neural_layer_maxpool.c

Go to the documentation of this file.

/*
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 */
 
#include "neural_layer_maxpool.h"
#include "neural_activations.h"
#include "utils.h"
#include <float.h>
 
static void
malloc_layer_arrays(struct Layer *l)
{
    layer_guard_outputs(l);
    l->indexes = calloc(l->n_outputs, sizeof(int));
    l->output = calloc(l->n_outputs, sizeof(double));
    l->delta = calloc(l->n_outputs, sizeof(double));
}
 
static void
realloc_layer_arrays(struct Layer *l)
{
    layer_guard_outputs(l);
    l->indexes = realloc(l->indexes, sizeof(int) * l->n_outputs);
    l->output = realloc(l->output, sizeof(double) * l->n_outputs);
    l->delta = realloc(l->delta, sizeof(double) * l->n_outputs);
}
 
void
neural_layer_maxpool_init(struct Layer *l, const struct ArgsLayer *args)
{
    l->height = args->height;
    l->width = args->width;
    l->channels = args->channels;
    l->pad = args->pad;
    l->size = args->size;
    l->stride = args->stride;
    l->out_w = (l->width + l->pad - l->size) / l->stride + 1;
    l->out_h = (l->height + l->pad - l->size) / l->stride + 1;
    l->out_c = l->channels;
    l->n_outputs = l->out_h * l->out_w * l->out_c;
    l->max_outputs = l->n_outputs;
    l->n_inputs = l->height * l->width * l->channels;
    malloc_layer_arrays(l);
}
 
struct Layer *
neural_layer_maxpool_copy(const struct Layer *src)
{
    if (src->type != MAXPOOL) {
        printf("neural_layer_maxpool_copy(): incorrect source layer type\n");
        exit(EXIT_FAILURE);
    }
    struct Layer *l = malloc(sizeof(struct Layer));
    layer_defaults(l);
    l->type = src->type;
    l->layer_vptr = src->layer_vptr;
    l->height = src->height;
    l->width = src->width;
    l->channels = src->channels;
    l->pad = src->pad;
    l->size = src->size;
    l->stride = src->stride;
    l->out_w = src->out_w;
    l->out_h = src->out_h;
    l->out_c = src->out_c;
    l->n_outputs = src->n_outputs;
    l->max_outputs = src->max_outputs;
    l->n_inputs = src->n_inputs;
    malloc_layer_arrays(l);
    return l;
}
 
void
neural_layer_maxpool_free(const struct Layer *l)
{
    free(l->indexes);
    free(l->output);
    free(l->delta);
}
 
void
neural_layer_maxpool_rand(struct Layer *l)
{
    (void) l;
}
 
static int
max_pool(const struct Layer *l, const double *input, const int i, const int j,
         const int k)
{
    const int w_offset = -l->pad / 2;
    const int h_offset = w_offset;
    double max = -DBL_MAX;
    int max_index = -1;
    for (int n = 0; n < l->size; ++n) {
        for (int m = 0; m < l->size; ++m) {
            const int cur_h = h_offset + i * l->stride + n;
            const int cur_w = w_offset + j * l->stride + m;
            const int index = cur_w + l->width * (cur_h + l->height * k);
            if (cur_h >= 0 && cur_h < l->height && cur_w >= 0 &&
                cur_w < l->width && index < l->n_inputs && input[index] > max) {
                max_index = index;
                max = input[index];
            }
        }
    }
    if (max_index < 0 || max_index >= l->n_inputs) {
        printf("max_pool() error: invalid max_index: (%d)\n", max_index);
        layer_print(l, false);
        exit(EXIT_FAILURE);
    }
    return max_index;
}
 
void
neural_layer_maxpool_forward(const struct Layer *l, const struct Net *net,
                             const double *input)
{
    (void) net;
    for (int k = 0; k < l->channels; ++k) {
        for (int i = 0; i < l->out_h; ++i) {
            for (int j = 0; j < l->out_w; ++j) {
                const int out_index = j + l->out_w * (i + l->out_h * k);
                if (out_index < l->n_outputs) {
                    const int max_index = max_pool(l, input, i, j, k);
                    l->indexes[out_index] = max_index;
                    l->output[out_index] = input[max_index];
                }
            }
        }
    }
}
 
void
neural_layer_maxpool_backward(const struct Layer *l, const struct Net *net,
                              const double *input, double *delta)
{
    (void) net;
    (void) input;
    if (delta) {
        for (int i = 0; i < l->n_outputs; ++i) {
            delta[l->indexes[i]] += l->delta[i];
        }
    }
}
 
void
neural_layer_maxpool_update(const struct Layer *l)
{
    (void) l;
}
 
bool
neural_layer_maxpool_mutate(struct Layer *l)
{
    (void) l;
    return false;
}
 
void
neural_layer_maxpool_resize(struct Layer *l, const struct Layer *prev)
{
    const int w = prev->out_w;
    const int h = prev->out_h;
    const int c = prev->out_c;
    l->height = h;
    l->width = w;
    l->channels = c;
    l->n_inputs = h * w * c;
    l->out_w = (w + l->pad - l->size) / l->stride + 1;
    l->out_h = (h + l->pad - l->size) / l->stride + 1;
    l->out_c = c;
    l->n_outputs = l->out_h * l->out_w * l->out_c;
    l->max_outputs = l->n_outputs;
    realloc_layer_arrays(l);
}
 
double *
neural_layer_maxpool_output(const struct Layer *l)
{
    return l->output;
}
 
void
neural_layer_maxpool_print(const struct Layer *l, const bool print_weights)
{
    char *json_str = neural_layer_maxpool_json_export(l, print_weights);
    printf("%s\n", json_str);
    free(json_str);
}
 
char *
neural_layer_maxpool_json_export(const struct Layer *l,
                                 const bool return_weights)
{
    (void) return_weights;
    cJSON *json = cJSON_CreateObject();
    cJSON_AddStringToObject(json, "type", "maxpool");
    cJSON_AddNumberToObject(json, "n_inputs", l->n_inputs);
    cJSON_AddNumberToObject(json, "n_outputs", l->n_outputs);
    cJSON_AddNumberToObject(json, "height", l->height);
    cJSON_AddNumberToObject(json, "width", l->width);
    cJSON_AddNumberToObject(json, "channels", l->channels);
    cJSON_AddNumberToObject(json, "size", l->size);
    cJSON_AddNumberToObject(json, "stride", l->stride);
    cJSON_AddNumberToObject(json, "pad", l->pad);
    cJSON_AddNumberToObject(json, "out_w", l->out_w);
    cJSON_AddNumberToObject(json, "out_h", l->out_h);
    cJSON_AddNumberToObject(json, "out_c", l->out_c);
    char *string = cJSON_Print(json);
    cJSON_Delete(json);
    return string;
}
 
size_t
neural_layer_maxpool_save(const struct Layer *l, FILE *fp)
{
    size_t s = 0;
    s += fwrite(&l->height, sizeof(int), 1, fp);
    s += fwrite(&l->width, sizeof(int), 1, fp);
    s += fwrite(&l->channels, sizeof(int), 1, fp);
    s += fwrite(&l->pad, sizeof(int), 1, fp);
    s += fwrite(&l->out_w, sizeof(int), 1, fp);
    s += fwrite(&l->out_h, sizeof(int), 1, fp);
    s += fwrite(&l->out_c, sizeof(int), 1, fp);
    s += fwrite(&l->n_outputs, sizeof(int), 1, fp);
    s += fwrite(&l->max_outputs, sizeof(int), 1, fp);
    s += fwrite(&l->n_inputs, sizeof(int), 1, fp);
    s += fwrite(&l->size, sizeof(int), 1, fp);
    s += fwrite(&l->stride, sizeof(int), 1, fp);
    return s;
}
 
size_t
neural_layer_maxpool_load(struct Layer *l, FILE *fp)
{
    size_t s = 0;
    s += fread(&l->height, sizeof(int), 1, fp);
    s += fread(&l->width, sizeof(int), 1, fp);
    s += fread(&l->channels, sizeof(int), 1, fp);
    s += fread(&l->pad, sizeof(int), 1, fp);
    s += fread(&l->out_w, sizeof(int), 1, fp);
    s += fread(&l->out_h, sizeof(int), 1, fp);
    s += fread(&l->out_c, sizeof(int), 1, fp);
    s += fread(&l->n_outputs, sizeof(int), 1, fp);
    s += fread(&l->max_outputs, sizeof(int), 1, fp);
    s += fread(&l->n_inputs, sizeof(int), 1, fp);
    s += fread(&l->size, sizeof(int), 1, fp);
    s += fread(&l->stride, sizeof(int), 1, fp);
    malloc_layer_arrays(l);
    return s;
}