jacobian

commit 74c8a4630056abd6ee6df40935fe391bece52f4f
parent 757192251eee3763c010b0bfb80b2912eaf0f2ca
Author: David Freifeld <freifeld.david@gmail.com>
Date:   Mon, 21 Sep 2020 11:40:14 -0700

Switched to faster file read, 30% overall speedup

Bug spotted, random stopping at epochs

Diffstat:
M
src/bpnn.cpp
 | 
20
+-------------------
M
src/bpnn.hpp
 | 
4
++++
M
src/data.cpp
 | 
20
+++++++++++++++++---
M
src/utils.cpp
 | 
14
--------------
M
src/utils.hpp
 | 
2
+-

5 files changed, 23 insertions(+), 37 deletions(-)
diff --git a/src/bpnn.cpp b/src/bpnn.cpp
@@ -339,24 +339,7 @@ float Network::validate(char* path)
     float costsum = 0;
     float accsum = 0;
     for (int i = 0; i <= val_instances-batch_size; i+=batch_size) {
-        char line[MAXLINE];
-        int inputs = layers[0].contents->cols();
-        int datalen = batch_size * inputs;
-        float batch[datalen];
-        int label = -1;
-        for (int i = 0; i < batch_size; i++) {
-
-            fgets(line, MAXLINE, val_data);
-            char *p;
-            p = strtok(line,",");
-            for (int j = 0; j < inputs; j++) {
-                batch[j + (i * inputs)] = strtod(p, NULL);
-                p = strtok(NULL,",");
-            }
-            (*labels)(i, 0) = strtod(p, NULL);
-        }
-        float* batchptr = batch;
-        update_layer(batchptr, datalen, 0);
+        next_batch();
         feedforward();
         costsum += cost();
         accsum += accuracy();
@@ -369,7 +352,6 @@ float Network::validate(char* path)
 
 void Network::train()
 {
-    rewind(data);
     float cost_sum = 0;
     float acc_sum = 0;
     for (int i = 0; i <= instances-batch_size; i+=batch_size) {
diff --git a/src/bpnn.hpp b/src/bpnn.hpp
@@ -14,6 +14,10 @@
 #include <fstream>
 #include <random>
 #include <algorithm>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <unistd.h>
 
 #define BUFFER_SIZE 16*1024
 enum Regularization {L1, L2};
diff --git a/src/data.cpp b/src/data.cpp
@@ -1,14 +1,28 @@
+typedef float val_t;
+inline float scan(char **p)
+{
+    float n;
+    int neg = 1;
+    while (!isdigit(**p) && **p != '-' && **p != '.') ++*p;
+    if (**p == '-') neg = -1, ++*p;
+    for (n=0; isdigit(**p); ++*p) (n *= 10) += (**p-'0');
+    if (*(*p)++ != '.') return n*neg;
+    float d = 1;
+    for (; isdigit(**p); ++*p) n += (d /= 10) * (**p-'0');
+    return n*neg;
+}
+
 int Network::next_batch()
 {
-    uintmax_t lines = 0
-    while(size_t bytes_read = read(fd, buf, BUFFER_SIZE))
+    uintmax_t lines = 0;
+    while(size_t bytes_read = read(data, buf, BUFFER_SIZE))
     {
         if (!bytes_read) break;
         for(char *p = buf; lines < 10;) {
             char* bound = (char*) memchr(p, '\n', (buf + bytes_read) - p);
             if (bound - p < 0) break; // Stop.
             for (int i=0; i<layers[0].contents->cols(); ++i) (*layers[0].contents)(lines,i) = scan(&p);
-            labels (lines,0) = scan(&p);
+            (*labels)(lines,0) = scan(&p);
             p = bound + 1;
             ++lines;
         }
diff --git a/src/utils.cpp b/src/utils.cpp
@@ -88,20 +88,6 @@ std::function<float(float)> rectifier(float (*activation)(float))
     return rectified;
 }
 
-typedef float val_t;
-inline float scan(char **p)
-{
-    float n;
-    int neg = 1;
-    while (!isdigit(**p) && **p != '-' && **p != '.') ++*p;
-    if (**p == '-') neg = -1, ++*p;
-    for (n=0; isdigit(**p); ++*p) (n *= 10) += (**p-'0');
-    if (*(*p)++ != '.') return n*neg;
-    float d = 1;
-    for (; isdigit(**p); ++*p) n += (d /= 10) * (**p-'0');
-    return n*neg;
-}
-
 #if (AVX)
 // Intel intrinsics for the win!
 Eigen::MatrixXf avx_product(Eigen::MatrixXf a, Eigen::MatrixXf b)
diff --git a/src/utils.hpp b/src/utils.hpp
@@ -46,7 +46,7 @@ Eigen::MatrixXf avx_cpow(Eigen::MatrixXf m, float exponent);
 Eigen::MatrixXf avx_tanh(Eigen::MatrixXf m);
 Eigen::MatrixXf avx_cosh(Eigen::MatrixXf m);
 
-inline float scan(char **p);
 Eigen::MatrixXf strassen_mul(Eigen::MatrixXf a, Eigen::MatrixXf b);
 
 #endif /* MODULE_H */
+