Implement XML export

This allows to save the neural network
once it has been trained.

include/cx.h

@@ -1,6 +1,8 @@
 #ifndef CX_H
 #define CX_H
 
+#define __STDC_WANT_IEC_60559_BFP_EXT__
+
 // Include standard headers
 #include <stdio.h>
 #include <stdlib.h>
@@ -10,6 +12,8 @@
 #include <unistd.h>
 #include <stdint.h>
 #include <pthread.h>
+#include <inttypes.h>
+#include <string.h>
 
 // Include GLEW
 #include <GL/glew.h>

include/model.h

@@ -22,7 +22,7 @@ int modelRegistry_register(ModelRegistry *, Model *);
 void modelRegistry_free(ModelRegistry *);
 GLfloat * model_applyTransformations(Model *);
 void model_colorFromPosition(Model *);
-void model_colorXYZ(Model *, int R, int G, int B);
+void model_colorXYZ(Model *, float R, float G, float B);
 void model_colorRed(Model *);
 void model_colorGreen(Model *);
 void model_colorBlue(Model *);

include/neural.h

@@ -32,6 +32,7 @@ float *neural_loadData(Neural_Network *, const char *);
 float *neural_process(Neural_Network *, float *);
 Neural_Data *neural_getData(Neural_Network *, size_t);
 int neural_getMesh(Neural_Network *, ModelRegistry *);
+char *neural_getXML(Neural_Network *);
 
 #endif
 

src/cx.c

@@ -279,26 +279,22 @@ cx_nnthread(void *self) {
     CX_Thread *self_t = self;
     CX_NN_CTX *nn_ctx = self_t->ctx;
     float *input, *output;
+    char *export;
 
     cx_nninit(&nn_ctx->nn);
     input = neural_loadData(nn_ctx->nn, "../training_data/0");
 
-    for (int i = 0; i < 64; i++) {
-        nn_ctx->nn->layers[0]->neurons[i].value = input[i];
-    }
-
     output = neural_process(nn_ctx->nn, input);
 
-    for (int i = 0; i < 4; i++) {
-        nn_ctx->nn->layers[7]->neurons[i].value = output[i];
-    }
+    export = neural_getXML(nn_ctx->nn);
 
-    return NULL;
+    return export;
 }
 
 int
 cx_run(CX_Context *ctx) {
     CX_ThreadGroup *tg[2];
+    void *neural_xml;
 
     // Establish a model registry
     ctx->gl_ctx->mr = modelRegistry_new();
@@ -308,7 +304,7 @@ cx_run(CX_Context *ctx) {
 
     tg[1] = cx_threadGroup_new(&cx_nnthread, ctx->nn_ctx);
 
-    pthread_join(tg[1]->group_manager->thread, NULL);
+    pthread_join(tg[1]->group_manager->thread, &neural_xml);
 
     ctx->gl_ctx->master_lock = 0;
 
@@ -323,6 +319,7 @@ cx_run(CX_Context *ctx) {
 
     free(ctx->threads);
     free(ctx);
+    free(neural_xml);
 
     return 0;
 }

src/main.c

@@ -4,7 +4,6 @@ int
 main(void) {
     // CX context (Window, neural network, threads.)
     CX_Context *cx_ctx;
-
     int retval;
 
     if (cx_init(&cx_ctx)) {

src/model.c

@@ -129,7 +129,7 @@ model_colorFromPosition(Model *self) {
     }
 }
 
-void model_colorXYZ(Model *self, int R, int G, int B) {
+void model_colorXYZ(Model *self, float R, float G, float B) {
     for (int i = 0; i < self->bufsize; i++) {
         for (int j = 0; j < 4; j++) {
             switch(j) {

src/neural.c

@@ -73,7 +73,6 @@ neural_randomize(Neural_Network *self) {
     Neural_Layer *nl;
     uint64_t *rand_vals;
 
-
     f = fopen("/dev/urandom", "r");
 
     for (int i = 0; i < self->layer_count; i++) {
@@ -83,7 +82,7 @@ neural_randomize(Neural_Network *self) {
             fread(rand_vals, sizeof(uint64_t),
               nl->layer_size_next, f);
             for (int k = 0; k < nl->layer_size_next; k++) {
-                nl->neurons[j].synapses[k] = (float)rand_vals[k] / UINT64_MAX;
+                nl->neurons[j].synapses[k] = (float)rand_vals[k] / UINT64_MAX / nl->layer_size;
             }
             free(rand_vals);
         }
@@ -135,8 +134,8 @@ neural_process(Neural_Network *self, float *input) {
     for (int i = 0; i < self->layers[0]->layer_size; i++) {
         nl->neurons[i].value = input[i];
     }
-    neural_vector = tensor_new(1, nl->layer_size, 0);
     for (int i = 0; i < self->layer_count; i++) {
+        neural_vector = tensor_new(nl->layer_size, 1, 0);
         nl = self->layers[i];
         synapse_matrix = tensor_new(nl->layer_size_next, nl->layer_size, 0);
         for (int j = 0; j < nl->layer_size; j++) {
@@ -147,9 +146,16 @@ neural_process(Neural_Network *self, float *input) {
         }
 
         temp_buffer = tensor_multip(synapse_matrix, neural_vector);
+        neural_vector = temp_buffer;
+        if (nl->layer_size_next) {
+            Neural_Layer *nl_next = self->layers[i+1];
+            for (int j = 0; j < nl_next->layer_size; j++) {
+                nl_next->neurons[j].value = neural_vector->data[j];
+            }
+
+        }
         tensor_free(neural_vector);
         tensor_free(synapse_matrix);
-        neural_vector = temp_buffer;
     }
 
     retval = malloc(nl->layer_size * sizeof(float));
@@ -160,48 +166,81 @@ neural_process(Neural_Network *self, float *input) {
     return retval;
 }
 
-// These two will be merged into one once I have
-// enough patience to create more dynamic objects.
 static void *
-neural_backprop_up(Neural_Network *self, size_t neuron, size_t layer) {
-    return NULL;
+neural_backpropagation(Neural_Network *self, int neuron, int layer, float ratio) {
+    Neural_Layer *nl;
+    Neural_Data *nd;
+    float *ratios;
+    int *neurons;
+    float *synapses;
+
+
+    for (int i = layer-1; i >= 0; i--) {
+        nl = self->layers[i];
+        for (int j = 0; j < nl->layer_size; j++) {
+            synapses = nl->neurons[j].synapses;
+            for (int k = 0; k < nl->layer_size_next; i++) {
+                synapses[k] = 0;
+            }
+        }
+
     }
 
-static void *
-neural_backprop_down(Neural_Network *self, size_t neuron, size_t layer) {
     return NULL;
 }
 
 int
 neural_train(Neural_Network *self,
+             const char *input_path,
              const float *expected_result) {
-    Neural_Data *input_data; // What the neural network received
     Neural_Data *result_data; // What the neural network computed
+    float backprop_ratio;
 
-    input_data = neural_getData(self, 0);
-    result_data = neural_getData(self, self->layer_count-1);
+    for (int i = self->layer_count-1; i >= 0; i--) {
+        Neural_Layer *nl = self->layers[i];
+        result_data = neural_getData(self, i);
+
+        for (int j = nl->layer_size-1; j >= 0; j--) {
+            backprop_ratio = nl->neurons[i].value / expected_result[i];
+            neural_backpropagation(self, j, i, backprop_ratio);
+        }
+    }
 
     return 0;
 }
 
+Neural_Data *
+neural_data_new(int layer_size, int layer_size_next) {
+    Neural_Data *self;
+
+    self = calloc(1, sizeof(Neural_Data));
+    self->neural_vector = malloc(layer_size * sizeof(float));
+    self->vect_len = layer_size;
+
+    if (layer_size_next) {
+        self->synapse_matrix = malloc(layer_size * layer_size_next
+                                        * sizeof(float));
+        self->mat_len = layer_size_next;
+    }
+    return self;
+}
+
 Neural_Data *
 neural_getData(Neural_Network *self, size_t layer) {
     Neural_Layer *nl;
     Neural_Data *retval;
 
-    retval = malloc(1 * sizeof(Neural_Data));
 
     nl = self->layers[layer];
 
-    retval->neural_vector = malloc(nl->layer_size * sizeof(float));
+    retval = neural_data_new(nl->layer_size, nl->layer_size_next);
+
     retval->vect_len = nl->layer_size;
     if (!nl->layer_size_next) {
         retval->synapse_matrix = NULL;
         retval->mat_len = 0;
     }
     else {
-        retval->synapse_matrix = malloc(nl->layer_size * nl->layer_size_next
-                                        * sizeof(float));
         for (int i = 0; i < nl->layer_size; i++) {
             for (int j = 0; j < nl->layer_size_next; j++) {
                 retval->synapse_matrix[i*j+i] = nl->neurons[i].synapses[j];
@@ -222,7 +261,7 @@ neural_getMesh(Neural_Network *nn, ModelRegistry *mr) {
     for (int j = 0; j < nn->layer_count; j++) {
         Neural_Layer *nl = nn->layers[j];
         for (int i = 0; i < nl->layer_size; i++) {
-            unsigned int brightness;
+            float brightness;
             for (int k = 0; k < nl->layer_size_next; k++) {
                 model = model_line((-.90)
                                         + ((GLfloat)2 * i * .90/(nl->layer_size-1)),
@@ -236,7 +275,7 @@ neural_getMesh(Neural_Network *nn, ModelRegistry *mr) {
 
                                   .001 // girth
                                   );
-                brightness =  nl->neurons[i].synapses[k] * 255;
+                brightness =  nl->neurons[i].synapses[k];
                 if (brightness) {
                     model_colorXYZ(model, brightness, 0, 0);
                 }
@@ -245,7 +284,7 @@ neural_getMesh(Neural_Network *nn, ModelRegistry *mr) {
 
             model = model_circle(0, (GLfloat)1/64);
             brightness = nl->neurons[i].value <= 1.0 ?
-                         nl->neurons[i].value : 255;
+                         nl->neurons[i].value : 1.0;
             model_colorXYZ(model, 0, brightness, 0);
             Tensor *translation_matrix = tensor_new(4, 4, 1);
             Tensor *aspectRatio_matrix = tensor_new(4, 4, 1);
@@ -270,3 +309,122 @@ neural_getMesh(Neural_Network *nn, ModelRegistry *mr) {
     return 0;
 }
 
+static char*
+indented_line(char *str, const char *line, int *indent) {
+    for (int m = 0; m < *indent; m++)
+        str = strcat(str, " ");
+    str = strcat(str, line);
+
+    return str;
+
+
+}
+
+static char*
+indented_tag(char *str, const char *tag, int *indent) {
+    if (tag[1] == '/') {
+        *indent -= 4;
+    }
+
+    indented_line(str, tag, indent);
+
+    if (tag[1] != '/') {
+        *indent += 4;
+    }
+
+    return str;
+}
+
+// TODO
+/* This XML implementation has potential bugs and has not
+ * been checked very thoroughly, fix, please.
+ */
+char *
+neural_getXML(Neural_Network *nn) {
+    char *retval;
+    const char *to_write;
+    int volume = 0;
+    int indent = 0;
+
+    retval = malloc(0xff * sizeof(char));
+
+    to_write =  "<?xml version=\"1.0\"?>\n\n";
+    retval = strcpy(retval, to_write);
+    to_write = "<Network>\n";
+    retval = indented_tag(retval, to_write, &indent);
+
+    for (int i = 0; i < nn->layer_count; i++) {
+        Neural_Layer *nl;
+        Neural_Data *nd;
+        char *line_prep;
+
+        nl = nn->layers[i];
+        nd = neural_getData(nn, i);
+
+        retval = realloc(retval, strlen(retval) 
+                                 + (nl->layer_size * 32 * nl->layer_size_next)// Matrix
+                                 + (nl->layer_size * 32)                      // Vector
+                                 + 0x3ff * nl->layer_size                     // Expected tag garbage.
+                                 + indent);                                   // Space waster
+
+        to_write = "<Layer>\n";
+        retval = indented_tag(retval, to_write, &indent);
+
+        to_write = "<Synapse_Matrix>\n";
+        retval = indented_tag(retval, to_write, &indent);
+        for (int j = 0; j < nd->mat_len; j++) {
+            char number_buffer[32];
+            line_prep = malloc((nl->layer_size * 32 * nl->layer_size_next) // Matrix
+                             + (nl->layer_size * 32));
+            *line_prep = '\0';
+
+            line_prep = strcat(line_prep, "[ ");
+            for (int k = 0; k < nd->vect_len; k++) {
+
+                strfromf(number_buffer, 32, "%.2f ", nd->synapse_matrix[k+j*nd->mat_len]);
+                line_prep = strcat(line_prep, number_buffer);
+                if (k < nd->vect_len - 1) {
+                    line_prep = strcat(line_prep, ", ");
+                }
+
+            }
+            line_prep = strcat(line_prep, " ]\n");
+            retval = indented_line(retval, line_prep, &indent);
+            free(line_prep);
+        }
+        to_write = "</Synapse_Matrix>\n";
+        retval = indented_tag(retval, to_write, &indent);
+
+
+        to_write = "<Neural_Vector>\n";
+        retval = indented_tag(retval, to_write, &indent);
+        char number_buffer[32];
+        line_prep = malloc((nl->layer_size * 32 * nl->layer_size_next) // Matrix
+                         + (nl->layer_size * 32));
+        *line_prep = '\0';
+        line_prep = strcat(line_prep, "[ ");
+
+        for (int k = 0; k < nd->vect_len; k++) {
+            strfromf(number_buffer, 32, "%.4f", nd->neural_vector[k]);
+            line_prep = strcat(line_prep, number_buffer);
+
+            if (k < nd->vect_len - 1) {
+                line_prep = strcat(line_prep, ", ");
+            }
+
+        }
+        line_prep = strcat(line_prep, " ]\n");
+        retval = indented_line(retval, line_prep, &indent);
+        free(line_prep);
+        to_write = "</Neural_Vector>\n";
+        retval = indented_tag(retval, to_write, &indent);
+
+        to_write = "</Layer>\n";
+        retval = indented_tag(retval, to_write, &indent);
+    }
+    to_write = "</Network>\n";
+    retval = indented_tag(retval, to_write, &indent);
+
+    return retval;
+}
+