feat(//core/converters): Add power layer conversion support and minor…

… README edits Signed-off-by: Dheeraj Peri <peri.dheeraj@gmail.com>
pytorch · Oct 5, 2020 · a801506 · a801506
1 parent e4a4574
commit a801506
Show file tree

Hide file tree

Showing 4 changed files with 88 additions and 19 deletions.
diff --git a/README.md b/README.md
@@ -115,10 +115,10 @@ then you have two options.
 1. You need to download the tarball distributions of TensorRT and cuDNN from the NVIDIA website.
     - https://developer.nvidia.com/cudnn
     - https://developer.nvidia.com/tensorrt
-2. Place these files in a directory (the directories `third_party/distdir/[x86_64-linux-gnu | aarch64-linux-gnu]` exist for this purpose)
+2. Place these files in a directory (the directories `third_party/dist_dir/[x86_64-linux-gnu | aarch64-linux-gnu]` exist for this purpose)
 3. Compile using:
 ``` shell
-bazel build //:libtrtorch --compilation_mode opt --distdir third_party/distdir/[x86_64-linux-gnu | aarch64-linux-gnu]
+bazel build //:libtrtorch --compilation_mode opt --distdir third_party/dist_dir/[x86_64-linux-gnu | aarch64-linux-gnu]
 ```
 
 #### 2. Building using locally installed cuDNN & TensorRT
@@ -175,7 +175,7 @@ bazel build //:libtrtorch --compilation_mode=dbg
 
 ### Native compilation on NVIDIA Jetson AGX
 ``` shell
-bazel build //:libtrtorch --distdir third_party/distdir/aarch64-linux-gnu
+bazel build //:libtrtorch --distdir third_party/dist_dir/aarch64-linux-gnu
 ```
 > Note: Please refer [installation](docs/tutorials/installation.html) instructions for Pre-requisites
 

diff --git a/core/conversion/converters/impl/element_wise.cpp b/core/conversion/converters/impl/element_wise.cpp
@@ -1,3 +1,4 @@
+#include <torch/torch.h>
 #include "core/util/prelude.h"
 #include "core/conversion/converters/converters.h"
 
@@ -180,6 +181,50 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
                 mul->setName(util::node_info(n).c_str());
                 auto out = ctx->AssociateValueAndTensor(n->outputs()[0], mul->getOutput(0));
 
+                LOG_DEBUG("Output tensor shape: " << out->getDimensions());
+                return true;
+             }
+         }).pattern({
+            "aten::pow.Tensor_Tensor(Tensor self, Tensor exponent) -> (Tensor)",
+            [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+                // TODO: Remove with functionalization
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto exponent = args[1].ITensorOrFreeze(ctx);
+                auto pow = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kPOW, self, exponent, util::node_info(n));
+                TRTORCH_CHECK(pow, "Unable to create Power layer from node: " << *n);
+
+                pow->setName(util::node_info(n).c_str());
+                auto out = ctx->AssociateValueAndTensor(n->outputs()[0], pow->getOutput(0));
+
+                LOG_DEBUG("Output tensor shape: " << out->getDimensions());
+                return true;
+             }
+         }).pattern({
+            "aten::pow.Tensor_Scalar(Tensor self, Scalar exponent) -> (Tensor)",
+            [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto exponentScalar = args[1].unwrapToScalar().to<float>();
+
+                // Calculate size of the input and define an exponent tensor of the same size
+                int volume = 1;
+                for (int i = 0; i < self->getDimensions().nbDims; i++) {
+                    volume = volume * (self->getDimensions().d[i]);
+                }
+
+                // Create a torch tensor with constant exponent values
+                LOG_DEBUG("Broadcasting the exponent in power layer");
+                torch::Tensor exponentBlob = torch::full({volume}, exponentScalar);
+
+                // Create a corresponding constant layer in TRT and get the layer output.
+                auto weights = converters::Weights(ctx, exponentBlob);
+                auto exponentTensor = ctx->net->addConstant(self->getDimensions(), weights.data)->getOutput(0);
+
+                auto pow = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kPOW, self, exponentTensor, util::node_info(n));
+                TRTORCH_CHECK(pow, "Unable to create Power layer from node: " << *n);
+
+                pow->setName(util::node_info(n).c_str());
+                auto out = ctx->AssociateValueAndTensor(n->outputs()[0], pow->getOutput(0));
+
                 LOG_DEBUG("Output tensor shape: " << out->getDimensions());
                 return true;
              }

diff --git a/py/README.md b/py/README.md
@@ -22,7 +22,7 @@ traced_model = torch.jit.trace(model, [data])
 
 # Compile module
 compiled_trt_model = trtorch.compile(model, {
-    "input_shape": [data.shape],
+    "input_shapes": [data.shape],
     "op_precision": torch.half, # Run in FP16
 })
 

diff --git a/tests/core/converters/test_element_wise.cpp b/tests/core/converters/test_element_wise.cpp
@@ -4,32 +4,39 @@
 #include "tests/util/util.h"
 #include "core/compiler.h"
 
-void pointwise_test_helper(std::string graph_ir) {
+void pointwise_test_helper(std::string graph_ir, bool singleInput) {
     auto g = std::make_shared<torch::jit::Graph>();
     torch::jit::parseIR(graph_ir, &*g);
-
-    auto in0 = at::randint(1, 5, {5}, {at::kCUDA});
-    auto in1 = at::randint(1, 5, {5}, {at::kCUDA});
+
+    // singleInput case is enabled when elementwise operation is performed 
+    // with an input and a constant embedded in graph
+    std::vector<at::Tensor> torch_inputs;
+    torch_inputs.push_back(at::randint(1, 5, {5}, {at::kCUDA}));
+    if (!singleInput) {
+        torch_inputs.push_back(at::randint(1, 5, {5}, {at::kCUDA}));
+    }
     auto params = trtorch::core::conversion::get_named_params(g->inputs(), {});
-    auto jit_results = trtorch::tests::util::RunGraph(g, params, {in0, in1});
+    auto jit_results = trtorch::tests::util::RunGraph(g, params, torch_inputs);
+
+    std::vector<at::Tensor> trt_inputs;
+    for (auto in : torch_inputs) {
+        trt_inputs.push_back(at::clone(in));
+    }
 
-    in0 = at::clone(in0);
-    in1 = at::clone(in1);
     params = trtorch::core::conversion::get_named_params(g->inputs(), {});
-    auto trt_results = trtorch::tests::util::RunGraphEngine(g, params, {in0, in1});
+    auto trt_results = trtorch::tests::util::RunGraphEngine(g, params, trt_inputs);
 
     ASSERT_TRUE(trtorch::tests::util::almostEqual(jit_results[0], trt_results[0], 2e-6));
 }
 
 
-
 TEST(Converters, ATenAddConvertsCorrectly) {
     const auto graph = R"IR(
       graph(%0 : Tensor, %1 : Tensor):
         %2 : int = prim::Constant[value=1]()
         %3 : Tensor = aten::add(%0, %1, %2)
         return (%3))IR";
-    pointwise_test_helper(graph);
+    pointwise_test_helper(graph, false);
 }
 
 
@@ -39,7 +46,7 @@ TEST(Converters, ATenAddConvertsCorrectly) {
 //         %2 : int = prim::Constant[value=2]()
 //         %3 : Tensor = aten::add(%0, %1, %2)
 //         return (%3))IR";
-//     pointwise_test_helper(graph);
+//     pointwise_test_helper(graph, false);
 // }
 
 TEST(Converters, ATenSubConvertsCorrectly) {
@@ -48,7 +55,7 @@ TEST(Converters, ATenSubConvertsCorrectly) {
         %2 : int = prim::Constant[value=1]()
         %3 : Tensor = aten::sub(%0, %1, %2)
         return (%3))IR";
-    pointwise_test_helper(graph);
+    pointwise_test_helper(graph, false);
 }
 
 // TEST(Converters, ATenSubWithScaleConvertsCorrectly) {
@@ -57,21 +64,38 @@ TEST(Converters, ATenSubConvertsCorrectly) {
 //         %2 : float = prim::Constant[value=0.5]()
 //         %3 : Tensor = aten::add(%0, %1, %2)
 //         return (%3))IR";
-//     pointwise_test_helper(graph);
+//     pointwise_test_helper(graph, false);
 // }
 
 TEST(Converters, ATenMulConvertsCorrectly) {
     const auto graph = R"IR(
       graph(%0 : Tensor, %1 : Tensor):
         %2 : Tensor = aten::mul(%0, %1)
         return (%2))IR";
-    pointwise_test_helper(graph);
+    pointwise_test_helper(graph, false);
 }
 
 TEST(Converters, ATenDivConvertsCorrectly) {
     const auto graph = R"IR(
       graph(%0 : Tensor, %1 : Tensor):
         %2 : Tensor = aten::div(%0, %1)
         return (%2))IR";
-    pointwise_test_helper(graph);
+    pointwise_test_helper(graph, false);
+}
+
+TEST(Converters, ATenPowTensorConvertsCorrectly) {
+    const auto graph = R"IR(
+       graph(%x.1 : Tensor, %x2.1 : Tensor):
+          %3 : Tensor = aten::pow(%x.1, %x2.1)
+          return (%3))IR";
+    pointwise_test_helper(graph, false);
+}
+
+TEST(Converters, ATenPowScalarConvertsCorrectly) {
+    const auto graph = R"IR(
+        graph(%x.1 : Tensor):
+          %2 : int = prim::Constant[value=2]()
+          %3 : Tensor = aten::pow(%x.1, %2)
+          return (%3))IR";
+    pointwise_test_helper(graph, true);
 }