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

src/nn/batchnorm2d.cpp
92 строки3 KB
kolkir
Make test convnet cpp app
07 мар 2025, 10:56
07 мар 2025, 10:56997ae5b
#include <adept/nn/batchnorm2d.hpp>
#include <adept/nn/init.hpp>
#include "../dispatch/batchnorm.hpp"

namespace adept {

BarchNorm2dImpl::BarchNorm2dImpl(index_t num_features,
                                 float32_t eps,
                                 float32_t momentum,
                                 device_t device,
                                 dtype_t dtype)
    : num_features_(num_features),
      eps_(eps),
      momentum_(momentum),
      weight_(Tensor::empty({.shape = {num_features}, .device = device, .dtype = dtype})),
      bias_(Tensor::empty({.shape = {num_features}, .device = device, .dtype = dtype})),
      running_mean_(Tensor::empty({.shape = {num_features}, .device = device, .dtype = dtype})),
      running_var_(Tensor::empty({.shape = {num_features}, .device = device, .dtype = dtype})) {
  init_weights();
  set_name("BatchNorm2d");
  register_parameter("weight", weight_);
  register_parameter("bias", bias_);

  // TODO: consider add buffers registering for load/store automating
  running_mean_.requires_grad(false);
  register_parameter("running_mean_", running_mean_);
  running_var_.requires_grad(false);
  register_parameter("running_var", running_var_);
}

void BarchNorm2dImpl::init_weights() {
  fill(weight_.data(), 1.f);
  fill_zero(bias_.data());
  fill_zero(running_mean_.data());
  fill_zero(running_var_.data());
}

void BarchNorm2dImpl::set_weights(const Tensor& weight) {
  weight_.data() = weight;
}

void BarchNorm2dImpl::set_bias(const Tensor& bias) {
  bias_.data() = bias;
}

Variable BarchNorm2dImpl::weights() const {
  return weight_;
}

Variable BarchNorm2dImpl::bias() const {
  return bias_;
}

void BarchNorm2dImpl::set_running_mean(const Tensor& mean) {
  running_mean_.data() = mean;
}

void BarchNorm2dImpl::set_running_var(const Tensor& var) {
  running_var_.data() = var;
}

Variable BarchNorm2dImpl::running_mean() const {
  return running_mean_;
}

Variable BarchNorm2dImpl::running_var() const {
  return running_var_;
}

Variable BarchNorm2dImpl::forward(const Variable& input) {
  CHECK(input.data().shape().dim(1) == num_features_,
        " BarchNorm2d input has invalid feature number ", num_features_,
        " != ", input.data().shape().dim(1));
  auto [result, save_mean, save_var] =
      batchnorm2d_fwd(input.data(), weight_.data(), bias_.data(), running_mean_.data(),
                      running_var_.data(), eps_, momentum_, is_train_mode_);
  Variable var(std::move(result), {input}, "batchnorm2d");
  var.set_backward_fn([this, input = input, save_mean = save_mean,
                       save_var = save_var](const auto& out_grad) mutable {
    Tensor input_grad;
    if (input.requires_grad())
      input_grad = input.grad();

    auto [weight_grad, bias_grad] = batchnorm2d_bwd(input_grad, out_grad, input.data(),
                                                    weight_.data(), save_mean, save_var, eps_);
    weight_.add_grad(weight_grad);
    bias_.add_grad(bias_grad);
  });
  return var;
}

}  // namespace adept