thxx - libtorch C++ API extentions

documentation

• sphinx https://shigekikarita.github.io/thxx/sphinx/html
• doxygen https://shigekikarita.github.io/thxx/

CI status

https://travis-ci.org/ShigekiKarita/thxx

compiler	conda package	latest zip
gcc-7
gcc-8
clang-5
clang-6
clang-7

requirements

• gcc 7, 8 or clang 5, 6, 7
• compiler option -std=c++17
• libtorch 1.0.0 (recommend conda install pytorch-cpu=1.0.0 -c pytorch)

how to run

• using conda (recommend)

$ conda install pytorch-cpu=1.0.0 -c pytorch
$ make test
$ make example-mnist

• using libtorch latest zip

$ cd 
$ make --directory ./third_party libtorch-shared-with-deps-latest.zip
$ make test
$ make example-mnist

for more details, see .travis.yml.

features

functional meta modules

• thxx::meta::Lambda<...> : torch::nn::ModuleHolder<...>

auto x = torch::rand({2, 3, 4});

// single input/output
Lambda f1 = lambda([](auto&& x) { return x * 2; });
auto x1 = f1->forward(x);
CHECK_THAT( x1, testing::TensorEq(x * 2) );

// multi input/output
Lambda f2 = lambda([](auto&& x1, auto&& x2) { return std::make_tuple(x1.relu(), x2 * 2); });
{
    auto [a, b] = f2->forward(x, x);
    CHECK_THAT( a, testing::TensorEq(torch::relu(x)) );
    CHECK_THAT( b, testing::TensorEq(x * 2) );
}
{
    // automatic tuple arg unpack
    auto [a, b] = f1->forward(std::make_tuple(x, x));
    CHECK_THAT( a, testing::TensorEq(torch::relu(x)) );
    CHECK_THAT( b, testing::TensorEq(x * 2) );
}

• thxx::meta::Seq<...> : torch::nn::ModuleHolder<...>

auto x1 = torch::rand({2, 3, 4});

// single input/output with Lambda
auto f1 = torch::nn::Linear(4, 5);
auto f2 = sequential(f1, lambda(torch::relu));
auto x2 = f2->forward(x1);
CHECK_THAT( x2, testing::TensorEq(torch::relu(f1->forward(x1)) );

// share submodules/parameters
CHECK_THAT( f2->parameters[0], testing::TensorEq(f1->weight) );
CHECK_THAT( f2->parameters[1], testing::TensorEq(f1->bias) );

// also support multi input/output with Lambda
auto f3 = lambda([](auto&& x1, auto&& x2) { return std::make_tuple(x1, x2 * 2); });
auto f4 = sequential(f3, f3);;
auto [x3, x4] = f4->forward(x1);
CHECK_THAT( x3, testing::TensorEq(x1) );
CHECK_THAT( x4, testing::TensorEq(x1 * 4) );

see test/test_meta.cpp

networks/modules

• Transformer
  • MultiHeadedAttention
  • PositionalEncoding
  • PositionwiseFeedforward
  • pad/masking functions
• Normalization
  • LayerNorm
• Math (wip)
  • copy C++ batched/complex linalg funcitons from https://github.com/ShigekiKarita/thxx-py
• Loss
  • label smoothing KLDivLoss

see test/test_net.cpp

serialization

• HDF5 (wip)
• numpy (wip)

acknowledgement

mnist example is forked from https://github.com/goldsborough/examples/tree/cpp/cpp/mnist

license

BSL-1.0