Merge pull request #67 from JDAI-CV/update_script_add_new_ops

[WIP] Make most boilerplate code auto-generated & refactor
JDAI-CV · Jan 28, 2020 · ad11c32 · ad11c32
2 parents a5d5627 + 2980add
commit ad11c32
Show file tree

Hide file tree

Showing 30 changed files with 9,009 additions and 2,654 deletions.
diff --git a/.daq_pm/configs/infer b/.daq_pm/configs/infer
@@ -1,6 +1,6 @@
 name DNNLibrary
 type cpp
 target dnn_retrieve_result
-build_dir build_infer
+build_dir build
 cmake_options -DCMAKE_SYSTEM_NAME=Android -DCMAKE_TOOLCHAIN_FILE=~/Android/Sdk/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_CPP_FEATURES=exceptions -DANDROID_PLATFORM=android-28 -DANDROID_ABI=arm64-v8a -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
 program_arguments ~/adb_push_and_run.sh binaries/dnn_retrieve_result
diff --git a/.gitignore b/.gitignore
@@ -17,3 +17,6 @@ dist/
 onnx2daq.egg-info/
 __pycache__/
 .eggs/
+
+.clangd/
+venv/
diff --git a/binaries/ex_model_builder.cpp b/binaries/ex_model_builder.cpp
@@ -1,14 +1,14 @@
 /**
  * It is an example showing how to use the ModelBuilder API to build an model
  */
-#include <chrono>
-#include <iostream>
-#include <vector>
-
 #include <common/helper.h>
 #include <dnnlibrary/ModelBuilder.h>
 #include <glog/logging.h>
 
+#include <chrono>
+#include <iostream>
+#include <vector>
+
 using namespace android::nn::wrapper;
 using dnn::ModelBuilder;
 
@@ -26,23 +26,24 @@ int main() {
             {Type::TENSOR_QUANT8_ASYMM, {3, 1, 1, 3}, 0.1, 150});
         builder.AddTensorFromBuffer("bias", bias_buf,
                                     {Type::TENSOR_INT32, {3}, 0.1, 0});
-        builder.AddDepthWiseConv("data", 1, 1, 0, 0, 0, 0,
-                                 ModelBuilder::ACTIVATION_NONE, 1, "weight",
-                                 "bias", "conv_fwd",
-                                 std::make_optional<ModelBuilder::QuantInfo>(
-                                     {Type::TENSOR_QUANT8_ASYMM, {0.5}, 100}));
-        builder.AddReLU("conv_fwd", "relu_fwd");
-        builder.AddOperationAdd("data", "relu_fwd", "output",
-                                std::make_optional<ModelBuilder::QuantInfo>(
-                                    {Type::TENSOR_QUANT8_ASYMM, {0.05}, 100}));
+        builder.AddLayer_DEPTHWISE_CONV_2D(
+            "data", "weight", "bias", 1, 1, 0, 0, 0, 0, 1, dnn::FuseCode::NONE,
+            "conv_fwd",
+            std::make_optional<ModelBuilder::QuantInfo>(
+                {Type::TENSOR_QUANT8_ASYMM, {0.5}, 100}));
+        builder.AddLayer_RELU("conv_fwd", "relu_fwd");
+        builder.AddLayer_ADD("data", "relu_fwd", dnn::FuseCode::NONE, "output",
+                             std::make_optional<ModelBuilder::QuantInfo>(
+                                 {Type::TENSOR_QUANT8_ASYMM, {0.05}, 100}));
     } else {
         builder.AddInput("data", {Type::TENSOR_FLOAT32, {1, 224, 224, 3}});
         builder.AddTensorFromBuffer("weight", weight_buf,
                                     {Type::TENSOR_FLOAT32, {3, 1, 1, 3}});
         builder.AddTensorFromBuffer("bias", bias_buf,
                                     {Type::TENSOR_FLOAT32, {3}});
-        builder.AddConv("data", 1, 1, 0, 0, 0, 0, ModelBuilder::ACTIVATION_NONE,
-                        "weight", "bias", "output");
+        builder.AddLayer_CONV_2D("data", "weight", "bias", 1, 1, 0, 0, 0, 0,
+                                 dnn::FuseCode::NONE, false, 1, 1, "output",
+                                 dnn::nullopt);
     }
     auto model = builder.AddOutput("output").Compile(
         ModelBuilder::PREFERENCE_FAST_SINGLE_ANSWER);

diff --git a/ci/download_and_test_models.sh b/ci/download_and_test_models.sh
@@ -2,19 +2,19 @@
 
 set -e
 
-wget "https://s3.amazonaws.com/onnx-model-zoo/squeezenet/squeezenet1.1/squeezenet1.1.tar.gz" -O squeezenet1.1.tar.gz
-tar xvf squeezenet1.1.tar.gz
+wget -q "https://s3.amazonaws.com/onnx-model-zoo/squeezenet/squeezenet1.1/squeezenet1.1.tar.gz" -O squeezenet1.1.tar.gz
+tar xf squeezenet1.1.tar.gz
 python3 ci/validate_onnx.py squeezenet1.1 build_dnnlibrary/binaries/dnn_retrieve_result
 
-wget "https://s3.amazonaws.com/onnx-model-zoo/mobilenet/mobilenetv2-1.0/mobilenetv2-1.0.tar.gz" -O mobilenetv2-1.0.tar.gz
-tar xvf mobilenetv2-1.0.tar.gz
+wget -q "https://s3.amazonaws.com/onnx-model-zoo/mobilenet/mobilenetv2-1.0/mobilenetv2-1.0.tar.gz" -O mobilenetv2-1.0.tar.gz
+tar xf mobilenetv2-1.0.tar.gz
 python3 ci/validate_onnx.py mobilenetv2-1.0 build_dnnlibrary/binaries/dnn_retrieve_result
 
-wget "https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet18v2/resnet18v2.tar.gz" -O resnet18v2.tar.gz
-tar xvf resnet18v2.tar.gz
+wget -q "https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet18v2/resnet18v2.tar.gz" -O resnet18v2.tar.gz
+tar xf resnet18v2.tar.gz
 python3 ci/validate_onnx.py resnet18v2 build_dnnlibrary/binaries/dnn_retrieve_result
 
-wget "https://s3.amazonaws.com/download.onnx/models/opset_9/bvlc_googlenet.tar.gz" -O bvlc_googlenet.tar.gz
-tar xvf bvlc_googlenet.tar.gz
+wget -q "https://s3.amazonaws.com/download.onnx/models/opset_9/bvlc_googlenet.tar.gz" -O bvlc_googlenet.tar.gz
+tar xf bvlc_googlenet.tar.gz
 python3 ci/validate_onnx.py bvlc_googlenet build_dnnlibrary/binaries/dnn_retrieve_result
 
diff --git a/common/Shaper.cpp b/common/Shaper.cpp
@@ -10,35 +10,39 @@ Shaper::len_t Shaper::total(const Shape &shape) {
     return Product(shape);
 }
 
-/**
- *  strides: [stride_y, stride_x]
- *  paddings: [top, left, bottom, right]
- */
-void Shaper::Conv(const std::string &input_name, const std::string &weight_name,
-                  const std::vector<int32_t> paddings,
-                  const std::vector<int32_t> strides,
-                  const std::string &output_name) {
-    Shaper::Conv(input_name, strides[1], strides[0], 1, 1, paddings[1],
-                 paddings[3], paddings[0], paddings[2], weight_name,
-                 output_name);
-}
-
-void Shaper::Conv(const std::string &input_name,
-                  const std::vector<int32_t> paddings,
-                  const std::vector<int32_t> strides,
-                  const std::vector<int32_t> dilations,
-                  const std::string &weight_name,
-                  const std::string &output_name) {
-    Shaper::Conv(input_name, strides[1], strides[0], dilations[1], dilations[0],
-                 paddings[1], paddings[3], paddings[0], paddings[2],
-                 weight_name, output_name);
-}
+// /**
+//  *  strides: [stride_y, stride_x]
+//  *  paddings: [top, left, bottom, right]
+//  */
+// void Shaper::Conv(const std::string &input_name, const std::string
+// &weight_name,
+//                   const std::vector<int32_t> paddings,
+//                   const std::vector<int32_t> strides,
+//                   const std::string &output_name) {
+//     Shaper::Conv(input_name, strides[1], strides[0], 1, 1, paddings[1],
+//                  paddings[3], paddings[0], paddings[2], weight_name,
+//                  output_name);
+// }
+//
+// void Shaper::Conv(const std::string &input_name,
+//                   const std::vector<int32_t> paddings,
+//                   const std::vector<int32_t> strides,
+//                   const std::vector<int32_t> dilations,
+//                   const std::string &weight_name,
+//                   const std::string &output_name) {
+//     Shaper::Conv(input_name, strides[1], strides[0], dilations[1],
+//     dilations[0],
+//                  paddings[1], paddings[3], paddings[0], paddings[2],
+//                  weight_name, output_name);
+// }
 void Shaper::Conv(const std::string &input, const std::string &weight,
                   int32_t padding_left, int32_t padding_right,
                   int32_t padding_top, int32_t padding_bottom, int32_t stride_x,
-                  int32_t stride_y, const std::string &output) {
-    Conv(input, stride_x, stride_y, 1, 1, padding_left, padding_right,
-         padding_top, padding_bottom, weight, output);
+                  int32_t stride_y, const bool nchw, const int32_t dilation_x,
+                  const int32_t dilation_y, const std::string &output) {
+    DNN_ASSERT_EQ(nchw, false);
+    Conv(input, stride_x, stride_y, dilation_x, dilation_y, padding_left,
+         padding_right, padding_top, padding_bottom, weight, output);
 }
 
 void Shaper::Conv(const std::string &input_name, int32_t strideX,
@@ -63,17 +67,18 @@ void Shaper::Conv(const std::string &input_name, int32_t strideX,
     shape_map_[output_name] = outputDimen;
 }
 
-void Shaper::DepthwiseConv(const std::string &input_name,
-                           const std::vector<int32_t> paddings,
-                           const std::vector<int32_t> strides,
-                           const std::vector<int32_t> dilations,
-                           const std::string &weight_name,
-                           const std::string &output_name) {
-    Shaper::DepthwiseConv(input_name, strides[1], strides[0], dilations[1],
-                          dilations[0], paddings[1], paddings[3], paddings[0],
-                          paddings[2], weight_name, output_name);
-}
-
+// void Shaper::DepthwiseConv(const std::string &input_name,
+//                            const std::vector<int32_t> paddings,
+//                            const std::vector<int32_t> strides,
+//                            const std::vector<int32_t> dilations,
+//                            const std::string &weight_name,
+//                            const std::string &output_name) {
+//     Shaper::DepthwiseConv(input_name, strides[1], strides[0], dilations[1],
+//                           dilations[0], paddings[1], paddings[3],
+//                           paddings[0], paddings[2], weight_name,
+//                           output_name);
+// }
+//
 void Shaper::DepthwiseConv(const std::string &input_name,
                            const std::string &weight_name, int32_t padding_left,
                            int32_t padding_right, int32_t padding_top,
@@ -185,7 +190,7 @@ void Shaper::Softmax(const std::string &input_name,
     shape_map_[output_name] = shape_map_.at(input_name);
 }
 
-void Shaper::Relu(const std::string &input_name,
+void Shaper::ReLU(const std::string &input_name,
                   const std::string &output_name) {
     shape_map_[output_name] = shape_map_.at(input_name);
 }