feat(//core/conversion/var): created ITensorOrFreeze() method, to rep…

…lace functionality of Var::ITensor() Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): updates to some comments on the PR Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): addressed PR comment Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): addressed PR comments Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): addressed PR comments Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): addressing PR comments Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): addressing PR comments Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): Addressed PR comments Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> fix(): bug in test_serialization, need to fix Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> Delete converters.h.orig delete .orig file Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Update activation.cpp addressing PR comments Signed-off-by: Abhiram Iyer <abhi.iyer.ai@gmail.com> Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Delete converters.h.orig delete .orig file Signed-off-by: Abhiram Iyer <abhirami@nvidia.com> Update activation.cpp addressing PR comments Signed-off-by: Abhiram Iyer <abhirami@nvidia.com>
pytorch · Jul 28, 2020 · 2ccf8d0 · 2ccf8d0
1 parent 362c932
commit 2ccf8d0
Show file tree

Hide file tree

Showing 17 changed files with 167 additions and 95 deletions.
diff --git a/core/conversion/converters/BUILD b/core/conversion/converters/BUILD
@@ -7,14 +7,32 @@ config_setting(
     }
 )
 
+cc_library(
+    name = "weights",
+    hdrs = [
+        "Weights.h"
+    ],
+    srcs = [
+        "Weights.cpp"
+    ],
+    deps = [
+        "@tensorrt//:nvinfer",
+        "//core/util:prelude",
+        "//core/conversion/conversionctx"
+    ] + select({
+        ":use_pre_cxx11_abi":  ["@libtorch_pre_cxx11_abi//:libtorch"],
+        "//conditions:default":  ["@libtorch//:libtorch"],
+    }),
+    alwayslink = True,
+)
+
 cc_library(
     name = "converters",
     hdrs = [
         "converters.h"
     ],
     srcs = [
         "NodeConverterRegistry.cpp",
-        "Weights.cpp",
         "impl/activation.cpp",
         "impl/batch_norm.cpp",
         "impl/concat.cpp",
@@ -51,5 +69,8 @@ load("@rules_pkg//:pkg.bzl", "pkg_tar")
 pkg_tar(
     name = "include",
     package_dir = "core/conversion/converters/",
-    srcs = ["converters.h"],
+    srcs = [
+        "converters.h",
+        "Weights.h"
+    ],
 )
diff --git a/core/conversion/converters/Weights.cpp b/core/conversion/converters/Weights.cpp
@@ -1,12 +1,11 @@
 #include "core/util/prelude.h"
-#include "core/conversion/converters/converters.h"
+#include "core/conversion/converters/Weights.h"
 
 namespace trtorch {
 namespace core {
 namespace conversion {
 namespace converters {
 
-
 Weights::Weights() {
     this->num_input_maps = 0;
     this->num_output_maps = 0;
@@ -18,20 +17,36 @@ Weights::Weights() {
 Weights::Weights(ConversionCtx* ctx, float val) {
     this->num_input_maps = 1;
     this->num_output_maps = 1;
+
     this->data.type = nvinfer1::DataType::kFLOAT;
     float* buf = reinterpret_cast<float*>(malloc(1 * sizeof(float)));
     buf[0] = val;
     this->data.values = buf;
     this->data.count = 1;
     ctx->builder_resources.push_back(buf);
-    this->kernel_shape.nbDims = 1;
-    this->kernel_shape.d[0] = 1;
+
+    this->shape.nbDims = 0;
+    this->kernel_shape.nbDims = 0;
+}
+
+Weights::Weights(ConversionCtx* ctx, int32_t val) {
+    this->num_input_maps = 1;
+    this->num_output_maps = 1;
+
+    this->data.type = nvinfer1::DataType::kINT32;
+    int32_t* buf = reinterpret_cast<int32_t*>(malloc(1 * sizeof(int32_t)));
+    buf[0] = val;
+    this->data.values = buf;
+    this->data.count = 1;
+    ctx->builder_resources.push_back(buf);
+
+    this->shape.nbDims = 0;
+    this->kernel_shape.nbDims = 0;
 }
 
 Weights::Weights(ConversionCtx* ctx, at::Tensor t) {
     if (t.sizes().size() > nvinfer1::Dims::MAX_DIMS) {
-        //TODO: Handle this with exceptions or whatever
-        LOG_INTERNAL_ERROR("The tensor requested to be converted to nvinfer1::Weights exceeds the max number of dimensions for TensorRT");
+        TRTORCH_THROW_ERROR("The tensor requested to be converted to nvinfer1::Weights exceeds the max number of dimensions for TensorRT");
     }
     this->shape = util::toDims(t.sizes());
     if (t.sizes().size() >= 2) {
@@ -59,9 +74,7 @@ Weights::Weights(ConversionCtx* ctx, at::Tensor t) {
     t_cpu = t_cpu.contiguous();
     auto dtype_optional = util::toTRTDataType(t_cpu.dtype());
     if (!dtype_optional) {
-        //TODO: Handle this with exceptions or whatever
-        //TODO: Implement handling for the Torch Types
-       LOG_INTERNAL_ERROR("The tensor requested to be converted to nvinfer1::Weights is of an unsupported type");
+       TRTORCH_THROW_ERROR("The tensor requested to be converted to nvinfer1::Weights is of an unsupported type");
     }
 
     // Store the data in the conversion context so it remains until building is complete

diff --git a/core/conversion/converters/Weights.h b/core/conversion/converters/Weights.h
@@ -0,0 +1,45 @@
+#pragma once
+
+#include "core/util/prelude.h"
+#include "core/conversion/conversionctx/ConversionCtx.h"
+
+namespace trtorch {
+namespace core {
+namespace conversion {
+namespace converters {
+
+struct Weights {
+    nvinfer1::Weights data;
+    nvinfer1::Dims kernel_shape;
+    nvinfer1::Dims shape;
+    int64_t num_input_maps;
+    int64_t num_output_maps;
+
+    Weights();
+    Weights(ConversionCtx* ctx, at::Tensor t);
+    Weights(ConversionCtx* ctx, float val);
+    Weights(ConversionCtx* ctx, int32_t val);
+    friend std::ostream& operator<<(std::ostream& os, const Weights& w);
+};
+
+inline nvinfer1::ITensor* tensor_to_const(ConversionCtx* ctx, at::Tensor t) {
+    auto t_weights = Weights(ctx, t);
+    auto const_layer = ctx->net->addConstant(t_weights.shape, t_weights.data);
+    TRTORCH_CHECK(const_layer, "Unable to freeze tensor");
+
+    auto out = const_layer->getOutput(0);
+
+    std::ostringstream tensor_id;
+    tensor_id << reinterpret_cast<int*>(out);
+
+    LOG_DEBUG(ctx->logger, "Freezing tensor " << tensor_id.str() << " as an IConstantLayer");
+    const_layer->setName(("[Freeze Tensor " + tensor_id.str() + " ]").c_str());
+
+    return out;
+}
+
+
+} // namespace converters
+} // namespace conversion
+} // namespace core
+} // namespace trtorch
diff --git a/core/conversion/converters/converters.h b/core/conversion/converters/converters.h
@@ -9,6 +9,7 @@
 #include "core/util/prelude.h"
 #include "core/conversion/var/Var.h"
 #include "core/conversion/conversionctx/ConversionCtx.h"
+#include "core/conversion/converters/Weights.h"
 
 namespace trtorch {
 namespace core {
@@ -39,28 +40,6 @@ class RegisterNodeConversionPatterns {
 bool node_is_convertable(const torch::jit::Node* n);
 OpConverter get_node_converter_for(const torch::jit::FunctionSchema* signature);
 
-struct Weights {
-    //TODO: Rebuild this in a way that makes sense for more than just conv2/3D and linear
-    nvinfer1::Weights data;
-    nvinfer1::Dims kernel_shape;
-    nvinfer1::Dims shape;
-    int64_t num_input_maps;
-    int64_t num_output_maps;
-
-    Weights();
-    Weights(ConversionCtx* ctx, at::Tensor t);
-    Weights(ConversionCtx* ctx, float val);
-    friend std::ostream& operator<<(std::ostream& os, const Weights& w);
-};
-
-inline nvinfer1::ITensor* tensor_to_const(ConversionCtx* ctx, at::Tensor t) {
-    auto t_weights = Weights(ctx, t);
-    auto const_layer = ctx->net->addConstant(t_weights.shape, t_weights.data);
-    TRTORCH_CHECK(const_layer, "Unable to freeze tensor");
-    const_layer->setName("[Freeze Tensor]");
-    return const_layer->getOutput(0);
-}
-
 } // namespace converters
 } // namespace conversion
 } // namespace core

diff --git a/core/conversion/converters/impl/activation.cpp b/core/conversion/converters/impl/activation.cpp
@@ -10,7 +10,7 @@ namespace {
 
 #define convert(act, trt_type)                                                 \
   bool act(ConversionCtx* ctx, const torch::jit::Node* n, args& args) {        \
-    auto in = args[0].ITensor();                                               \
+    auto in = args[0].ITensorOrFreeze(ctx);                                    \
                                                                                \
     auto new_layer =                                                           \
         ctx->net->addActivation(*in, nvinfer1::ActivationType::trt_type);      \
@@ -46,7 +46,7 @@ auto acthardtanh TRTORCH_UNUSED = RegisterNodeConversionPatterns()
   .pattern({
     "aten::hardtanh(Tensor self, Scalar min_val=-1, Scalar max_val=1) -> (Tensor)",
     [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
-      auto in = args[0].ITensor();
+      auto in = args[0].ITensorOrFreeze(ctx);
       auto min = args[1].unwrapToDouble();
       auto max = args[2].unwrapToDouble();
 
@@ -66,7 +66,7 @@ auto acthardtanh TRTORCH_UNUSED = RegisterNodeConversionPatterns()
     //TODO: Remove after functionalization
     "aten::hardtanh_(Tensor(a!) self, Scalar min_val=-1, Scalar max_val=1) -> (Tensor(a!))",
     [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
-      auto in = args[0].ITensor();
+      auto in = args[0].ITensorOrFreeze(ctx);
       auto min = args[1].unwrapToDouble();
       auto max = args[2].unwrapToDouble();
 

diff --git a/core/conversion/converters/impl/batch_norm.cpp b/core/conversion/converters/impl/batch_norm.cpp
@@ -15,7 +15,7 @@ auto batch_norm_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
                             Tensor? mean, Tensor? var,
                             bool training, float momentum, float eps, bool cudnn_enabled) -> (Tensor))SIG",
     [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
-      auto input = args[0].ITensor();
+      auto input = args[0].ITensor(); // assumes non-static input Tensor
       auto orig_shape = input->getDimensions();
       auto shape = util::toVec(orig_shape);
       auto options = torch::TensorOptions().dtype(torch::kFloat32);

diff --git a/core/conversion/converters/impl/conv_deconv.cpp b/core/conversion/converters/impl/conv_deconv.cpp
@@ -17,7 +17,7 @@ auto conv_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
                                  int[] output_padding, int groups, bool benchmark,
                                  bool deterministic, bool cudnn_enabled) -> (Tensor))SIG",
         [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
-            auto in = args[0].ITensor();
+            auto in = args[0].ITensor(); // assumes non-static input Tensor
 
             auto w = Weights(ctx, args[1].unwrapToTensor());
             auto stride = util::toDims(args[3].unwrapToIntList());

diff --git a/core/conversion/converters/impl/element_wise.cpp b/core/conversion/converters/impl/element_wise.cpp
@@ -26,7 +26,6 @@ nvinfer1::ILayer* add_elementwise(ConversionCtx* ctx, nvinfer1::ElementWiseOpera
         self = self_shuffle->getOutput(0);
     }
 
-
     nvinfer1::ILayer* ele;
     if (scalar != 1) {
         LOG_WARNING("Please verify scalar handling in add converter, channel axis set to 3 but scaling is uniform");
@@ -73,8 +72,8 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
             "aten::add.Tensor(Tensor self, Tensor other, Scalar alpha=1) -> Tensor",
             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                 // Should implement self + alpha * other
-                auto self = args[0].ITensor();
-                auto other = args[1].ITensor();
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto other = args[1].ITensorOrFreeze(ctx);
                 auto scalar = args[2].unwrapToScalar().to<float>();
                 auto add = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUM, self, other, util::node_info(n), scalar);
 
@@ -90,8 +89,8 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
             "aten::add_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> (Tensor(a!))",
             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                 // Should implement self + alpha * other
-                auto self = args[0].ITensor();
-                auto other = args[1].ITensor();
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto other = args[1].ITensorOrFreeze(ctx);
                 auto scalar = args[2].unwrapToScalar().to<float>();
                 auto add = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUM, self, other, util::node_info(n), scalar);
 
@@ -107,8 +106,8 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
             "aten::sub.Tensor(Tensor self, Tensor other, Scalar alpha=1) -> Tensor",
             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                 // Should implement self - alpha * other
-                auto self = args[0].ITensor();
-                auto other = args[1].ITensor();
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto other = args[1].ITensorOrFreeze(ctx);
                 auto scalar = args[2].unwrapToScalar().to<float>();
                 auto sub = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kSUB, self, other, util::node_info(n), scalar);
 
@@ -124,8 +123,8 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
             "aten::div.Tensor(Tensor self, Tensor other) -> Tensor",
             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                 // Should implement self / other
-                auto self = args[0].ITensor();
-                auto other = args[1].ITensor();
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto other = args[1].ITensorOrFreeze(ctx);
                 auto div = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kDIV, self, other, util::node_info(n));
 
                 TRTORCH_CHECK(div, "Unable to create div layer from node: " << *n);
@@ -140,8 +139,8 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
             "aten::div_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)",
             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                 // TODO: Remove with functionalization
-                auto self = args[0].ITensor();
-                auto other = args[1].ITensor();
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto other = args[1].ITensorOrFreeze(ctx);
                 auto div = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kDIV, self, other, util::node_info(n));
 
                 TRTORCH_CHECK(div, "Unable to create div layer from node: " << *n);
@@ -156,8 +155,8 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
             "aten::mul.Tensor(Tensor self, Tensor other) -> Tensor",
             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                 // Should implement self * other
-                auto self = args[0].ITensor();
-                auto other = args[1].ITensor();
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto other = args[1].ITensorOrFreeze(ctx);
                 auto mul = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kPROD, self, other, util::node_info(n));
 
                 TRTORCH_CHECK(mul, "Unable to create mul layer from node: " << *n);
@@ -172,8 +171,8 @@ auto element_wise_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns(
             "aten::mul_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)",
             [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
                 // TODO: Remove with functionalization
-                auto self = args[0].ITensor();
-                auto other = args[1].ITensor();
+                auto self = args[0].ITensorOrFreeze(ctx);
+                auto other = args[1].ITensorOrFreeze(ctx);
                 auto mul = add_elementwise(ctx, nvinfer1::ElementWiseOperation::kPROD, self, other, util::node_info(n));
 
                 TRTORCH_CHECK(mul, "Unable to create mul layer from node: " << *n);

diff --git a/core/conversion/converters/impl/linear.cpp b/core/conversion/converters/impl/linear.cpp
@@ -14,7 +14,7 @@ auto linear_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
         [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
             // PyTorch follows in: Nx*xIN, W: OUTxIN, B: OUT, out: Nx*xOUT
             // TensorRT inserts a flatten in when following conv
-            auto in = args[0].ITensor();
+            auto in = args[0].ITensorOrFreeze(ctx);
             auto shape = util::toVec(in->getDimensions());
 
             LOG_DEBUG("Input tensor shape: " << in->getDimensions());

diff --git a/core/conversion/converters/impl/matrix_multiply.cpp b/core/conversion/converters/impl/matrix_multiply.cpp
@@ -12,30 +12,10 @@ auto mm_registrations TRTORCH_UNUSED = RegisterNodeConversionPatterns()
   .pattern({
     "aten::matmul(Tensor self, Tensor other) -> (Tensor)",
     [](ConversionCtx* ctx, const torch::jit::Node* n, args& args) -> bool {
-      nvinfer1::ITensor* self;
-      if (args[0].isIValue()) {
-        auto t = args[0].unwrapToTensor();
-        auto t_weights = Weights(ctx, t);
-        auto const_layer = ctx->net->addConstant(t_weights.shape, t_weights.data);
-        TRTORCH_CHECK(const_layer, "Unable to freeze tensor self for node: " << *n);
-        const_layer->setName((util::node_info(n) + " [Freeze Tensor(self)]").c_str());
-        self = const_layer->getOutput(0);
-      } else {
-        self = args[0].ITensor();
-      }
+      auto self = args[0].ITensorOrFreeze(ctx);
       LOG_DEBUG("self tensor shape: " << self->getDimensions());
 
-      nvinfer1::ITensor* other;
-      if (args[1].isIValue()) {
-        auto t = args[1].unwrapToTensor();
-        auto t_weights = Weights(ctx, t);
-        auto const_layer = ctx->net->addConstant(t_weights.shape, t_weights.data);
-        TRTORCH_CHECK(const_layer, "Unable to freeze tensor other for node: " << *n);
-        const_layer->setName((util::node_info(n) + " [Freeze Tensor(other)]").c_str());
-        other = const_layer->getOutput(0);
-      } else {
-        other = args[1].ITensor();
-      }
+      auto other = args[1].ITensorOrFreeze(ctx);
       LOG_DEBUG("other tensor shape: " << other->getDimensions());
 
       auto mm_layer = ctx->net->addMatrixMultiply(*self, nvinfer1::MatrixOperation::kNONE, *other, nvinfer1::MatrixOperation::kNONE);