Update freezing API - changes from 52337 (pytorch#52392)

Co-authored-by: eellison <[email protected]>

Author: eellison

test/cpp/jit/test_module_api.cpp

@@ -3,9 +3,11 @@
 #include <test/cpp/jit/test_utils.h>
 
 #include <ATen/core/qualified_name.h>
+#include <torch/csrc/jit/api/module.h>
 #include <torch/csrc/jit/frontend/resolver.h>
 #include <torch/csrc/jit/serialization/import.h>
 #include <torch/csrc/jit/serialization/import_source.h>
+#include <torch/csrc/jit/testing/file_check.h>
 #include <torch/torch.h>
 
 namespace torch {
@@ -341,6 +343,20 @@ TEST(ModuleAPITest, Define) {
   AT_ASSERT(result.toTensor().item<float>() == 6);
 }
 
+TEST(ModuleAPITest, Freezing) {
+  Module m("m");
+  m.register_parameter("foo", torch::ones({}), false);
+  m.define(R"(
+    def forward(self, x, b : int = 4):
+      return self.foo + x + b
+  )");
+  m.eval();
+  auto frozen_mod = torch::jit::freeze(m);
+  auto forward_g = frozen_mod.get_method("forward").graph();
+  testing::FileCheck().check_not("GetAttr")->run(*forward_g);
+  ;
+}
+
 TEST(ModuleAPITest, To_CUDA) {
   Module m("test");
   {

test/jit/test_freezing.py

@@ -1508,7 +1508,7 @@ def test_optimize_freeze_module(self):
         bn = torch.nn.BatchNorm2d(out_channels, eps=.001)
         mod = torch.nn.Sequential(conv, bn)
         # set optimize to False here, by default freezing runs optimize_frozen_module
-        frozen_mod = torch.jit.freeze(torch.jit.script(mod.eval()), optimize=False)
+        frozen_mod = torch.jit.freeze(torch.jit.script(mod.eval()), optimize_numerics=False)
         # inspect frozen mod
         FileCheck().check("batch_norm").run(frozen_mod.graph)
         torch.jit.optimize_frozen_module(frozen_mod)

torch/_C/__init__.pyi.in

@@ -174,6 +174,11 @@ def _freeze_module(module: ScriptModule,
                    freeze_interfaces: _bool = True,
                    preserveParameters: _bool = True) -> ScriptModule: ...
 def _jit_pass_optimize_frozen_graph(Graph) -> None: ...
+def _jit_pass_fold_frozen_conv_bn(graph: Graph): ...
+def _jit_pass_fold_frozen_conv_add_or_sub(graph: Graph): ...
+def _jit_pass_fold_frozen_conv_mul_or_div(graph: Graph): ...
+def _jit_pass_remove_dropout(module: 'torch.jit.ScriptModule'): ...
+
 def _is_tracing() -> _bool: ...
 def _jit_init() -> _bool: ...
 def _jit_flatten(arg: Any) -> Tuple[List[Tensor], IODescriptor]: ...

torch/csrc/jit/api/module.cpp

@@ -8,6 +8,8 @@
 #include <torch/csrc/jit/frontend/schema_matching.h>
 #include <torch/csrc/jit/jit_log.h>
 #include <torch/csrc/jit/passes/dead_code_elimination.h>
+#include <torch/csrc/jit/passes/freeze_module.h>
+#include <torch/csrc/jit/passes/frozen_graph_optimizations.h>
 #include <torch/csrc/jit/passes/inliner.h>
 #include <torch/csrc/jit/runtime/operator.h>
 
@@ -336,6 +338,21 @@ IValue Module::create_class(const c10::QualifiedName& name, Stack stack) const {
   return obj;
 }
 
+Module freeze(
+    const Module& module,
+    c10::optional<std::vector<std::string>> preserved_attrs,
+    bool optimize_numerics) {
+  TORCH_CHECK(
+      module.is_training(),
+      "Freezing is currently only implemented for modules in eval mode. Please call .eval() before freezing");
+
+  Module out_mod = freeze_module(
+      module, preserved_attrs.value_or(std::vector<std::string>({})));
+  auto graph = module.get_method("forward").graph();
+  OptimizeFrozenGraph(graph, optimize_numerics);
+  return out_mod;
+}
+
 buffer_list Module::buffers(bool recurse) const {
   return buffer_list(*this, recurse, /*return_module=*/false);
 }

torch/csrc/jit/api/module.h

@@ -276,6 +276,13 @@ struct TORCH_API Module : public Object {
       bool non_blocking);
 };
 
+// C++ equivalent api of `torch.jit.freeze`. See documentation there for
+// details.
+TORCH_API Module freeze(
+    const Module& module,
+    c10::optional<std::vector<std::string>> preserved_attrs = c10::nullopt,
+    bool optimize_numerics = true);
+
 namespace detail {
 
 struct TORCH_API SlotCursor {

torch/csrc/jit/passes/frozen_graph_optimizations.cpp

@@ -8,12 +8,16 @@
 namespace torch {
 namespace jit {
 
-void OptimizeFrozenGraph(std::shared_ptr<Graph>& graph) {
+void OptimizeFrozenGraph(
+    std::shared_ptr<Graph>& graph,
+    bool optimize_numerics) {
   // run a couple times to capture Conv -> Mul -> Add etc
-  for (size_t i = 0; i < 2; i++) {
-    FoldFrozenConvBatchnorm(graph);
-    FoldFrozenConvAddOrSub(graph);
-    FoldFrozenConvMulOrDiv(graph);
+  if (optimize_numerics) {
+    for (size_t i = 0; i < 2; i++) {
+      FoldFrozenConvBatchnorm(graph);
+      FoldFrozenConvAddOrSub(graph);
+      FoldFrozenConvMulOrDiv(graph);
+    }
   }
 }
 

torch/csrc/jit/passes/frozen_graph_optimizations.h

@@ -13,7 +13,9 @@
 namespace torch {
 namespace jit {
 
-TORCH_API void OptimizeFrozenGraph(std::shared_ptr<Graph>& graph);
+TORCH_API void OptimizeFrozenGraph(
+    std::shared_ptr<Graph>& graph,
+    bool optimize_numerics = true);
 
 } // namespace jit
 } // namespace torch

torch/jit/_freeze.py

@@ -10,7 +10,7 @@
 from torch.jit._script import RecursiveScriptModule, ScriptModule
 
 
-def freeze(mod, preserved_attrs: Optional[List[str]] = None, optimize: bool = True):
+def freeze(mod, preserved_attrs: Optional[List[str]] = None, optimize_numerics: bool = True):
     r"""
     Freezing a :class:`ScriptModule` will clone it and attempt to inline the cloned
     module's submodules, parameters, and attributes as constants in the TorchScript IR Graph.
@@ -26,10 +26,8 @@ def freeze(mod, preserved_attrs: Optional[List[str]] = None, optimize: bool = Tr
         preserved_attrs (Optional[List[str]]): a list of attributes to preserve in addition to the forward method.
         Attributes modified in preserved methods will also be preserved.
 
-        optimize (bool): If ``True``, a set of optimization passes will be run to prepare the graph for inference,
-        in addition to the graph cleanup that already occurs. The details of the optimizations can be found in
-        `torch.jit.optimize_frozen_module.`
-
+        optimize_numerics (bool): If ``True``, a set of optimization passes will be run that does not strictly 
+        preserve numerics. Full details of optimization can be found at `torch.jit.optimize_frozen_module`.
 
     Returns:
         Frozen :class:`ScriptModule`.
@@ -102,23 +100,29 @@ def forward(self, input):
 
     out = RecursiveScriptModule(torch._C._freeze_module(mod._c, preserved_attrs))
     RecursiveScriptModule._finalize_scriptmodule(out)
-    if optimize:
-        optimize_frozen_module(out)
+    optimize_frozen_module(out, optimize_numerics)
 
     return out
 
 
-def optimize_frozen_module(mod):
+def optimize_frozen_module(mod, optimize_numerics: bool = True):
     r"""
     Runs a series of optimizations looking for patterns that occur in frozen graphs.
     The current set of optimizations is:
+        - Dropout Removal
         - Conv -> Batchnorm folding
         - Conv -> Add/Sub folding
         - Conv -> Mul/Div folding
 
     Args:
         mod (:class:`ScriptModule`): a frozen module to be optimized
 
+        optimize_numerics (bool): If ``True``, a set of optimization passes will be run that does not strictly 
+        preserve numerics. These optimizations preserve default rtol and atol of `torch.testing.assert_allclose` 
+        when applied on a single transformation, however in a module where many transformations are applied 
+        the rtol or atol may no longer fall within the default `assert_allclose` tolerance. Conv -> Batchnorm folding, 
+        Conv-Add/Sub, and Conv -> Mul/Div folding all may alter numerics.
+
     Returns:
         None
 
@@ -140,4 +144,12 @@ def optimize_frozen_module(mod):
         assert "batch_norm" not in str(frozen_mod.graph)
 
     """
-    torch._C._jit_pass_optimize_frozen_graph(mod.graph)
+    # xxx: keep in sync with frozen_graph_optimization.cpp
+    # intentionally duplicated to make to make it easier to create custom optimization sequence
+    torch._C._jit_pass_remove_dropout(mod._c)
+    if optimize_numerics:
+        # run a couple times to capture Conv -> Mul -> Add etc
+        for _ in range(2):
+            torch._C._jit_pass_fold_frozen_conv_bn(mod.graph)
+            torch._C._jit_pass_fold_frozen_conv_add_or_sub(mod.graph)
+            torch._C._jit_pass_fold_frozen_conv_mul_or_div(mod.graph)