Merge pull request #7 from mfl28/deep-residual-network

Add Deep Residual Network tutorial

CMakeLists.txt

@@ -19,6 +19,7 @@ add_subdirectory("tutorials/basics/linear_regression")
 add_subdirectory("tutorials/basics/logistic_regression")
 add_subdirectory("tutorials/basics/pytorch_basics")
 add_subdirectory("tutorials/intermediate/convolutional_neural_network")
+add_subdirectory("tutorials/intermediate/deep_residual_network")
 
 # The following code block is suggested to be used on Windows.
 # According to https://github.com/pytorch/pytorch/issues/25457,

README.md

@@ -30,7 +30,7 @@ $ ./scripts.sh build
 
 #### 2. Intermediate
 * [Convolutional Neural Network](https://github.com/prabhuomkar/pytorch-cpp/tree/master/tutorials/intermediate/convolutional_neural_network/src/main.cpp)
-* [Deep Residual Network]()
+* [Deep Residual Network](https://github.com/prabhuomkar/pytorch-cpp/tree/master/tutorials/intermediate/deep_residual_network/src/main.cpp)
 * [Recurrent Neural Network]()
 * [Bidirectional Recurrent Neural Network]()
 * [Language Model (RNN-LM)]()

scripts.sh

@@ -1,9 +1,9 @@
 #!/bin/bash
 
 function install() {
-	wget https://download.pytorch.org/libtorch/nightly/cpu/libtorch-shared-with-deps-latest.zip
-	unzip libtorch-shared-with-deps-latest.zip
-	rm -rf libtorch-shared-with-deps-latest.zip
+	wget https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-1.3.1%2Bcpu.zip
+	unzip libtorch-shared-with-deps-1.3.1+cpu.zip
+	rm -rf libtorch-shared-with-deps-1.3.1+cpu.zip
 }
 
 function build() {

tutorials/intermediate/deep_residual_network/CMakeLists.txt

@@ -0,0 +1,42 @@
+cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
+
+project(deep-residual-network VERSION 1.0.0 LANGUAGES CXX)
+
+# Files
+set(SOURCES src/main.cpp
+            src/cifar10.cpp
+            src/residual_block.cpp
+            src/transform.cpp
+)
+
+set(HEADERS include/residual_block.h
+            include/resnet.h
+            include/cifar10.h
+            include/transform.h
+)
+
+set(EXECUTABLE_NAME deep-residual-network)
+
+
+add_executable(${EXECUTABLE_NAME} ${SOURCES} ${HEADERS})
+target_include_directories(${EXECUTABLE_NAME} PRIVATE include)
+
+target_link_libraries(${EXECUTABLE_NAME} "${TORCH_LIBRARIES}")
+
+set_target_properties(${EXECUTABLE_NAME} PROPERTIES
+  CXX_STANDARD 11
+  CXX_STANDARD_REQUIRED YES
+)
+
+# The following code block is suggested to be used on Windows.
+# According to https://github.com/pytorch/pytorch/issues/25457,
+# the DLLs need to be copied to avoid memory errors.
+# See https://pytorch.org/cppdocs/installing.html.
+if (MSVC)
+  file(GLOB TORCH_DLLS "${TORCH_INSTALL_PREFIX}/lib/*.dll")
+  add_custom_command(TARGET ${EXECUTABLE_NAME}
+                     POST_BUILD
+                     COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                     ${TORCH_DLLS}
+                     $<TARGET_FILE_DIR:${EXECUTABLE_NAME}>)
+endif (MSVC)

tutorials/intermediate/deep_residual_network/include/cifar10.h

@@ -0,0 +1,43 @@
+// Copyright 2019 Markus Fleischhacker
+#pragma once
+
+#include <torch/data/datasets/base.h>
+#include <torch/data/example.h>
+#include <torch/types.h>
+#include <cstddef>
+#include <string>
+
+// CIFAR10 dataset
+// based on: https://github.com/pytorch/pytorch/blob/master/torch/csrc/api/include/torch/data/datasets/mnist.h.
+class CIFAR10 : public torch::data::datasets::Dataset<CIFAR10> {
+ public:
+    // The mode in which the dataset is loaded
+    enum Mode { kTrain, kTest };
+
+    // Loads the CIFAR10 dataset from the `root` path.
+    //
+    // The supplied `root` path should contain the *content* of the unzipped
+    // CIFAR10 dataset (binary version), available from http://www.cs.toronto.edu/~kriz/cifar.html.
+    explicit CIFAR10(const std::string& root, Mode mode = Mode::kTrain);
+
+    // Returns the `Example` at the given `index`.
+    torch::data::Example<> get(size_t index) override;
+
+    // Returns the size of the dataset.
+    torch::optional<size_t> size() const override;
+
+    // Returns true if this is the training subset of CIFAR10.
+    bool is_train() const noexcept;
+
+    // Returns all images stacked into a single tensor.
+    const torch::Tensor& images() const;
+
+    // Returns all targets stacked into a single tensor.
+    const torch::Tensor& targets() const;
+
+ private:
+    torch::Tensor images_;
+    torch::Tensor targets_;
+    Mode mode_;
+};
+

tutorials/intermediate/deep_residual_network/include/residual_block.h

@@ -0,0 +1,25 @@
+// Copyright 2019 Markus Fleischhacker
+#pragma once
+
+#include <torch/torch.h>
+
+namespace resnet {
+class ResidualBlockImpl : public torch::nn::Module {
+ public:
+    ResidualBlockImpl(int64_t in_channels, int64_t out_channels, int64_t stride = 1,
+        torch::nn::Sequential downsample = nullptr);
+    torch::Tensor forward(torch::Tensor x);
+
+ private:
+    torch::nn::Conv2d conv1;
+    torch::nn::BatchNorm bn1;
+    torch::nn::Functional relu;
+    torch::nn::Conv2d conv2;
+    torch::nn::BatchNorm bn2;
+    torch::nn::Sequential downsampler;
+};
+
+torch::nn::Conv2d conv3x3(int64_t in_channels, int64_t out_channels, int64_t stride = 1);
+
+TORCH_MODULE(ResidualBlock);
+}  // namespace resnet

tutorials/intermediate/deep_residual_network/include/resnet.h

@@ -0,0 +1,91 @@
+// Copyright 2019 Markus Fleischhacker
+#pragma once
+
+#include <torch/torch.h>
+#include <vector>
+#include "residual_block.h"
+
+namespace resnet {
+template<typename Block>
+class ResNetImpl : public torch::nn::Module {
+ public:
+    explicit ResNetImpl(const std::array<int64_t, 3>& layers, int64_t num_classes = 10);
+    torch::Tensor forward(torch::Tensor x);
+
+ private:
+    int64_t in_channels = 16;
+    torch::nn::Conv2d conv{conv3x3(3, 16)};
+    torch::nn::BatchNorm bn{16};
+    torch::nn::Functional relu{torch::relu};
+    torch::nn::Sequential layer1;
+    torch::nn::Sequential layer2;
+    torch::nn::Sequential layer3;
+    torch::nn::AvgPool2d avg_pool{8};
+    torch::nn::Linear fc;
+
+    torch::nn::Sequential make_layer(int64_t out_channels, int64_t blocks, int64_t stride = 1);
+};
+
+template<typename Block>
+ResNetImpl<Block>::ResNetImpl(const std::array<int64_t, 3>& layers, int64_t num_classes) :
+    layer1(make_layer(16, layers[0])),
+    layer2(make_layer(32, layers[1], 2)),
+    layer3(make_layer(64, layers[2], 2)),
+    fc(64, num_classes) {
+    register_module("conv", conv);
+    register_module("bn", bn);
+    register_module("relu", relu);
+    register_module("layer1", layer1);
+    register_module("layer2", layer2);
+    register_module("layer3", layer3);
+    register_module("avg_pool", avg_pool);
+    register_module("fc", fc);
+}
+
+template<typename Block>
+torch::Tensor ResNetImpl<Block>::forward(torch::Tensor x) {
+    auto out = conv->forward(x);
+    out = bn->forward(out);
+    out = relu->forward(out);
+    out = layer1->forward(out);
+    out = layer2->forward(out);
+    out = layer3->forward(out);
+    out = avg_pool->forward(out);
+    out = out.view({out.size(0), -1});
+    out = fc->forward(out);
+
+    return torch::log_softmax(out, 1);
+}
+
+template<typename Block>
+torch::nn::Sequential ResNetImpl<Block>::make_layer(int64_t out_channels, int64_t blocks, int64_t stride) {
+    torch::nn::Sequential layers;
+    torch::nn::Sequential downsample{nullptr};
+
+    if (stride != 1 || in_channels != out_channels) {
+        downsample = torch::nn::Sequential{
+            conv3x3(in_channels, out_channels, stride),
+            torch::nn::BatchNorm(out_channels)
+        };
+    }
+
+    layers->push_back(Block(in_channels, out_channels, stride, downsample));
+
+    in_channels = out_channels;
+
+    for (int64_t i = 1; i != blocks; ++i) {
+        layers->push_back(Block(out_channels, out_channels));
+    }
+
+    return layers;
+}
+
+// Wrap class into ModuleHolder (a shared_ptr wrapper),
+// see https://github.com/pytorch/pytorch/blob/master/torch/csrc/api/include/torch/nn/pimpl.h
+template<typename Block = ResidualBlock>
+class ResNet : public torch::nn::ModuleHolder<ResNetImpl<Block>> {
+ public:
+    using torch::nn::ModuleHolder<ResNetImpl<Block>>::ModuleHolder;
+};
+}  // namespace resnet
+

tutorials/intermediate/deep_residual_network/include/transform.h

@@ -0,0 +1,55 @@
+// Copyright 2019 Markus Fleischhacker
+#pragma once
+
+#include <torch/torch.h>
+#include <random>
+
+namespace transform {
+class RandomHorizontalFlip : public torch::data::transforms::TensorTransform<torch::Tensor> {
+ public:
+    // Creates a transformation that randomly horizontally flips a tensor.
+    //
+    // The parameter `p` determines the probability that a tensor is flipped (default = 0.5).
+    explicit RandomHorizontalFlip(double p = 0.5);
+
+    torch::Tensor operator()(torch::Tensor input) override;
+
+ private:
+    double p_;
+};
+
+class ConstantPad : public torch::data::transforms::TensorTransform<torch::Tensor> {
+ public:
+    // Creates a transformation that pads a tensor.
+    //
+    // `padding` is expected to be a vector of size 4 whose entries correspond to the
+    // padding of the sides, i.e {left, right, top, bottom}. `value` determines the value
+    // for the padded pixels.
+    explicit ConstantPad(const std::vector<int64_t>& padding, torch::Scalar value = 0);
+
+    // Creates a transformation that pads a tensor.
+    //
+    // The padding will be performed using the size `padding` for all 4 sides.
+    // `value` determines the value for the padded pixels.
+    explicit ConstantPad(int64_t padding, torch::Scalar value = 0);
+
+    torch::Tensor operator()(torch::Tensor input) override;
+
+ private:
+    std::vector<int64_t> padding_;
+    torch::Scalar value_;
+};
+
+class RandomCrop : public torch::data::transforms::TensorTransform<torch::Tensor> {
+ public:
+    // Creates a transformation that randomly crops a tensor.
+    //
+    // The parameter `size` is expected to be a vector of size 2
+    // and determines the output size {height, width}.
+    explicit RandomCrop(const std::vector<int64_t>& size);
+    torch::Tensor operator()(torch::Tensor input) override;
+
+ private:
+    std::vector<int64_t> size_;
+};
+}  // namespace transform

tutorials/intermediate/deep_residual_network/src/cifar10.cpp

@@ -0,0 +1,98 @@
+// Copyright 2019 Markus Fleischhacker
+#include "cifar10.h"
+
+namespace {
+// CIFAR10 dataset description can be found at https://www.cs.toronto.edu/~kriz/cifar.html.
+constexpr uint32_t kTrainSize = 50000;
+constexpr uint32_t kTestSize = 10000;
+constexpr uint32_t kSizePerBatch = 10000;
+constexpr uint32_t kImageRows = 32;
+constexpr uint32_t kImageColumns = 32;
+constexpr uint32_t kBytesPerRow = 3073;
+constexpr uint32_t kBytesPerChannelPerRow = (kBytesPerRow - 1) / 3;
+constexpr uint32_t kBytesPerBatchFile = kBytesPerRow * kSizePerBatch;
+
+const std::vector<std::string> kTrainDataBatchFiles = {
+    "data_batch_1.bin",
+    "data_batch_2.bin",
+    "data_batch_3.bin",
+    "data_batch_4.bin",
+    "data_batch_5.bin",
+};
+
+const std::vector<std::string> kTestDataBatchFiles = {
+    "test_batch.bin"
+};
+
+// Source: https://github.com/pytorch/pytorch/blob/master/torch/csrc/api/src/data/datasets/mnist.cpp.
+std::string join_paths(std::string head, const std::string& tail) {
+    if (head.back() != '/') {
+        head.push_back('/');
+    }
+    head += tail;
+    return head;
+}
+// Partially based on https://github.com/pytorch/pytorch/blob/master/torch/csrc/api/src/data/datasets/mnist.cpp.
+std::pair<torch::Tensor, torch::Tensor> read_data(const std::string& root, bool train) {
+    const auto& files = train ? kTrainDataBatchFiles : kTestDataBatchFiles;
+    const auto num_samples = train ? kTrainSize : kTestSize;
+
+    std::vector<char> data_buffer;
+    data_buffer.reserve(files.size() * kBytesPerBatchFile);
+
+    for (const auto& file : files) {
+        const auto path = join_paths(root, file);
+        std::ifstream data(path, std::ios::binary);
+        TORCH_CHECK(data, "Error opening data file at", path);
+
+        data_buffer.insert(data_buffer.end(), std::istreambuf_iterator<char>(data), {});
+    }
+
+    TORCH_CHECK(data_buffer.size() == files.size() * kBytesPerBatchFile, "Unexpected file sizes");
+
+    auto targets = torch::empty(num_samples, torch::kByte);
+    auto images = torch::empty({num_samples, 3, kImageRows, kImageColumns}, torch::kByte);
+
+    for (uint32_t i = 0; i != num_samples; ++i) {
+        // The first byte of each row is the target class index.
+        uint32_t start_index = i * kBytesPerRow;
+        targets[i] = data_buffer[start_index];
+
+        // The next bytes correspond to the rgb channel values in the following order:
+        // red (32 *32 = 1024 bytes) | green (1024 bytes) | blue (1024 bytes)
+        uint32_t image_start = start_index + 1;
+        uint32_t image_end = image_start + 3 * kBytesPerChannelPerRow;
+        std::copy(&data_buffer[image_start], &data_buffer[image_end],
+            reinterpret_cast<char*>(images[i].data_ptr()));
+    }
+
+    return {images.to(torch::kFloat32).div_(255), targets.to(torch::kInt64)};
+}
+}  // namespace
+
+CIFAR10::CIFAR10(const std::string& root, Mode mode) : mode_(mode) {
+    auto data = read_data(root, mode == Mode::kTrain);
+
+    images_ = std::move(data.first);
+    targets_ = std::move(data.second);
+}
+
+torch::data::Example<> CIFAR10::get(size_t index) {
+    return {images_[index], targets_[index]};
+}
+
+torch::optional<size_t> CIFAR10::size() const {
+    return images_.size(0);
+}
+
+bool CIFAR10::is_train() const noexcept {
+    return mode_ == Mode::kTrain;
+}
+
+const torch::Tensor& CIFAR10::images() const {
+    return images_;
+}
+
+const torch::Tensor& CIFAR10::targets() const {
+    return targets_;
+}