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

tests/conv2d_tests.cpp
151 строка7 KB
kolkir
Conv2d backward pass implementation
02 фев 2025, 21:49
02 фев 2025, 21:499fe8c9a
#include <adept/nn/conv2d.hpp>

#include "catch.hpp"

using namespace adept;

TEST_CASE("Conv2dOptions default construction", "[conv2d]") {
  Conv2dOptions opt(1, 1, 3);
  REQUIRE(opt.in_channels == 1);
  REQUIRE(opt.out_channels == 1);
  REQUIRE(opt.bias == true);
  REQUIRE(opt.kernel == ParamArray<2>{3, 3});
  REQUIRE(opt.stride == ParamArray<2>{1, 1});
  REQUIRE(opt.padding == ParamArray<2>{0, 0});
  REQUIRE(opt.dilation == ParamArray<2>{1, 1});
}

TEST_CASE("Conv2dOptions extended construction", "[conv2d]") {
  auto opt = Conv2dOptions(1, 1, 3)
                 .with_padding({3, 1})
                 .with_bias(false)
                 .with_stride({3, 1})
                 .with_dilation({3, 1});
  REQUIRE(opt.in_channels == 1);
  REQUIRE(opt.out_channels == 1);
  REQUIRE(opt.bias == false);
  REQUIRE(opt.kernel == ParamArray<2>{3, 3});
  REQUIRE(opt.stride == ParamArray<2>{3, 1});
  REQUIRE(opt.padding == ParamArray<2>{3, 1});
  REQUIRE(opt.dilation == ParamArray<2>{3, 1});
}

TEST_CASE("Conv2dO construction", "[conv2d]") {
  Conv2d conv(Conv2dOptions(1, 1, 3));
}

TEST_CASE("Conv2d im3x3 k2x2 ch1x1", "[conv2d]") {
  std::vector<float32_t> im_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
  std::vector<float32_t> kernel_data = {7, 8, 1, 3};

  auto im = Tensor::from_blob(
      im_data.data(),
      {.shape = {1, 1, 3, 3}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto kernels = Tensor::from_blob(
      kernel_data.data(),
      {.shape = {1, 1, 2, 2}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});

  Conv2d conv(Conv2dOptions(1, 1, 2).with_bias(false));
  conv->set_weights(kernels);
  auto res = conv(im);

  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 0}), Catch::WithinRel(42.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 1}), Catch::WithinRel(61.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 0}), Catch::WithinRel(99.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 1}), Catch::WithinRel(118.f, 0.001f));

  mean(res).backward();

  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 0, 0}), Catch::WithinRel(3.f, 0.001f));
  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 0, 1}), Catch::WithinRel(4.f, 0.001f));
  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 1, 0}), Catch::WithinRel(6.f, 0.001f));
  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 1, 1}), Catch::WithinRel(7.f, 0.001f));
}

TEST_CASE("Conv2d im3x3 k2x2 ch1x2", "[conv2d]") {
  std::vector<float32_t> im_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
  std::vector<float32_t> kernel_data = {7, 8, 1, 3, 7, 8, 1, 3};

  auto im = Tensor::from_blob(
      im_data.data(),
      {.shape = {1, 1, 3, 3}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto kernels = Tensor::from_blob(
      kernel_data.data(),
      {.shape = {2, 1, 2, 2}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});

  Conv2d conv(Conv2dOptions(1, 2, 2).with_bias(false));
  conv->set_weights(kernels);
  auto res = conv(im);

  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 0}), Catch::WithinRel(42.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 1}), Catch::WithinRel(61.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 0}), Catch::WithinRel(99.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 1}), Catch::WithinRel(118.f, 0.001f));

  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 0, 0}), Catch::WithinRel(42.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 0, 1}), Catch::WithinRel(61.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 1, 0}), Catch::WithinRel(99.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 1, 1}), Catch::WithinRel(118.f, 0.001f));
}

TEST_CASE("Conv2d im3x3 k2x2 ch1x1 bias", "[conv2d]") {
  std::vector<float32_t> im_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
  std::vector<float32_t> kernel_data = {7, 8, 1, 3};

  auto im = Tensor::from_blob(
      im_data.data(),
      {.shape = {1, 1, 3, 3}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto kernels = Tensor::from_blob(
      kernel_data.data(),
      {.shape = {1, 1, 2, 2}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});

  auto bias = Tensor::from_values({1.f}, {1, 1}, device_t::CPU);

  Conv2d conv(Conv2dOptions(1, 1, 2).with_bias(true));
  conv->set_weights(kernels);
  conv->set_bias(bias);
  auto res = conv(im);

  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 0}), Catch::WithinRel(43.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 1}), Catch::WithinRel(62.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 0}), Catch::WithinRel(100.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 1}), Catch::WithinRel(119.f, 0.001f));

  mean(res).backward();

  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 0, 0}), Catch::WithinRel(3.f, 0.001f));
  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 0, 1}), Catch::WithinRel(4.f, 0.001f));
  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 1, 0}), Catch::WithinRel(6.f, 0.001f));
  REQUIRE_THAT(conv->weights().grad().at<float32_t>({0, 0, 1, 1}), Catch::WithinRel(7.f, 0.001f));

  REQUIRE_THAT(conv->bias().grad().at<float32_t>({0, 0}), Catch::WithinRel(1.f, 0.001f));
}

TEST_CASE("Conv2d im3x3 k2x2 ch1x2 bias", "[conv2d]") {
  std::vector<float32_t> im_data = {1, 2, 3, 4, 5, 6, 7, 8, 9};
  std::vector<float32_t> kernel_data = {7, 8, 1, 3, 7, 8, 1, 3};

  auto im = Tensor::from_blob(
      im_data.data(),
      {.shape = {1, 1, 3, 3}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});
  auto kernels = Tensor::from_blob(
      kernel_data.data(),
      {.shape = {2, 1, 2, 2}, .device = device_t::CPU, .dtype = to_dtype<float32_t>()});

  auto bias = Tensor::from_values({1.f, 1.f}, {1, 2}, device_t::CPU);

  Conv2d conv(Conv2dOptions(1, 2, 2).with_bias(true));
  conv->set_weights(kernels);
  conv->set_bias(bias);
  auto res = conv(im);

  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 0}), Catch::WithinRel(43.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 0, 1}), Catch::WithinRel(62.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 0}), Catch::WithinRel(100.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 0, 1, 1}), Catch::WithinRel(119.f, 0.001f));

  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 0, 0}), Catch::WithinRel(43.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 0, 1}), Catch::WithinRel(62.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 1, 0}), Catch::WithinRel(100.f, 0.001f));
  REQUIRE_THAT(res.data().at<float32_t>({0, 1, 1, 1}), Catch::WithinRel(119.f, 0.001f));
}