vuron/adept: Платформа для обучения и вывода нейронных сетей, с API приближенным к PyTorch.

tests/transpose_tests.cpp
150 строк5 KB
kolkir
revert compiler compatibility to gcc12
03 мар 2025, 23:52
03 мар 2025, 23:5226a9274
#include <adept/tensor.hpp>
#include <adept/tensor_print.hpp>
#include <iterator>
#include <tuple>
#include <vector>

#include "adept/tensor_print.hpp"
#include "adept/types.hpp"
#include "catch.hpp"

using namespace adept;

namespace {

template <typename T>
using Data = std::vector<T>;

template <typename T, template <class> class Op>
std::tuple<Data<T>, Data<T>, Data<T>> make_test_data() {
  Data<T> x = {1, 2, 3, 4, 5, 6, 7, 8, 9};
  Data<T> y = {7, 8, 1, 3, 9, 4, 5, 2, 6};
  Data<T> z;
  std::transform(x.begin(), x.end(), y.begin(), std::back_inserter(z), Op<T>());
  return {x, y, z};
}
}  // namespace

TEST_CASE("Tensor transpose 1xN", "[transpose]") {
  auto [dx, dy, dz] = make_test_data<float32_t, std::plus>();
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {1, 9}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(1));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(0));

  for (index_t i = 0; i < x.properties().shape.numel(); ++i) {
    REQUIRE_THAT(y.at<float32_t>({i, 0}), Catch::WithinRel(x.at<float32_t>({0, i}), 0.001f));
  }
}

TEST_CASE("Tensor transpose Nx1", "[transpose]") {
  auto [dx, dy, dz] = make_test_data<float32_t, std::plus>();
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {9, 1}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(1));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(0));

  for (index_t i = 0; i < x.properties().shape.numel(); ++i) {
    REQUIRE_THAT(y.at<float32_t>({0, i}), Catch::WithinRel(x.at<float32_t>({i, 0}), 0.001f));
  }
}

TEST_CASE("Tensor transpose NxN", "[transpose]") {
  auto [dx, dy, dz] = make_test_data<float32_t, std::plus>();
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {3, 3}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(1));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(0));

  for (index_t i = 0; i < 3; ++i) {
    for (index_t j = 0; j < 3; ++j) {
      REQUIRE_THAT(y.at<float32_t>({i, j}), Catch::WithinRel(x.at<float32_t>({j, i}), 0.001f));
    }
  }
}

TEST_CASE("Tensor transpose MxN", "[transpose]") {
  auto [dx, dy, dz] = make_test_data<float32_t, std::plus>();
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {2, 3}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(1));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(0));

  for (index_t i = 0; i < 3; ++i) {
    for (index_t j = 0; j < 2; ++j) {
      REQUIRE_THAT(y.at<float32_t>({i, j}), Catch::WithinRel(x.at<float32_t>({j, i}), 0.001f));
    }
  }
}

TEST_CASE("Tensor transpose NxM", "[transpose]") {
  auto [dx, dy, dz] = make_test_data<float32_t, std::plus>();
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {3, 2}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(1));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(0));

  for (index_t i = 0; i < 2; ++i) {
    for (index_t j = 0; j < 3; ++j) {
      REQUIRE_THAT(y.at<float32_t>({i, j}), Catch::WithinRel(x.at<float32_t>({j, i}), 0.001f));
    }
  }
}

TEST_CASE("Tensor transpose NxM batched", "[transpose]") {
  auto [dx, dy, dz] = make_test_data<float32_t, std::plus>();
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {2, 1, 2}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(0));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(2));
  REQUIRE(y.properties().shape.dim(2) == x.properties().shape.dim(1));

  for (index_t n = 0; n < 2; ++n) {
    for (index_t i = 0; i < 2; ++i) {
      for (index_t j = 0; j < 1; ++j) {
        REQUIRE_THAT(y.at<float32_t>({n, i, j}),
                     Catch::WithinRel(x.at<float32_t>({n, j, i}), 0.001f));
      }
    }
  }
}

TEST_CASE("Tensor transpose NxM 2", "[transpose]") {
  Data<float32_t> dx = {1, 2, 3, 4, 5, 6, 7, 8};
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {2, 4}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(1));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(0));

  for (index_t i = 0; i < 4; ++i) {
    for (index_t j = 0; j < 2; ++j) {
      REQUIRE_THAT(y.at<float32_t>({i, j}), Catch::WithinRel(x.at<float32_t>({j, i}), 0.001f));
    }
  }
}

TEST_CASE("Tensor transpose NxM batched 2", "[transpose]") {
  Data<float32_t> dx = {1, 2, 3, 4, 5, 6, 7, 8, 1, 2, 3, 4, 5, 6, 7, 8};
  auto x = Tensor::from_blob(
      dx.data(), {.shape = {2, 2, 4}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto y = x.transpose2d();
  REQUIRE(y.properties().shape.dim(0) == x.properties().shape.dim(0));
  REQUIRE(y.properties().shape.dim(1) == x.properties().shape.dim(2));
  REQUIRE(y.properties().shape.dim(2) == x.properties().shape.dim(1));

  for (index_t n = 0; n < 2; ++n) {
    for (index_t i = 0; i < 4; ++i) {
      for (index_t j = 0; j < 2; ++j) {
        REQUIRE_THAT(y.at<float32_t>({n, i, j}),
                     Catch::WithinRel(x.at<float32_t>({n, j, i}), 0.001f));
      }
    }
  }
}