Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-12-23 14:54:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

Migrate batch norm layer

Browse Source
This commit is contained in:
Matúš Námešný
2025-11-21 23:24:14 +01:00
parent 5679dc0a50
commit fd4775faa4
11 changed files with 181 additions and 364 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
include/backend.hpp
View File

@@ -16,6 +16,7 @@ class Backend {
// Tensor ops
virtual void print(const CUDANet::Tensor& input) = 0;
virtual void zero(CUDANet::Tensor& input) = 0;
virtual void fill(CUDANet::Tensor& input, int data) = 0;
virtual void
copy_to_device(CUDANet::Tensor& tensor, void* data, size_t size) = 0;
@@ -53,7 +54,7 @@ class Backend {
const CUDANet::Shape out_shape
) = 0;
virtual CUDANet::Tensor& maxPool2d(
virtual CUDANet::Tensor& max_pool2d(
const CUDANet::Tensor& input,
CUDANet::Tensor& output,
CUDANet::Shape input_shape,
@@ -63,7 +64,7 @@ class Backend {
CUDANet::Shape output_shape
) = 0;
virtual CUDANet::Tensor& avgPool2d(
virtual CUDANet::Tensor& avg_pool2d(
const CUDANet::Tensor& input,
CUDANet::Tensor& output,
CUDANet::Shape input_shape,
@@ -72,6 +73,17 @@ class Backend {
CUDANet::Shape padding_shape,
CUDANet::Shape output_shape
) = 0;
virtual CUDANet::Tensor& batch_norm(
const CUDANet::Tensor& input,
CUDANet::Tensor& output,
CUDANet::Shape input_shape,
CUDANet::Tensor& weights,
CUDANet::Tensor& biases,
CUDANet::Tensor& running_mean,
CUDANet::Tensor& running_var,
CUDANet::Tensor& epsilon
) = 0;
};
} // namespace CUDANet

18
include/backend/cuda.cuh
View File

@@ -14,6 +14,7 @@ class CUDA : public Backend {
// Tensor ops
void print(const CUDANet::Tensor& input) override;
void zero(CUDANet::Tensor& input) override;
void fill(CUDANet::Tensor &input, int value) override;
void
copy_to_device(CUDANet::Tensor& tensor, void* data, size_t size) override;
void sum(const CUDANet::Tensor& input, CUDANet::Tensor& sum) override;
@@ -49,7 +50,7 @@ class CUDA : public Backend {
const CUDANet::Shape out_shape
) override;
CUDANet::Tensor& maxPool2d(
CUDANet::Tensor& max_pool2d(
const CUDANet::Tensor& input,
CUDANet::Tensor& output,
CUDANet::Shape input_shape,
@@ -59,7 +60,7 @@ class CUDA : public Backend {
CUDANet::Shape output_shape
) override;
CUDANet::Tensor& avgPool2d(
CUDANet::Tensor& avg_pool2d(
const CUDANet::Tensor& input,
CUDANet::Tensor& output,
CUDANet::Shape input_shape,
@@ -67,7 +68,18 @@ class CUDA : public Backend {
CUDANet::Shape stride_shape,
CUDANet::Shape padding_shape,
CUDANet::Shape output_shape
) = 0;
) override;
CUDANet::Tensor& batch_norm(
const CUDANet::Tensor& input,
CUDANet::Tensor& output,
CUDANet::Shape input_shape,
CUDANet::Tensor& weights,
CUDANet::Tensor& biases,
CUDANet::Tensor& running_mean,
CUDANet::Tensor& running_var,
CUDANet::Tensor& epsilon
) override;
};
} // namespace CUDANet::Backend

166
include/layers/batch_norm.hpp
View File

@@ -1,170 +1,54 @@
#ifndef CUDANET_BATCH_NORM_H
#define CUDANET_BATCH_NORM_H
#pragma once
#include <vector>
#include "activation.hpp"
#include "layer.hpp"
namespace CUDANet::Layers {
class BatchNorm2d : public WeightedLayer, public TwoDLayer {
class BatchNorm2d : public Layer {
public:
BatchNorm2d(
shape2d inputSize,
int inputChannels,
float epsilon,
ActivationType activationType
);
BatchNorm2d(CUDANet::Shape input_shape, float epsilon, CUDANet::Backend *backend);
~BatchNorm2d();
/**
* @brief Compute the forward pass of the batchnorm layer
*
* @param d_input Device pointer to the input
* @return float* Device pointer to the output
*/
float* forward(const float* d_input);
CUDANet::Tensor& forward(CUDANet::Tensor& input) override;
/**
* @brief Set the weights of the batchnorm layer
*
* @param weights_input Pointer to the weights
*/
void setWeights(const float* weights_input);
CUDANet::Shape input_shape() override;
/**
* @brief Get the weights of the batchnorm layer
*
* @return std::vector<float>
*/
std::vector<float> getWeights();
CUDANet::Shape output_shape() override;
/**
* @brief Set the biases of the batchnorm layer
*
* @param biases_input Pointer to the biases
*/
void setBiases(const float* biases_input);
size_t input_size() override;
/**
* @brief Get the biases of the batchnorm layer
*
* @return std::vector<float>
*/
std::vector<float> getBiases();
size_t output_size() override;
/**
* @brief Set the Running Mean
*
* @param running_mean_input
*/
void setRunningMean(const float* running_mean_input);
void set_weights(void* input) override;
/**
* @brief Get the Running Mean
*
*/
std::vector<float> getRunningMean();
CUDANet::Tensor& get_weights() override;
/**
* @brief Set the Running Var
*
* @param running_mean_input
*/
void setRunningVar(const float* running_mean_input);
void set_biases(void* input) override;
/**
* @brief Get the Running Var
*
*/
std::vector<float> getRunningVar();
CUDANet::Tensor& get_biases() override;
/**
* @brief Get output size
*
* @return int output size
*/
int getOutputSize();
void set_running_mean(void* input);
/**
* @brief Get input size
*
* @return int input size
*/
int getInputSize();
CUDANet::Tensor& get_running_mean();
shape2d getOutputDims();
void set_running_var(void* input);
CUDANet::Tensor& get_running_var();
private:
shape2d inputSize;
int inputChannels;
float epsilon;
CUDANet::Shape in_shape;
CUDANet::Tensor epsilon;
int gridSize;
CUDANet::Tensor running_mean;
CUDANet::Tensor running_var;
#ifdef USE_CUDA
CUDANet::Tensor weights;
CUDANet::Tensor biases;
float* d_output;
CUDANet::Tensor output;
float* d_running_mean;
float* d_running_var;
float* d_length;
float* d_epsilon;
float* d_weights;
float* d_biases;
void initCUDA();
void delCUDA();
/**
* @brief Copy weights and biases to the device
*
*/
void toCuda();
float* forwardCUDA(const float* d_input);
#endif
std::vector<float> weights;
std::vector<float> biases;
std::vector<float> running_mean;
std::vector<float> running_var;
Activation* activation;
float* forwardCPU(const float* input);
/**
* @brief Initialize weights of the batchnorm layer with zeros
*
*/
void initializeWeights();
/**
* @brief Initialize biases of the batchnorm layer with zeros
*
*/
void initializeBiases();
/**
* @brief Initialize mean of the batchnorm layer with zeros
*
*/
void initializeRunningMean();
/**
* @brief Initialize sqrt of variance of the batchnorm layer with ones
*
*/
void initializeRunningVar();
CUDANet::Backend *backend;
};
} // namespace CUDANet::Layers
#endif // CUDANET_BATCH_NORM_H

5
include/tensor.hpp
View File

@@ -45,6 +45,11 @@ public:
void zero();
template <typename T>
void fill(T value) {
backend->fill(*this, value);
}
template <typename T>
void set_data(T *data) {
backend->copy_to_device(*this, data, total_size);