im2col.h

Go to the documentation of this file.

/*
 * bc_im2col.h
 *
 *  Created on: Jan 9, 2020
 *      Author: joseph
 */

#ifndef BC_IM2COL_H_
#define BC_IM2COL_H_

namespace bc {
namespace nn {
namespace functions {

inline bool is_a_ge_zero_and_a_lt_b(int a, int b) {
    return static_cast<unsigned>(a) < static_cast<unsigned>(b);
}

template <typename Dtype>
void im2col(
        bc::host_tag,
        const Dtype* data_im, const int channels,
        const int height, const int width,
        const int kernel_h, const int kernel_w,

        const int pad_h, const int pad_w,
        const int stride_h, const int stride_w,

        const int dilation_h, const int dilation_w,
        Dtype* data_col) {

    int image_h_end = (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1));
    int image_w_end =  (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1));

    const int output_h = image_h_end / stride_h + 1;
    const int output_w = image_w_end / stride_w + 1;

    using bc::tensors::exprs::Kernel_Array;

    using column_tensor_type = Kernel_Array<bc::Shape<5>, Dtype, bc::host_tag>;
    using image_tensor_type = Kernel_Array<bc::Shape<3>, Dtype, bc::host_tag>;

    auto column_tensor = column_tensor_type(
            {kernel_h, kernel_w, channels, output_w, output_h},
            data_col);

    auto image_tensor = image_tensor_type(
            {height, width, channels},
            const_cast<Dtype*>(data_im));

    //TODO add support for strides, and padding
    for (int channel = 0; channel < channels; channel++) {
        for (int image_col = 0; image_col < width-kernel_w+1; image_col++) {
            for (int image_row = 0; image_row < height-kernel_w+1; image_row++) {
                for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
                    for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
                        column_tensor(image_col, image_row, channel, kernel_col, kernel_row)
                                = image_tensor(channel, image_col+kernel_col, image_row+kernel_row);
                    }
                }
            }
        }
    }
}

template <typename Dtype>
void col2im(
        bc::host_tag,
        const Dtype* data_im, const int channels,
        const int height, const int width,
        const int kernel_h, const int kernel_w,

        const int pad_h, const int pad_w,
        const int stride_h, const int stride_w,

        const int dilation_h, const int dilation_w,
        Dtype* data_col) {

    int image_h_end = (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1));
    int image_w_end =  (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1));

    const int output_h = image_h_end / stride_h + 1;
    const int output_w = image_w_end / stride_w + 1;

    using bc::tensors::exprs::Kernel_Array;

    using column_tensor_type = Kernel_Array<bc::Shape<5>, Dtype, bc::host_tag>;
    using image_tensor_type = Kernel_Array<bc::Shape<3>, Dtype, bc::host_tag>;

    auto column_tensor = column_tensor_type(
            {kernel_h, kernel_w, channels, output_w, output_h},
            data_col);

    auto image_tensor = image_tensor_type(
            {height, width, channels},
            const_cast<Dtype*>(data_im));

    //TODO add support for strides, and padding
    for (int channel = 0; channel < channels; channel++) {
        for (int image_col = 0; image_col < width-kernel_w+1; image_col++) {
            for (int image_row = 0; image_row < height-kernel_w+1; image_row++) {
                for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
                    for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
                        image_tensor(channel, image_col+kernel_col, image_row+kernel_row)
                                += column_tensor(image_col, image_row, channel, kernel_col, kernel_row);
                    }
                }
            }
        }
    }
}


}
}
}



#endif /* BC_IM2COL_H_ */