fix: Final working version of QAT in TRTorch

peri044 · peri044 · commit 521a0cbc227a · 2021-07-25T18:10:27.000-07:00
Signed-off-by: Dheeraj Peri &lt;peri.dheeraj@gmail.com&gt;
diff --git a/core/compiler.cpp b/core/compiler.cpp
@@ -119,7 +119,7 @@ void AddEngineToGraph(
 
 bool CheckMethodOperatorSupport(const torch::jit::script::Module& mod, std::string method_name) {
   // Go through Lowering to simplify graph and extract weight parameters
-  auto graph_and_parameters = lowering::Lower(mod, method_name);
+  auto graph_and_parameters = lowering::Lower(mod, method_name, false);
 
   auto g = graph_and_parameters.first;
   LOG_DEBUG(*g << "(CheckMethodOperatorSupport)\n");
@@ -129,7 +129,7 @@ bool CheckMethodOperatorSupport(const torch::jit::script::Module& mod, std::stri
 
 std::string ConvertGraphToTRTEngine(const torch::jit::script::Module& mod, std::string method_name, CompileSpec cfg) {
   // Go through Lowering to simplify graph and extract weight parameters
-  auto graph_and_parameters = lowering::Lower(mod, method_name);
+  auto graph_and_parameters = lowering::Lower(mod, method_name, cfg.convert_info.engine_settings.unfreeze_module);
 
   auto convert_cfg = std::move(cfg.convert_info);
   auto g = graph_and_parameters.first;
@@ -187,7 +187,7 @@ torch::jit::script::Module CompileGraphWithFallback(const torch::jit::script::Mo
     // Compile only forward methods. forward method contains the entire graph.
     if (method.name().compare("forward") == 0) {
       auto new_g = std::make_shared<torch::jit::Graph>();
-      auto graph_and_parameters = lowering::Lower(mod, method.name());
+      auto graph_and_parameters = lowering::Lower(mod, method.name(), cfg.convert_info.engine_settings.unfreeze_module);
 
       auto g = graph_and_parameters.first;
       auto params = graph_and_parameters.second;
diff --git a/core/conversion/conversionctx/ConversionCtx.cpp b/core/conversion/conversionctx/ConversionCtx.cpp
@@ -72,7 +72,7 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
         if (!settings.calibrator) {
           LOG_WARNING(
               "Int8 precision has been enabled but no calibrator provided. This assumes the network has Q/DQ nodes obtained from Quantization aware training. For more details, refer to https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#work-with-qat-networks");
-        } else{
+        } else {
           cfg->setInt8Calibrator(settings.calibrator);
         }
         break;
diff --git a/core/conversion/conversionctx/ConversionCtx.h b/core/conversion/conversionctx/ConversionCtx.h
@@ -27,6 +27,8 @@ struct BuilderSettings {
   bool sparse_weights = false;
   std::set<nvinfer1::DataType> enabled_precisions = {nvinfer1::DataType::kFLOAT};
   bool disable_tf32 = false;
+  // Internal flag to ensure torch.jit.Module does not get freezed in lowering.cpp. This is required for QAT models.
+  bool unfreeze_module = false;
   bool refit = false;
   bool debug = false;
   bool strict_types = false;
diff --git a/core/conversion/converters/impl/matrix_multiply.cpp b/core/conversion/converters/impl/matrix_multiply.cpp
@@ -26,6 +26,7 @@ auto mm_registrations TRTORCH_UNUSED =
 
                     auto mm_layer = ctx->net->addMatrixMultiply(
                         *self, nvinfer1::MatrixOperation::kNONE, *other, nvinfer1::MatrixOperation::kNONE);
+
                     TRTORCH_CHECK(mm_layer, "Unable to create matrix multiplication node: " << *n);
                     mm_layer->setName(util::node_info(n).c_str());
                     auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], mm_layer->getOutput(0));
@@ -73,51 +74,6 @@ auto mm_registrations TRTORCH_UNUSED =
 
                LOG_DEBUG("Output tensor shape: " << out_tensor->getDimensions());
                return true;
-             }})
-        .pattern(
-            {"aten::addmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> (Tensor)",
-             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
-               auto self = args[0].ITensorOrFreeze(ctx);
-               auto mat1 = args[1].ITensorOrFreeze(ctx);
-               auto mat2 = args[2].ITensorOrFreeze(ctx);
-               auto beta = args[3].unwrapToScalar().to<float>();
-               auto betaTensor = tensor_to_const(ctx, torch::tensor({beta}));
-               auto alpha = args[4].unwrapToScalar().to<float>();
-               auto alphaTensor = tensor_to_const(ctx, torch::tensor({alpha}));
-
-               // Ensure self and other tensors have same nbDims by expanding the dimensions (from 0 axis) if
-               // necessary.
-               if (mat1->getDimensions().nbDims < mat2->getDimensions().nbDims) {
-                 mat1 = addPadding(ctx, n, mat1, mat2->getDimensions().nbDims, false, false);
-               } else {
-                 mat2 = addPadding(ctx, n, mat2, mat1->getDimensions().nbDims, false, false);
-               }
-
-               auto mm_layer = ctx->net->addMatrixMultiply(
-                   *mat1, nvinfer1::MatrixOperation::kNONE, *mat2, nvinfer1::MatrixOperation::kNONE);
-               TRTORCH_CHECK(mm_layer, "Unable to create matrix multiplication layer in node: " << *n);
-               auto mm_scale_layer = add_elementwise(
-                   ctx,
-                   nvinfer1::ElementWiseOperation::kPROD,
-                   mm_layer->getOutput(0),
-                   alphaTensor,
-                   util::node_info(n) + "_alphaScale");
-               TRTORCH_CHECK(mm_scale_layer, "Unable to create alpha scaling layer in node: " << *n);
-               auto beta_scale_layer = add_elementwise(
-                   ctx, nvinfer1::ElementWiseOperation::kPROD, self, betaTensor, util::node_info(n) + "_betaScale");
-               TRTORCH_CHECK(beta_scale_layer, "Unable to create beta scaling layer in node: " << *n);
-               auto add_mm_layer = add_elementwise(
-                   ctx,
-                   nvinfer1::ElementWiseOperation::kSUM,
-                   beta_scale_layer->getOutput(0),
-                   mm_scale_layer->getOutput(0),
-                   util::node_info(n));
-               TRTORCH_CHECK(add_mm_layer, "Unable to create addmm layer in node: " << *n);
-
-               auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], mm_layer->getOutput(0));
-
-               LOG_DEBUG("[AddMM layer] Output tensor shape: " << out_tensor->getDimensions());
-               return true;
              }});
 } // namespace
 } // namespace impl
diff --git a/core/conversion/converters/impl/shuffle.cpp b/core/conversion/converters/impl/shuffle.cpp
@@ -131,15 +131,22 @@ static auto shuffle_registrations TRTORCH_UNUSED =
                   [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                     auto in = args[0].ITensorOrFreeze(ctx);
                     auto input_dims = in->getDimensions();
-                    nvinfer1::Dims transposed_input_dims;
-                    transposed_input_dims.nbDims = input_dims.nbDims;
-                    for (int i = input_dims.nbDims - 1; i >= 0; i--) {
-                      transposed_input_dims.d[i] = input_dims.d[input_dims.nbDims - 1 - i];
+                    // For input tensors < 2D, return them as is
+                    // For a 2D input tensor, return transpose(input, 0, 1) which is a general 2d matrix transpose.
+                    if (input_dims.nbDims < 2) {
+                      auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], in);
+                      LOG_DEBUG("Output tensor shape: " << out_tensor->getDimensions());
+                      return true;
                     }
+
                     auto shuffle_layer = ctx->net->addShuffle(*in);
                     TRTORCH_CHECK(shuffle_layer, "Unable to create shuffle layer from node: " << *n);
-                    shuffle_layer->setReshapeDimensions(transposed_input_dims);
-                    shuffle_layer->setZeroIsPlaceholder(true);
+                    nvinfer1::Permutation firstPerm;
+                    firstPerm.order[0] = 1;
+                    firstPerm.order[1] = 0;
+
+                    shuffle_layer->setFirstTranspose(firstPerm);
+                    shuffle_layer->setZeroIsPlaceholder(false);
                     shuffle_layer->setName(util::node_info(n).c_str());
 
                     auto out_tensor = ctx->AssociateValueAndTensor(n->outputs()[0], shuffle_layer->getOutput(0));
diff --git a/core/conversion/evaluators/aten.cpp b/core/conversion/evaluators/aten.cpp
@@ -427,36 +427,6 @@ auto aten_registrations TRTORCH_UNUSED =
                     EvalOptions().validSchemas({
                         "aten::numel(Tensor self) -> int",
                     })})
-        // .evaluator({c10::Symbol::fromQualString("aten::t"),
-        //             [](const torch::jit::Node* n, kwargs& args) -> c10::optional<torch::jit::IValue> {
-        //               auto tensor_var = args.at(n->input(0));
-        //               if (tensor_var.isIValue() && tensor_var.IValue()->isTensor()) {
-        //                 auto tensor = tensor_var.unwrapToTensor();
-        //                 return tensor.t();
-        //               } else if (tensor_var.isITensor()) {
-        //                 auto input_tensor = tensor_var.ITensor();
-        //                 auto input_dims = input_tensor->getDimensions();
-        //                 LOG_DEBUG("[aten::t] INPUT TENSOR DIMS: " << input_dims);
-        //                 // nvinfer1::Dims transposed_input_dims;
-        //                 // for (int i = input_dims.nbDims - 1; i >= 0; i--) {
-        //                 //   transposed_input_dims.d[i] = input_dims.d[input_dims.nbDims - 1 - i];
-        //                 // }
-        //                 // auto shuffle_layer = ctx->net->addShuffle(*input_tensor);
-        //                 // shuffle_layer->setReshapeDimensions(transposed_input_dims);
-        //                 // shuffle_layer->setZeroIsPlaceholder(true);
-        //                 // auto output_tensor = shuffle_layer->getOutput(0);
-        //                 auto tensor_holder = TensorContainer();
-        //                 tensor_holder.hold_tensor(input_tensor);
-        //                 auto ival = c10::IValue(std::move(c10::make_intrusive<TensorContainer>(tensor_holder)));
-        //                 return ival;
-        //               } else {
-        //                 TRTORCH_THROW_ERROR("Unimplemented data type for aten::t evaluator: ITensor");
-        //                 return {};
-        //               }
-        //             },
-        //             EvalOptions().validSchemas({
-        //                 "aten::t(Tensor self) -> Tensor",
-        //             })})
         .evaluator({c10::Symbol::fromQualString("aten::dim"),
                     [](const torch::jit::Node* n, kwargs& args) -> c10::optional<torch::jit::IValue> {
                       auto tensor_var = args.at(n->input(0));
diff --git a/core/lowering/lowering.cpp b/core/lowering/lowering.cpp
@@ -24,7 +24,7 @@ void LowerBlock(torch::jit::Block* b) {
   DropUnusedNodes(b);
 }
 
-void LowerGraph(std::shared_ptr<torch::jit::Graph>& g) {
+void LowerGraph(std::shared_ptr<torch::jit::Graph>& g, bool disable_cse) {
   passes::UnpackHardSwish(g);
   torch::jit::EliminateRedundantGuards(g);
   torch::jit::RemoveListMutation(g);
@@ -42,9 +42,9 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g) {
   passes::Conv3DToConvolution(g);
   passes::FuseAddMMBranches(g);
   passes::RemoveBNDimCheck(g);
-  LOG_INFO("====PRE CSE =====" << *g);
-  // torch::jit::EliminateCommonSubexpression(g);
-  LOG_INFO("====POST CSE =====" << *g);
+  if (!disable_cse) {
+    torch::jit::EliminateCommonSubexpression(g);
+  }
   // torch::jit::UnrollLoops(g);
   passes::UnpackAddMM(g);
   // passes::UnpackBatchNorm(g);
@@ -57,25 +57,36 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g) {
 }
 
 torch::jit::Module LowerModule(const torch::jit::script::Module& mod) {
+  LOG_DEBUG("Input module is being frozen by torch::jit::freeze_module");
   auto mod_ = torch::jit::freeze_module(mod);
   return mod_;
 }
 
 std::pair<std::shared_ptr<torch::jit::Graph>, std::vector<torch::jit::IValue>> Lower(
     const torch::jit::script::Module& mod,
-    std::string method_name) {
-  auto lowered_mod = mod; // LowerModule(mod);
+    std::string method_name,
+    bool unfreeze_module = false) {
+  auto lowered_mod = unfreeze_module ? mod : LowerModule(mod);
   auto g = lowered_mod.get_method(method_name).graph();
   LOG_GRAPH(*g);
 
   // Go through TRTorch Lowering to reformat graph to be conversion friendly
   // and also segment for accelerators and executors (TRT-DLA, TRT-GPU, PYT)
-  LOG_GRAPH("TRTorch Graph Lowering");
-  // lowering::LowerGraph(g);
+  // unfreeze_module is used to not perform constant folding on weights in the network.
+  // In quantization aware trained (QAT) models, weights are passed through quantize and
+  // dequantize nodes which should not be folded. So unfreeze_module is set to True for QAT models.
+  if (!unfreeze_module) {
+    LOG_GRAPH("TRTorch Graph Lowering");
+    lowering::LowerGraph(g, false);
+  }
 
   LOG_GRAPH("LibTorch Lowering");
   auto graph_and_ivalues = torch::jit::LowerGraph(*g, lowered_mod._ivalue());
-  lowering::LowerGraph(graph_and_ivalues.first);
+
+  if (unfreeze_module) {
+    LOG_GRAPH("TRTorch Graph Lowering");
+    lowering::LowerGraph(graph_and_ivalues.first, true);
+  }
   // Is this necessary?
   lowering::LowerBlock(g->block());
 
diff --git a/core/lowering/lowering.h b/core/lowering/lowering.h
@@ -7,11 +7,12 @@ namespace core {
 namespace lowering {
 
 void LowerBlock(torch::jit::Block* b);
-void LowerGraph(std::shared_ptr<torch::jit::Graph>& g);
+void LowerGraph(std::shared_ptr<torch::jit::Graph>& g, bool disable_cse /*=false*/);
 torch::jit::Module LowerModule(const torch::jit::script::Module& mod);
 std::pair<std::shared_ptr<torch::jit::Graph>, std::vector<torch::jit::IValue>> Lower(
     const torch::jit::script::Module& mod,
-    std::string method_name);
+    std::string method_name,
+    bool unfreeze_module /*=false*/);
 
 } // namespace lowering
 } // namespace core
diff --git a/cpp/api/include/trtorch/trtorch.h b/cpp/api/include/trtorch/trtorch.h
@@ -262,9 +262,9 @@ struct TRTORCH_API CompileSpec {
    * Emum for selecting engine capability
    */
   enum class EngineCapability : int8_t {
-    kDEFAULT,
-    kSAFE_GPU,
-    kSAFE_DLA,
+    kSTANDARD,
+    kSAFETY,
+    kDLA_STANDALONE,
   };
 
   class TRTORCH_API TensorFormat {
@@ -686,12 +686,12 @@ struct TRTORCH_API CompileSpec {
    * This is the behavior of FP32 layers by default.
    */
   bool disable_tf32 = false;
-  
-   /**
+
+  /**
    * Enable sparsity for weights of conv and FC layers
    */
   bool sparse_weights = false;
-  
+
   /**
    * Build a refitable engine
    */
diff --git a/cpp/api/src/compile_spec.cpp b/cpp/api/src/compile_spec.cpp
@@ -405,7 +405,13 @@ core::CompileSpec to_internal_compile_spec(CompileSpec external) {
 
   if (internal.convert_info.engine_settings.enabled_precisions.find(nvinfer1::DataType::kINT8) !=
       internal.convert_info.engine_settings.enabled_precisions.end()) {
-    internal.convert_info.engine_settings.calibrator = external.ptq_calibrator;
+    if (external.ptq_calibrator) {
+      internal.convert_info.engine_settings.calibrator = external.ptq_calibrator;
+    } else {
+      ;
+      internal.convert_info.engine_settings.unfreeze_module = true;
+      internal.convert_info.engine_settings.calibrator = nullptr;
+    }
   } else {
     internal.convert_info.engine_settings.calibrator = nullptr;
   }
diff --git a/py/trtorch/csrc/register_tensorrt_classes.cpp b/py/trtorch/csrc/register_tensorrt_classes.cpp
@@ -47,7 +47,7 @@ void RegisterTRTCompileSpec() {
           .def("_set_torch_fallback", &trtorch::pyapi::CompileSpec::setTorchFallbackIntrusive)
           .def("_set_ptq_calibrator", &trtorch::pyapi::CompileSpec::setPTQCalibratorViaHandle)
           .def("__str__", &trtorch::pyapi::CompileSpec::stringify);
-  
+
   ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, sparse_weights);
   ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, disable_tf32);
   ADD_FIELD_GET_SET_REGISTRATION(TRTCompileSpecTSRegistration, trtorch::pyapi::CompileSpec, refit);
diff --git a/py/trtorch/csrc/tensorrt_backend.cpp b/py/trtorch/csrc/tensorrt_backend.cpp
@@ -32,7 +32,7 @@ c10::impl::GenericDict TensorRTBackend::compile(c10::IValue mod_val, c10::impl::
     const auto& method_name = it->key();
     auto method = mod.get_method(method_name);
     auto graph = method.graph();
-    core::lowering::LowerGraph(graph);
+    core::lowering::LowerGraph(graph, false);
   }
 
   auto handles = c10::impl::GenericDict(
diff --git a/py/trtorch/csrc/tensorrt_classes.cpp b/py/trtorch/csrc/tensorrt_classes.cpp
@@ -181,8 +181,15 @@ core::CompileSpec CompileSpec::toInternalCompileSpec() {
   for (auto p : enabled_precisions) {
     info.convert_info.engine_settings.enabled_precisions.insert(toTRTDataType(p));
   }
-
-  info.convert_info.engine_settings.calibrator = ptq_calibrator;
+  if (ptq_calibrator) {
+    info.convert_info.engine_settings.calibrator = ptq_calibrator;
+  } else {
+    if (info.convert_info.engine_settings.enabled_precisions.find(nvinfer1::DataType::kINT8) !=
+        info.convert_info.engine_settings.enabled_precisions.end()) {
+      std::cout << "===INTERNAL UNFREEZE MODULE TRUE===" << std::endl;
+      info.convert_info.engine_settings.unfreeze_module = true;
+    }
+  }
   info.convert_info.engine_settings.sparse_weights = sparse_weights;
   info.convert_info.engine_settings.disable_tf32 = disable_tf32;
   info.convert_info.engine_settings.refit = refit;
diff --git a/tests/core/conversion/converters/test_shuffle.cpp b/tests/core/conversion/converters/test_shuffle.cpp
@@ -241,6 +241,30 @@ TEST(Converters, ATenTransposeConvertsCorrectly) {
   ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[0], trt, 2e-6));
 }
 
+TEST(Converters, ATenTConvertsCorrectly) {
+  const auto graph = R"IR(
+    graph(%x.1 : Tensor):
+      %out : Tensor = aten::t(%x.1)
+      return (%out))IR";
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(graph, &*g);
+
+  auto in = at::randint(0, 5, {3, 4}, {at::kCUDA});
+  auto params = trtorch::core::conversion::get_named_params(g->inputs(), {});
+
+  std::cout << "Running JIT" << std::endl;
+  auto jit_results = trtorch::tests::util::RunGraph(g, params, {in});
+
+  std::cout << "Running TRT" << std::endl;
+  in = at::clone(in);
+  params = trtorch::core::conversion::get_named_params(g->inputs(), {});
+  auto trt_results = trtorch::tests::util::RunGraphEngine(g, params, {in});
+  auto trt = trt_results[0].reshape_as(jit_results[0]);
+
+  ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[0], trt, 2e-6));
+}
+
 TEST(Converters, ATenTransposeNegativeConvertsCorrectly) {
   const auto graph = R"IR(
     graph(%x.1 : Tensor):
@@ -312,7 +336,6 @@ TEST(Converters, ATenPixelShuffle3DConvertsCorrectly) {
   in = at::clone(in);
   params = trtorch::core::conversion::get_named_params(g->inputs(), {});
   auto trt_results = trtorch::tests::util::RunGraphEngine(g, params, {in});
-  // auto trt = trt_results[0].reshape_as(jit_results[0]);
 
   ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
 }
diff --git a/tests/modules/hub.py b/tests/modules/hub.py
@@ -54,18 +54,10 @@
         "model": torch.hub.load('pytorch/vision:v0.9.0', 'resnet50', pretrained=True),
         "path": "both"
     },
-    "fcn_resnet101": {
-        "model": torch.hub.load('pytorch/vision:v0.9.0', 'fcn_resnet101', pretrained=True),
-        "path": "script"
-    },
     "ssd": {
         "model": torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_ssd', model_math="fp32"),
         "path": "trace"
     },
-    "faster_rcnn": {
-        "model": models.detection.fasterrcnn_resnet50_fpn(pretrained=True),
-        "path": "script"
-    },
     "efficientnet_b0": {
         "model": timm.create_model('efficientnet_b0', pretrained=True),
         "path": "script"
diff --git a/tests/util/run_graph_engine.cpp b/tests/util/run_graph_engine.cpp
@@ -69,7 +69,7 @@ std::vector<at::Tensor> RunGraphEngine(
   auto in = toInputs(inputs);
   auto info = core::conversion::ConversionInfo(in);
   info.engine_settings.workspace_size = 1 << 20;
-  info.engine_settings.op_precision = op_precision;
+  info.engine_settings.enabled_precisions.insert(op_precision);
   std::string eng = core::conversion::ConvertBlockToEngine(g->block(), info, named_params);
   return RunEngine(eng, inputs);
 }

Original file line number	Diff line number	Diff line change
`@@ -72,7 +72,7 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)`
`72`	`72`	`if (!settings.calibrator) {`
`73`	`73`	`LOG_WARNING(`
`74`	`74`	`"Int8 precision has been enabled but no calibrator provided. This assumes the network has Q/DQ nodes obtained from Quantization aware training. For more details, refer to https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#work-with-qat-networks");`
`75`		`- } else{`
	`75`	`+ } else {`
`76`	`76`	`cfg->setInt8Calibrator(settings.calibrator);`
`77`	`77`	`}`
`78`	`78`	`break;`