, input=(x,), batch_size=BATCH_SIZE,
+                            example_outputs=(x,))
+
+    def test_nested_tuple_input_output(self):
+        class NestedTupleModel(torch.jit.ScriptModule):
+            @torch.jit.script_method
+            def forward(self, a, b):
+                # type: (Tensor, Tuple[Tensor, Tuple[Tensor, Tensor]]) -> Tensor
+                return a + b[0] + b[1][0] + b[1][1]
+
+        x = torch.randn(4, 5)
+        y = (torch.randn(4, 5), (torch.randn(4, 5), torch.randn(4, 5)))
+        self.run_model_test(NestedTupleModel(), train=False, input=(x, y), batch_size=BATCH_SIZE,
+                            example_outputs=x + y[0] + y[1][0] + y[1][1])
+
     def test_topk(self):
         class TopKModel(torch.nn.Module):
             def forward(self, input):

torch/csrc/jit/script/init.cpp

@@ -178,11 +178,58 @@ static Self moduleSelf(
   };
 }
 
-static void setInputTensorTypes(Graph& g, const Stack& stack) {
-  AT_ASSERT(stack.size() == g.inputs().size());
-  for (size_t i = 0; i < stack.size(); ++i) {
-    g.inputs().at(i)->setType(
-        DimensionedTensorType::create(stack.at(i).toTensor()));
+static TypePtr getTensorType(
+    const at::Tensor& t,
+    const TypeKind type_kind) {
+  switch (type_kind) {
+    case TypeKind::DimensionedTensorType:
+      return DimensionedTensorType::create(t);
+    case TypeKind::CompleteTensorType: {
+      auto scalar_type = t.scalar_type();
+      auto sizes = t.sizes();
+      return CompleteTensorType::create(scalar_type, at::kCPU, sizes);
+    }
+    default:
+      throw std::runtime_error(
+          "Attempted to call getTensorType for type kind other than DimensionedTensorType or CompleteTensorType.");
+  }
+}
+
+static TupleTypePtr getTupleTensorType(
+    const Stack::const_iterator& s_iter,
+    const Stack::const_iterator& s_iter_end,
+    const TypePtr& tupleType,
+    const TypeKind type_kind) {
+  AT_ASSERT(tupleType->kind() == TupleType::Kind);
+  AT_ASSERT(s_iter != s_iter_end);
+
+  std::vector<TypePtr> types;
+  for (const auto& subType : tupleType->containedTypes()) {
+    if (subType->kind() == TupleType::Kind) {
+      types.push_back(getTupleTensorType(s_iter+1, s_iter_end, subType, type_kind));
+    } else {
+      types.push_back(getTensorType(s_iter->toTensor(), type_kind));
+    }
+  }
+  return TupleType::create(types);
+}
+
+static void setInputTensorTypes(
+    Graph& g,
+    const Stack& stack,
+    const TypeKind type_kind = TypeKind::DimensionedTensorType) {
+  at::ArrayRef<Value*> input_values = g.inputs();
+  auto s_iter = stack.begin();
+  for (auto v : input_values) {
+    AT_ASSERT(s_iter != stack.end());
+    if (v->type()->kind() == TupleType::Kind) {
+      AT_ASSERT(v->node()->kind() == prim::Param);
+      v->setType(
+          getTupleTensorType(s_iter, stack.end(), v->type(), type_kind));
+    } else {
+      v->setType(getTensorType(s_iter->toTensor(), type_kind));
+      s_iter++;
+    }
   }
 }
 
@@ -197,38 +244,32 @@ static std::shared_ptr<Graph> _propagate_shapes(
   return retval;
 }
 
-static std::shared_ptr<Graph> _propagate_and_assign_input_and_output_shapes(
+static std::shared_ptr<Graph> _propagate_and_assign_input_shapes(
     Graph& graph,
-    std::vector<at::Tensor> inputs,
-    std::vector<at::Tensor> outputs,
+    const std::vector<at::Tensor>& inputs,
     bool with_grad = false,
     bool propagate = true) {
   auto retval = graph.copy();
   if (propagate) {
-    setInputTensorTypes(*retval, fmap<IValue>(inputs));
+    setInputTensorTypes(*retval, fmap<IValue>(inputs), TypeKind::DimensionedTensorType);
     PropagateInputShapes(retval);
   }
-  AT_ASSERT(retval->inputs().size() == inputs.size());
-  for (size_t i = 0; i < retval->inputs().size(); ++i) {
-    auto scalar_type = inputs[i].scalar_type();
-    auto sizes = inputs[i].sizes();
-    auto type =
-        torch::jit::CompleteTensorType::create(scalar_type, at::kCPU, sizes);
-    retval->inputs()[i]->setType(type);
-  }
-  at::ArrayRef<Value*> output_values = retval->outputs();
-  // patch this to still work if we are returning a tuple of multiple values
-  if (output_values.at(0)->type()->kind() == TupleType::Kind) {
-    AT_ASSERT(output_values.at(0)->node()->kind() == prim::TupleConstruct);
-    output_values = output_values.at(0)->node()->inputs();
-  }
-  AT_ASSERT(output_values.size() == outputs.size());
+  setInputTensorTypes(*retval, fmap<IValue>(inputs), TypeKind::CompleteTensorType);
+
+  return retval;
+}
+
+static std::shared_ptr<Graph> _assign_output_shapes(
+    Graph& graph,
+    std::vector<at::Tensor> outputs) {
+  auto retval = graph.copy();
+  AT_ASSERT(retval->outputs().size() == outputs.size());
   for (size_t i = 0; i < outputs.size(); ++i) {
     auto scalar_type = outputs[i].scalar_type();
     auto sizes = outputs[i].sizes();
     auto type =
         torch::jit::CompleteTensorType::create(scalar_type, at::kCPU, sizes);
-    output_values[i]->setType(type);
+    retval->outputs()[i]->setType(type);
   }
   return retval;
 }
@@ -679,8 +720,11 @@ void initJitScriptBindings(PyObject* module) {
       debugSetAutodiffSubgraphInlining);
   m.def("_propagate_shapes", _propagate_shapes);
   m.def(
-      "_propagate_and_assign_input_and_output_shapes",
-      _propagate_and_assign_input_and_output_shapes);
+      "_propagate_and_assign_input_shapes",
+      _propagate_and_assign_input_shapes);
+  m.def(
+      "_assign_output_shapes",
+      _assign_output_shapes);
   m.def("_jit_python_print", [](py::object obj) {
     std::ostringstream ss;
     std::vector<at::Tensor> constants;

torch/onnx/utils.py

@@ -16,7 +16,7 @@
 from torch._six import string_classes
 from torch.jit import _unique_state_dict
 from torch.onnx import ONNX_ARCHIVE_MODEL_PROTO_NAME, ExportTypes, OperatorExportTypes
-from torch._C import ListType, _propagate_and_assign_input_and_output_shapes
+from torch._C import ListType, _propagate_and_assign_input_shapes, _assign_output_shapes
 
 
 # the flag to tell the user whether it's in the middle of ONNX export or not
@@ -214,10 +214,10 @@ def _optimize_graph(graph, operator_export_type, _disable_torch_constant_prop=Fa
 
     # onnx only supports tensors, but 1 / 2 = 0.5 and tensor(1) / tensor(2) = 0
     torch._C._jit_pass_prepare_division_for_onnx(graph)
-    # onnx only supports tensors, so we turn all out number types into tensors
-    torch._C._jit_pass_erase_number_types(graph)
     # onnx does not support tuples, so try to remove them
     torch._C._jit_pass_lower_all_tuples(graph)
+    # onnx only supports tensors, so we turn all out number types into tensors
+    torch._C._jit_pass_erase_number_types(graph)
     torch._C._jit_pass_peephole(graph, True)
     torch._C._jit_pass_lint(graph)
 
@@ -288,16 +288,18 @@ def _model_to_graph(model, args, verbose=False, training=False,
         assert example_outputs is not None, "example_outputs must be provided when exporting a ScriptModule"
         try:
             method_graph, params = model.forward._lowered_graph()
-            graph = _propagate_and_assign_input_and_output_shapes(
-                method_graph, tuple(args) + tuple(params), example_outputs, False, propagate)
+            in_vars, in_desc = torch.jit._flatten(tuple(args) + tuple(params))
+            graph = _propagate_and_assign_input_shapes(
+                method_graph, tuple(in_vars), False, propagate)
         except AttributeError:
             raise RuntimeError('\'forward\' method must be a script method')
     elif isinstance(model, torch.jit.Function):
         assert example_outputs is not None, "example_outputs must be provided when exporting a TorchScript Function"
         method = model
         params = ()
-        graph = _propagate_and_assign_input_and_output_shapes(
-            model.graph, tuple(args), example_outputs, False, propagate)
+        in_vars, in_desc = torch.jit._flatten(tuple(args))
+        graph = _propagate_and_assign_input_shapes(
+            model.graph, tuple(in_vars), False, propagate)
     else:
         graph, torch_out = _trace_and_get_graph_from_model(model, args, training)
         state_dict = _unique_state_dict(model)
@@ -313,6 +315,10 @@ def _model_to_graph(model, args, verbose=False, training=False,
     graph = _optimize_graph(graph, operator_export_type,
                             _disable_torch_constant_prop=_disable_torch_constant_prop)
 
+    if isinstance(model, torch.jit.ScriptModule) or isinstance(model, torch.jit.Function):
+        out_vars, _ = torch.jit._flatten(tuple(example_outputs))
+        graph = _assign_output_shapes(graph, out_vars)
+
     # NB: ONNX requires complete information about output types, which might be
     # erased by some optimizations, so we need to set it explicitly again.
     if torch_out is not None: