aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/bin/CMakeLists.txt3
-rw-r--r--src/bin/mnist/CMakeLists.txt11
-rw-r--r--src/bin/mnist/src/main.c473
-rw-r--r--src/lib/CMakeLists.txt37
-rw-r--r--src/lib/include/neuralnet/matrix.h111
-rw-r--r--src/lib/include/neuralnet/neuralnet.h64
-rw-r--r--src/lib/include/neuralnet/train.h42
-rw-r--r--src/lib/include/neuralnet/types.h3
-rw-r--r--src/lib/src/activation.h21
-rw-r--r--src/lib/src/matrix.c298
-rw-r--r--src/lib/src/neuralnet.c228
-rw-r--r--src/lib/src/neuralnet_impl.h36
-rw-r--r--src/lib/src/train.c346
-rw-r--r--src/lib/test/matrix_test.c350
-rw-r--r--src/lib/test/neuralnet_test.c92
-rw-r--r--src/lib/test/test.h185
-rw-r--r--src/lib/test/test_main.c3
-rw-r--r--src/lib/test/test_util.h22
-rw-r--r--src/lib/test/train_linear_perceptron_non_origin_test.c67
-rw-r--r--src/lib/test/train_linear_perceptron_test.c62
-rw-r--r--src/lib/test/train_sigmoid_test.c66
-rw-r--r--src/lib/test/train_xor_test.c66
22 files changed, 2586 insertions, 0 deletions
diff --git a/src/bin/CMakeLists.txt b/src/bin/CMakeLists.txt
new file mode 100644
index 0000000..051a56f
--- /dev/null
+++ b/src/bin/CMakeLists.txt
@@ -0,0 +1,3 @@
1cmake_minimum_required(VERSION 3.0)
2
3add_subdirectory(mnist)
diff --git a/src/bin/mnist/CMakeLists.txt b/src/bin/mnist/CMakeLists.txt
new file mode 100644
index 0000000..a6c54f2
--- /dev/null
+++ b/src/bin/mnist/CMakeLists.txt
@@ -0,0 +1,11 @@
1cmake_minimum_required(VERSION 3.0)
2
3add_executable(mnist
4 src/main.c)
5
6target_link_libraries(mnist PRIVATE
7 neuralnet
8 bsd
9 z)
10
11target_compile_options(mnist PRIVATE -Wall -Wextra)
diff --git a/src/bin/mnist/src/main.c b/src/bin/mnist/src/main.c
new file mode 100644
index 0000000..4d268ac
--- /dev/null
+++ b/src/bin/mnist/src/main.c
@@ -0,0 +1,473 @@
1#include <neuralnet/matrix.h>
2#include <neuralnet/neuralnet.h>
3#include <neuralnet/train.h>
4
5#include <zlib.h>
6
7#include <assert.h>
8#include <bsd/string.h>
9#include <linux/limits.h>
10#include <math.h>
11#include <stdbool.h>
12#include <stdint.h>
13#include <stdio.h>
14#include <stdlib.h>
15
16static const int TRAIN_ITERATIONS = 100;
17
18static const int32_t IMAGE_FILE_MAGIC = 0x00000803;
19static const int32_t LABEL_FILE_MAGIC = 0x00000801;
20
21// Inputs of 0 cancel weights during training. This value is used to rescale the
22// input pixels from [0,255] to [PIXEL_LOWER_BOUND, 1.0].
23static const double PIXEL_LOWER_BOUND = 0.01;
24
25// Scale the outputs to (0,1) since the sigmoid cannot produce 0 or 1.
26static const double LABEL_LOWER_BOUND = 0.01;
27static const double LABEL_UPPER_BOUND = 0.99;
28
29// Epsilon used to compare R values.
30static const double EPS = 1e-10;
31
32#define min(a,b) ((a) < (b) ? (a) : (b))
33
34typedef struct ImageSet {
35 nnMatrix images; // Images flattened into row vectors of the matrix.
36 nnMatrix labels; // One-hot-encoded labels.
37 int count; // Number of images and labels.
38 int rows; // Rows in an image.
39 int cols; // Columns in an image.
40} ImageSet;
41
42static void usage(const char* argv0) {
43 fprintf(stderr, "Usage: %s <path to mnist files directory> [num images]\n", argv0);
44 fprintf(stderr, "\n");
45 fprintf(stderr, " Use -1 for [num images] to use all the images in the data set\n");
46}
47
48static bool R_eq(R a, R b) {
49 return fabs(a-b) <= EPS;
50}
51
52static void PrintImage(const nnMatrix* images, int rows, int cols, int image_index) {
53 assert(images);
54 assert((0 <= image_index) && (image_index < images->rows));
55
56 // Top line.
57 for (int j = 0; j < cols/2; ++j) {
58 printf(" -");
59 }
60 printf("\n");
61
62 // Image.
63 const R* value = nnMatrixRow(images, image_index);
64 for (int i = 0; i < rows; ++i) {
65 printf("|");
66 for (int j = 0; j < cols; ++j) {
67 if (*value > 0.8) {
68 printf("#");
69 } else if (*value > 0.5) {
70 printf("*");
71 }
72 else if (*value > PIXEL_LOWER_BOUND) {
73 printf(":");
74 } else if (*value == 0.0) {
75 // Values should not be exactly 0, otherwise they cancel out weights
76 // during training.
77 printf("X");
78 } else {
79 printf(" ");
80 }
81 value++;
82 }
83 printf("|\n");
84 }
85
86 // Bottom line.
87 for (int j = 0; j < cols/2; ++j) {
88 printf(" -");
89 }
90 printf("\n");
91}
92
93static void PrintLabel(const nnMatrix* labels, int label_index) {
94 assert(labels);
95 assert((0 <= label_index) && (label_index < labels->rows));
96
97 // Compute the label from the one-hot encoding.
98 const R* value = nnMatrixRow(labels, label_index);
99 int label = -1;
100 for (int i = 0; i < 10; ++i) {
101 if (R_eq(*value++, LABEL_UPPER_BOUND)) {
102 label = i;
103 break;
104 }
105 }
106 assert((0 <= label) && (label <= 9));
107
108 printf("Label: %d ( ", label);
109 value = nnMatrixRow(labels, label_index);
110 for (int i = 0; i < 10; ++i) {
111 printf("%.3f ", *value++);
112 }
113 printf(")\n");
114}
115
116static R lerp(R a, R b, R t) {
117 return a + t*(b-a);
118}
119
120/// Rescales a pixel from [0,255] to [PIXEL_LOWER_BOUND, 1.0].
121static R FormatPixel(uint8_t pixel) {
122 const R value = (R)(pixel) / 255.0 * (1.0 - PIXEL_LOWER_BOUND) + PIXEL_LOWER_BOUND;
123 assert(value >= PIXEL_LOWER_BOUND);
124 assert(value <= 1.0);
125 return value;
126}
127
128/// Rescales a one-hot-encoded label value to (0,1).
129static R FormatLabel(R label) {
130 const R value = lerp(LABEL_LOWER_BOUND, LABEL_UPPER_BOUND, label);
131 assert(value > 0.0);
132 assert(value < 1.0);
133 return value;
134}
135
136static int32_t ReverseEndian32(int32_t x) {
137 const int32_t x0 = x & 0xff;
138 const int32_t x1 = (x >> 8) & 0xff;
139 const int32_t x2 = (x >> 16) & 0xff;
140 const int32_t x3 = (x >> 24) & 0xff;
141 return (x0 << 24) | (x1 << 16) | (x2 << 8) | x3;
142}
143
144static void ImageToMatrix(
145 const uint8_t* pixels, int num_pixels, int row, nnMatrix* images) {
146 assert(pixels);
147 assert(images);
148
149 for (int i = 0; i < num_pixels; ++i) {
150 const R pixel = FormatPixel(pixels[i]);
151 nnMatrixSet(images, row, i, pixel);
152 }
153}
154
155static bool ReadImages(gzFile images_file, int max_num_images, ImageSet* image_set) {
156 assert(images_file != Z_NULL);
157 assert(image_set);
158
159 bool success = false;
160
161 uint8_t* pixels = 0;
162
163 int32_t magic, total_images, rows, cols;
164 if ( (gzread(images_file, (char*)&magic, sizeof(int32_t)) != sizeof(int32_t)) ||
165 (gzread(images_file, (char*)&total_images, sizeof(int32_t)) != sizeof(int32_t)) ||
166 (gzread(images_file, (char*)&rows, sizeof(int32_t)) != sizeof(int32_t)) ||
167 (gzread(images_file, (char*)&cols, sizeof(int32_t)) != sizeof(int32_t)) ) {
168 fprintf(stderr, "Failed to read header\n");
169 goto cleanup;
170 }
171
172 magic = ReverseEndian32(magic);
173 total_images = ReverseEndian32(total_images);
174 rows = ReverseEndian32(rows);
175 cols = ReverseEndian32(cols);
176
177 if (magic != IMAGE_FILE_MAGIC) {
178 fprintf(stderr, "Magic number mismatch. Got %x, expected: %x\n",
179 magic, IMAGE_FILE_MAGIC);
180 goto cleanup;
181 }
182
183 printf("Magic: %.8x\nTotal images: %d\nRows: %d\nCols: %d\n",
184 magic, total_images, rows, cols);
185
186 total_images = max_num_images >= 0 ? min(total_images, max_num_images) : total_images;
187
188 // Images are flattened into single row vectors.
189 const int num_pixels = rows * cols;
190 image_set->images = nnMatrixMake(total_images, num_pixels);
191 image_set->count = total_images;
192 image_set->rows = rows;
193 image_set->cols = cols;
194
195 pixels = calloc(1, num_pixels);
196 if (!pixels) {
197 fprintf(stderr, "Failed to allocate image buffer\n");
198 goto cleanup;
199 }
200
201 for (int i = 0; i < total_images; ++i) {
202 const int bytes_read = gzread(images_file, pixels, num_pixels);
203 if (bytes_read < num_pixels) {
204 fprintf(stderr, "Failed to read image %d\n", i);
205 goto cleanup;
206 }
207 ImageToMatrix(pixels, num_pixels, i, &image_set->images);
208 }
209
210 success = true;
211
212cleanup:
213 if (pixels) {
214 free(pixels);
215 }
216 if (!success) {
217 nnMatrixDel(&image_set->images);
218 }
219 return success;
220}
221
222static void OneHotEncode(const uint8_t* labels_bytes, int num_labels, nnMatrix* labels) {
223 assert(labels_bytes);
224 assert(labels);
225 assert(labels->rows == num_labels);
226 assert(labels->cols == 10);
227
228 static const R one_hot[10][10] = {
229 { 1, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
230 { 0, 1, 0, 0, 0, 0, 0, 0, 0, 0 },
231 { 0, 0, 1, 0, 0, 0, 0, 0, 0, 0 },
232 { 0, 0, 0, 1, 0, 0, 0, 0, 0, 0 },
233 { 0, 0, 0, 0, 1, 0, 0, 0, 0, 0 },
234 { 0, 0, 0, 0, 0, 1, 0, 0, 0, 0 },
235 { 0, 0, 0, 0, 0, 0, 1, 0, 0, 0 },
236 { 0, 0, 0, 0, 0, 0, 0, 1, 0, 0 },
237 { 0, 0, 0, 0, 0, 0, 0, 0, 1, 0 },
238 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 1 },
239 };
240
241 R* value = labels->values;
242
243 for (int i = 0; i < num_labels; ++i) {
244 const uint8_t label = labels_bytes[i];
245 const R* one_hot_value = one_hot[label];
246
247 for (int j = 0; j < 10; ++j) {
248 *value++ = FormatLabel(*one_hot_value++);
249 }
250 }
251}
252
253static int OneHotDecode(const nnMatrix* label_matrix) {
254 assert(label_matrix);
255 assert(label_matrix->cols == 1);
256 assert(label_matrix->rows == 10);
257
258 R max_value = 0;
259 int pos_max = 0;
260 for (int i = 0; i < 10; ++i) {
261 const R value = nnMatrixAt(label_matrix, 0, i);
262 if (value > max_value) {
263 max_value = value;
264 pos_max = i;
265 }
266 }
267 assert(pos_max >= 0);
268 assert(pos_max <= 10);
269 return pos_max;
270}
271
272static bool ReadLabels(gzFile labels_file, int max_num_labels, ImageSet* image_set) {
273 assert(labels_file != Z_NULL);
274 assert(image_set != 0);
275
276 bool success = false;
277
278 uint8_t* labels = 0;
279
280 int32_t magic, total_labels;
281 if ( (gzread(labels_file, (char*)&magic, sizeof(int32_t)) != sizeof(int32_t)) ||
282 (gzread(labels_file, (char*)&total_labels, sizeof(int32_t)) != sizeof(int32_t)) ) {
283 fprintf(stderr, "Failed to read header\n");
284 goto cleanup;
285 }
286
287 magic = ReverseEndian32(magic);
288 total_labels = ReverseEndian32(total_labels);
289
290 if (magic != LABEL_FILE_MAGIC) {
291 fprintf(stderr, "Magic number mismatch. Got %x, expected: %x\n",
292 magic, LABEL_FILE_MAGIC);
293 goto cleanup;
294 }
295
296 printf("Magic: %.8x\nTotal labels: %d\n", magic, total_labels);
297
298 total_labels = max_num_labels >= 0 ? min(total_labels, max_num_labels) : total_labels;
299
300 assert(image_set->count == total_labels);
301
302 // One-hot encoding of labels, 10 values (digits) per label.
303 image_set->labels = nnMatrixMake(total_labels, 10);
304
305 labels = calloc(total_labels, sizeof(uint8_t));
306 if (!labels) {
307 fprintf(stderr, "Failed to allocate labels buffer\n");
308 goto cleanup;
309 }
310
311 if (gzread(labels_file, labels, total_labels * sizeof(uint8_t)) != total_labels) {
312 fprintf(stderr, "Failed to read labels\n");
313 goto cleanup;
314 }
315
316 OneHotEncode(labels, total_labels, &image_set->labels);
317
318 success = true;
319
320cleanup:
321 if (labels) {
322 free(labels);
323 }
324 if (!success) {
325 nnMatrixDel(&image_set->labels);
326 }
327 return success;
328}
329
330int main(int argc, const char** argv) {
331 if (argc < 2) {
332 usage(argv[0]);
333 return 1;
334 }
335
336 bool success = false;
337
338 gzFile train_images_file = Z_NULL;
339 gzFile train_labels_file = Z_NULL;
340 gzFile test_images_file = Z_NULL;
341 gzFile test_labels_file = Z_NULL;
342 ImageSet train_set = { 0 };
343 ImageSet test_set = { 0 };
344 nnNeuralNetwork* net = 0;
345 nnQueryObject* query = 0;
346
347 const char* mnist_files_dir = argv[1];
348 const int max_num_images = argc > 2 ? atoi(argv[2]) : -1;
349
350 char train_labels_path[PATH_MAX];
351 char train_images_path[PATH_MAX];
352 char test_labels_path[PATH_MAX];
353 char test_images_path[PATH_MAX];
354 strlcpy(train_labels_path, mnist_files_dir, PATH_MAX);
355 strlcpy(train_images_path, mnist_files_dir, PATH_MAX);
356 strlcpy(test_labels_path, mnist_files_dir, PATH_MAX);
357 strlcpy(test_images_path, mnist_files_dir, PATH_MAX);
358 strlcat(train_labels_path, "/train-labels-idx1-ubyte.gz", PATH_MAX);
359 strlcat(train_images_path, "/train-images-idx3-ubyte.gz", PATH_MAX);
360 strlcat(test_labels_path, "/t10k-labels-idx1-ubyte.gz", PATH_MAX);
361 strlcat(test_images_path, "/t10k-images-idx3-ubyte.gz", PATH_MAX);
362
363 train_images_file = gzopen(train_images_path, "r");
364 if (train_images_file == Z_NULL) {
365 fprintf(stderr, "Failed to open file: %s\n", train_images_path);
366 goto cleanup;
367 }
368
369 train_labels_file = gzopen(train_labels_path, "r");
370 if (train_labels_file == Z_NULL) {
371 fprintf(stderr, "Failed to open file: %s\n", train_labels_path);
372 goto cleanup;
373 }
374
375 test_images_file = gzopen(test_images_path, "r");
376 if (test_images_file == Z_NULL) {
377 fprintf(stderr, "Failed to open file: %s\n", test_images_path);
378 goto cleanup;
379 }
380
381 test_labels_file = gzopen(test_labels_path, "r");
382 if (test_labels_file == Z_NULL) {
383 fprintf(stderr, "Failed to open file: %s\n", test_labels_path);
384 goto cleanup;
385 }
386
387 if (!ReadImages(train_images_file, max_num_images, &train_set)) {
388 goto cleanup;
389 }
390 if (!ReadLabels(train_labels_file, max_num_images, &train_set)) {
391 goto cleanup;
392 }
393
394 if (!ReadImages(test_images_file, max_num_images, &test_set)) {
395 goto cleanup;
396 }
397 if (!ReadLabels(test_labels_file, max_num_images, &test_set)) {
398 goto cleanup;
399 }
400
401 printf("\nTraining image/label pair examples:\n");
402 for (int i = 0; i < min(3, train_set.images.rows); ++i) {
403 PrintImage(&train_set.images, train_set.rows, train_set.cols, i);
404 PrintLabel(&train_set.labels, i);
405 printf("\n");
406 }
407
408 // Network definition.
409 const int image_size_pixels = train_set.rows * train_set.cols;
410 const int num_layers = 2;
411 const int layer_sizes[3] = { image_size_pixels, 100, 10 };
412 const nnActivation layer_activations[2] = { nnSigmoid, nnSigmoid };
413 if (!(net = nnMakeNet(num_layers, layer_sizes, layer_activations))) {
414 fprintf(stderr, "Failed to create neural network\n");
415 goto cleanup;
416 }
417
418 // Train.
419 printf("Training with up to %d images from the data set\n\n", max_num_images);
420 const nnTrainingParams training_params = {
421 .learning_rate = 0.1,
422 .max_iterations = TRAIN_ITERATIONS,
423 .seed = 0,
424 .weight_init = nnWeightInitNormal,
425 .debug = true,
426 };
427 nnTrain(net, &train_set.images, &train_set.labels, &training_params);
428
429 // Test.
430 int hits = 0;
431 query = nnMakeQueryObject(net, /*num_inputs=*/1);
432 for (int i = 0; i < test_set.count; ++i) {
433 const nnMatrix test_image = nnMatrixBorrowRows(&test_set.images, i, 1);
434 const nnMatrix test_label = nnMatrixBorrowRows(&test_set.labels, i, 1);
435
436 nnQuery(net, query, &test_image);
437
438 const int test_label_expected = OneHotDecode(&test_label);
439 const int test_label_actual = OneHotDecode(nnNetOutputs(query));
440
441 if (test_label_actual == test_label_expected) {
442 ++hits;
443 }
444 }
445 const R hit_ratio = (R)hits / (R)test_set.count;
446 printf("Test images: %d\n", test_set.count);
447 printf("Hits: %d/%d (%.3f%%)\n", hits, test_set.count, hit_ratio*100);
448
449 success = true;
450
451cleanup:
452 if (query) {
453 nnDeleteQueryObject(&query);
454 }
455 if (net) {
456 nnDeleteNet(&net);
457 }
458 nnMatrixDel(&train_set.images);
459 nnMatrixDel(&test_set.images);
460 if (train_images_file != Z_NULL) {
461 gzclose(train_images_file);
462 }
463 if (train_labels_file != Z_NULL) {
464 gzclose(train_labels_file);
465 }
466 if (test_images_file != Z_NULL) {
467 gzclose(test_images_file);
468 }
469 if (test_labels_file != Z_NULL) {
470 gzclose(test_labels_file);
471 }
472 return success ? 0 : 1;
473}
diff --git a/src/lib/CMakeLists.txt b/src/lib/CMakeLists.txt
new file mode 100644
index 0000000..9e0e924
--- /dev/null
+++ b/src/lib/CMakeLists.txt
@@ -0,0 +1,37 @@
1cmake_minimum_required(VERSION 3.0)
2
3# Library
4
5add_library(neuralnet
6 src/matrix.c
7 src/neuralnet.c
8 src/train.c)
9
10target_include_directories(neuralnet PUBLIC
11 include)
12
13target_link_libraries(neuralnet PRIVATE
14 math # System math library.
15 random)
16
17target_compile_options(neuralnet PRIVATE -Wall -Wextra)
18
19# Test
20
21add_executable(neuralnet-test
22 test/matrix_test.c
23 test/neuralnet_test.c
24 test/test_main.c
25 test/train_linear_perceptron_test.c
26 test/train_linear_perceptron_non_origin_test.c
27 test/train_sigmoid_test.c
28 test/train_xor_test.c)
29
30target_link_libraries(neuralnet-test PRIVATE
31 neuralnet)
32
33# So that we can include header files from the private implementation.
34target_include_directories(neuralnet-test PRIVATE
35 src)
36
37target_compile_options(neuralnet-test PRIVATE -DUNIT_TEST -Wall -Wextra)
diff --git a/src/lib/include/neuralnet/matrix.h b/src/lib/include/neuralnet/matrix.h
new file mode 100644
index 0000000..9816b81
--- /dev/null
+++ b/src/lib/include/neuralnet/matrix.h
@@ -0,0 +1,111 @@
1#pragma once
2
3#include <neuralnet/types.h>
4
5#include <assert.h>
6
7/// NxM matrix.
8typedef struct nnMatrix {
9 int rows;
10 int cols;
11 R* values;
12} nnMatrix;
13
14/// Construct a matrix.
15nnMatrix nnMatrixMake(int rows, int cols);
16
17/// Delete a matrix and free its internal memory.
18void nnMatrixDel(nnMatrix*);
19
20/// Move a matrix.
21///
22/// |in| is an empty matrix after the move.
23/// |out| is a matrix like |in| before the move.
24void nnMatrixMove(nnMatrix* in, nnMatrix* out);
25
26/// Deep-copy a matrix.
27void nnMatrixCopy(const nnMatrix* in, nnMatrix* out);
28
29/// Write the matrix values into an array in a row-major fashion.
30void nnMatrixToArray(const nnMatrix* in, R* out);
31
32/// Write the given row of a matrix into an array.
33void nnMatrixRowToArray(const nnMatrix* in, int row, R* out);
34
35/// Copy a column from a source to a target matrix.
36void nnMatrixCopyCol(const nnMatrix* in, nnMatrix* out, int col_in, int col_out);
37
38/// Mutable borrow of a matrix.
39nnMatrix nnMatrixBorrow(nnMatrix* in);
40
41/// Mutable borrow of a subrange of rows of a matrix.
42nnMatrix nnMatrixBorrowRows(nnMatrix* in, int row_start, int num_rows);
43
44/// Initialize the matrix from an array of values.
45///
46/// The array must hold values in a row-major fashion.
47void nnMatrixInit(nnMatrix*, const R* values);
48
49/// Initialize all matrix values to a given constant.
50void nnMatrixInitConstant(nnMatrix*, R value);
51
52/// Multiply two matrices.
53void nnMatrixMul(const nnMatrix* left, const nnMatrix* right, nnMatrix* out);
54
55/// Matrix multiply-add.
56///
57/// out = left + (right * scale)
58void nnMatrixMulAdd(const nnMatrix* left, const nnMatrix* right, R scale, nnMatrix* out);
59
60/// Matrix multiply-subtract.
61///
62/// out = left - (right * scale)
63void nnMatrixMulSub(const nnMatrix* left, const nnMatrix* right, R scale, nnMatrix* out);
64
65/// Hadamard product of two matrices.
66void nnMatrixMulPairs(const nnMatrix* left, const nnMatrix* right, nnMatrix* out);
67
68/// Add two matrices.
69void nnMatrixAdd(const nnMatrix* left, const nnMatrix* right, nnMatrix* out);
70
71/// Subtract two matrices.
72void nnMatrixSub(const nnMatrix* left, const nnMatrix* right, nnMatrix* out);
73
74/// Adds a row vector to all rows of the matrix.
75void nnMatrixAddRow(const nnMatrix* matrix, const nnMatrix* row, nnMatrix* out);
76
77/// Scale a matrix.
78void nnMatrixScale(nnMatrix*, R scale);
79
80/// Transpose a matrix.
81/// |in| must be different than |out|.
82void nnMatrixTranspose(const nnMatrix* in, nnMatrix* out);
83
84/// Threshold the values of a matrix using a greater-than operator.
85///
86/// out[x,y] = 1 if in[x,y] > threshold else 0
87void nnMatrixGt(const nnMatrix* in, R threshold, nnMatrix* out);
88
89/// Return the matrix value at the given row and column.
90static inline R nnMatrixAt(const nnMatrix* matrix, int row, int col) {
91 assert(matrix);
92 return matrix->values[row * matrix->cols + col];
93}
94
95/// Set the matrix value at the given row and column.
96static inline void nnMatrixSet(nnMatrix* matrix, int row, int col, R value) {
97 assert(matrix);
98 matrix->values[row * matrix->cols + col] = value;
99}
100
101/// Return a pointer to the given row in the matrix.
102static inline const R* nnMatrixRow(const nnMatrix* matrix, int row) {
103 assert(matrix);
104 return &matrix->values[row * matrix->cols];
105}
106
107/// Return a mutable pointer to the given row in the matrix.
108static inline R* nnMatrixRow_mut(nnMatrix* matrix, int row) {
109 assert(matrix);
110 return &matrix->values[row * matrix->cols];
111}
diff --git a/src/lib/include/neuralnet/neuralnet.h b/src/lib/include/neuralnet/neuralnet.h
new file mode 100644
index 0000000..1cf1c53
--- /dev/null
+++ b/src/lib/include/neuralnet/neuralnet.h
@@ -0,0 +1,64 @@
1#pragma once
2
3#include <neuralnet/types.h>
4
5typedef struct nnMatrix nnMatrix;
6
7typedef struct nnNeuralNetwork nnNeuralNetwork;
8typedef struct nnQueryObject nnQueryObject;
9
10/// Neuron activation.
11typedef enum nnActivation {
12 nnIdentity,
13 nnSigmoid,
14 nnRelu,
15} nnActivation;
16
17/// Create a network.
18nnNeuralNetwork* nnMakeNet(int num_layers, const int* layer_sizes, const nnActivation* activations);
19
20/// Delete the network and free its internal memory.
21void nnDeleteNet(nnNeuralNetwork**);
22
23/// Set the network's weights.
24void nnSetWeights(nnNeuralNetwork*, const R* weights);
25
26/// Set the network's biases.
27void nnSetBiases(nnNeuralNetwork*, const R* biases);
28
29/// Query the network.
30///
31/// |input| is a matrix of inputs, one row per input and as many columns as the
32/// input's dimension.
33///
34/// The query object's output matrix (see nnQueryOutputs()) is a matrix of
35/// outputs, one row per output and as many columns as the output's dimension.
36void nnQuery(const nnNeuralNetwork*, nnQueryObject*, const nnMatrix* input);
37
38/// Query the network, array version.
39void nnQueryArray(const nnNeuralNetwork*, nnQueryObject*, const R* input, R* output);
40
41/// Create a query object.
42///
43/// The query object holds all the internal memory required to query a network.
44/// Query objects allocate all memory up front so that network queries can run
45/// without additional memory allocation.
46nnQueryObject* nnMakeQueryObject(const nnNeuralNetwork*, int num_inputs);
47
48/// Delete the query object and free its internal memory.
49void nnDeleteQueryObject(nnQueryObject**);
50
51/// Return the outputs of the query.
52const nnMatrix* nnNetOutputs(const nnQueryObject*);
53
54/// Return the network's input size.
55int nnNetInputSize(const nnNeuralNetwork*);
56
57/// Return the network's output size.
58int nnNetOutputSize(const nnNeuralNetwork*);
59
60/// Return the layer's input size.
61int nnLayerInputSize(const nnMatrix* weights);
62
63/// Return the layer's output size.
64int nnLayerOutputSize(const nnMatrix* weights);
diff --git a/src/lib/include/neuralnet/train.h b/src/lib/include/neuralnet/train.h
new file mode 100644
index 0000000..79f8e7b
--- /dev/null
+++ b/src/lib/include/neuralnet/train.h
@@ -0,0 +1,42 @@
1#pragma once
2
3#include <neuralnet/neuralnet.h>
4
5#include <stdbool.h>
6#include <stdint.h>
7
8typedef struct nnMatrix nnMatrix;
9
10/// Weight initialization strategy.
11///
12/// Note that regardless of strategy, a layer's weights are scaled by the
13/// layer's size. This is to avoid saturation when, e.g., using a sigmoid
14/// activation with many inputs. Thus, a (0,1) initialization is really
15/// (0,scale), for example.
16typedef enum nnWeightInitStrategy {
17 nnWeightInit01, // (0,1) range.
18 nnWeightInit11, // (-1,+1) range.
19 nnWeightInitNormal, // Normal distribution.
20} nnWeightInitStrategy;
21
22/// Network training parameters.
23typedef struct nnTrainingParams {
24 R learning_rate;
25 int max_iterations;
26 uint64_t seed;
27 nnWeightInitStrategy weight_init;
28 bool debug;
29} nnTrainingParams;
30
31/// Train the network.
32///
33/// |inputs| is a matrix of inputs, one row per input and as many columns as
34/// the input's dimension.
35///
36/// |targets| is a matrix of targets, one row per target and as many columns as
37/// the target's dimension.
38void nnTrain(
39 nnNeuralNetwork*,
40 const nnMatrix* inputs,
41 const nnMatrix* targets,
42 const nnTrainingParams*);
diff --git a/src/lib/include/neuralnet/types.h b/src/lib/include/neuralnet/types.h
new file mode 100644
index 0000000..e8d3942
--- /dev/null
+++ b/src/lib/include/neuralnet/types.h
@@ -0,0 +1,3 @@
1#pragma once
2
3typedef double R;
diff --git a/src/lib/src/activation.h b/src/lib/src/activation.h
new file mode 100644
index 0000000..42ab73f
--- /dev/null
+++ b/src/lib/src/activation.h
@@ -0,0 +1,21 @@
1#pragma once
2
3#include <neuralnet/types.h>
4
5#include <math.h>
6
7static inline R sigmoid(R x) {
8 return 1. / (1. + exp(-x));
9}
10
11static inline R relu(R x) {
12 return fmax(0, x);
13}
14
15#define NN_MAP_ARRAY(f, in, out, size) \
16 for (int i = 0; i < size; ++i) { \
17 out[i] = f(in[i]); \
18 }
19
20#define sigmoid_array(in, out, size) NN_MAP_ARRAY(sigmoid, in, out, size)
21#define relu_array(in, out, size) NN_MAP_ARRAY(relu, in, out, size)
diff --git a/src/lib/src/matrix.c b/src/lib/src/matrix.c
new file mode 100644
index 0000000..a7a4ce6
--- /dev/null
+++ b/src/lib/src/matrix.c
@@ -0,0 +1,298 @@
1#include <neuralnet/matrix.h>
2
3#include <assert.h>
4#include <stdlib.h>
5#include <string.h>
6
7nnMatrix nnMatrixMake(int rows, int cols) {
8 R* values = calloc(rows * cols, sizeof(R));
9 assert(values != 0);
10
11 return (nnMatrix) {
12 .rows = rows,
13 .cols = cols,
14 .values = values,
15 };
16}
17
18void nnMatrixDel(nnMatrix* matrix) {
19 assert(matrix != 0);
20
21 if (matrix->values != 0) {
22 free(matrix->values);
23 matrix->values = 0;
24 matrix->rows = 0;
25 matrix->cols = 0;
26 }
27}
28
29void nnMatrixMove(nnMatrix* in, nnMatrix* out) {
30 assert(in);
31 assert(out);
32
33 out->rows = in->rows;
34 out->cols = in->cols;
35 out->values = in->values;
36
37 in->rows = 0;
38 in->cols = 0;
39 in->values = 0;
40}
41
42void nnMatrixCopy(const nnMatrix* in, nnMatrix* out) {
43 assert(in);
44 assert(out);
45 assert(in->rows == out->rows);
46 assert(in->cols == out->cols);
47
48 const R* in_value = in->values;
49 R* out_value = out->values;
50
51 for (int i = 0; i < in->rows * in->cols; ++i) {
52 *out_value++ = *in_value++;
53 }
54}
55
56void nnMatrixToArray(const nnMatrix* in, R* out) {
57 assert(in);
58 assert(out);
59
60 const R* values = in->values;
61 for (int i = 0; i < in->rows * in->cols; ++i) {
62 *out++ = *values++;
63 }
64}
65
66void nnMatrixRowToArray(const nnMatrix* in, int row, R* out) {
67 assert(in);
68 assert(out);
69
70 const R* values = in->values + row * in->cols;
71 for (int i = 0; i < in->cols; ++i) {
72 *out++ = *values++;
73 }
74}
75
76void nnMatrixCopyCol(const nnMatrix* in, nnMatrix* out, int col_in, int col_out) {
77 assert(in);
78 assert(out);
79 assert(in->rows == out->rows);
80 assert(col_in < in->cols);
81 assert(col_out < out->cols);
82
83 for (int row = 0; row < in->rows; ++row) {
84 nnMatrixSet(out, row, col_out, nnMatrixAt(in, row, col_in));
85 }
86}
87
88nnMatrix nnMatrixBorrow(nnMatrix* in) {
89 assert(in);
90
91 nnMatrix out;
92 out.rows = in->rows;
93 out.cols = in->cols;
94 out.values = in->values;
95 return out;
96}
97
98nnMatrix nnMatrixBorrowRows(nnMatrix* in, int row_start, int num_rows) {
99 assert(in);
100 assert(row_start < in->rows);
101 assert(row_start + num_rows <= in->rows);
102
103 nnMatrix out;
104 out.rows = num_rows;
105 out.cols = in->cols;
106 out.values = nnMatrixRow_mut(in, row_start);
107 return out;
108}
109
110void nnMatrixInit(nnMatrix* matrix, const R* values) {
111 assert(matrix);
112 assert(values);
113 memcpy(matrix->values, values, matrix->rows * matrix->cols * sizeof(R));
114}
115
116void nnMatrixInitConstant(nnMatrix* matrix, R value) {
117 assert(matrix);
118 for (int i = 0; i < matrix->rows * matrix->cols; ++i) {
119 matrix->values[i] = value;
120 }
121}
122
123void nnMatrixMul(const nnMatrix* left, const nnMatrix* right, nnMatrix* out) {
124 assert(left != 0);
125 assert(right != 0);
126 assert(out != 0);
127 assert(out != left);
128 assert(out != right);
129 assert(left->cols == right->rows);
130 assert(out->rows == left->rows);
131 assert(out->cols == right->cols);
132
133 R* out_value = out->values;
134
135 for (int i = 0; i < left->rows; ++i) {
136 const R* left_row = &left->values[i * left->cols];
137
138 for (int j = 0; j < right->cols; ++j) {
139 const R* right_col = &right->values[j];
140 *out_value = 0;
141
142 // Vector dot product.
143 for (int k = 0; k < left->cols; ++k) {
144 *out_value += left_row[k] * right_col[0];
145 right_col += right->cols; // Next row in the column.
146 }
147
148 out_value++;
149 }
150 }
151}
152
153void nnMatrixMulAdd(const nnMatrix* left, const nnMatrix* right, R scale, nnMatrix* out) {
154 assert(left);
155 assert(right);
156 assert(out);
157 assert(left->rows == right->rows);
158 assert(left->cols == right->cols);
159 assert(left->rows == out->rows);
160 assert(left->cols == out->cols);
161
162 const R* left_value = left->values;
163 const R* right_value = right->values;
164 R* out_value = out->values;
165
166 for (int i = 0; i < left->rows * left->cols; ++i) {
167 *out_value++ = *left_value++ + *right_value++ * scale;
168 }
169}
170
171void nnMatrixMulSub(const nnMatrix* left, const nnMatrix* right, R scale, nnMatrix* out) {
172 assert(left);
173 assert(right);
174 assert(out);
175 assert(left->rows == right->rows);
176 assert(left->cols == right->cols);
177 assert(left->rows == out->rows);
178 assert(left->cols == out->cols);
179
180 const R* left_value = left->values;
181 const R* right_value = right->values;
182 R* out_value = out->values;
183
184 for (int i = 0; i < left->rows * left->cols; ++i) {
185 *out_value++ = *left_value++ - *right_value++ * scale;
186 }
187}
188
189void nnMatrixMulPairs(const nnMatrix* left, const nnMatrix* right, nnMatrix* out) {
190 assert(left != 0);
191 assert(right != 0);
192 assert(out != 0);
193 assert(left->rows == right->rows);
194 assert(left->cols == right->cols);
195 assert(left->rows == out->rows);
196 assert(left->cols == out->cols);
197
198 R* left_value = left->values;
199 R* right_value = right->values;
200 R* out_value = out->values;
201
202 for (int i = 0; i < left->rows * left->cols; ++i) {
203 *out_value++ = *left_value++ * *right_value++;
204 }
205}
206
207void nnMatrixAdd(const nnMatrix* left, const nnMatrix* right, nnMatrix* out) {
208 assert(left);
209 assert(right);
210 assert(out);
211 assert(left->rows == right->rows);
212 assert(left->cols == right->cols);
213 assert(left->rows == out->rows);
214 assert(left->cols == out->cols);
215
216 const R* left_value = left->values;
217 const R* right_value = right->values;
218 R* out_value = out->values;
219
220 for (int i = 0; i < left->rows * left->cols; ++i) {
221 *out_value++ = *left_value++ + *right_value++;
222 }
223}
224
225void nnMatrixSub(const nnMatrix* left, const nnMatrix* right, nnMatrix* out) {
226 assert(left);
227 assert(right);
228 assert(out);
229 assert(left->rows == right->rows);
230 assert(left->cols == right->cols);
231 assert(left->rows == out->rows);
232 assert(left->cols == out->cols);
233
234 const R* left_value = left->values;
235 const R* right_value = right->values;
236 R* out_value = out->values;
237
238 for (int i = 0; i < left->rows * left->cols; ++i) {
239 *out_value++ = *left_value++ - *right_value++;
240 }
241}
242
243void nnMatrixAddRow(const nnMatrix* matrix, const nnMatrix* row, nnMatrix* out) {
244 assert(matrix);
245 assert(row);
246 assert(out);
247 assert(row->rows == 1);
248 assert(matrix->cols == row->cols);
249 assert(matrix->rows == out->rows);
250 assert(matrix->cols == out->cols);
251
252 const R* matrix_value = matrix->values;
253 R* out_value = out->values;
254
255 for (int i = 0; i < matrix->rows; ++i) {
256 const R* row_value = row->values;
257 for (int j = 0; j < row->cols; ++j) {
258 *out_value++ = *matrix_value++ + *row_value++;
259 }
260 }
261}
262
263void nnMatrixScale(nnMatrix* matrix, R scale) {
264 assert(matrix);
265
266 R* value = matrix->values;
267 for (int i = 0; i < matrix->rows * matrix->cols; ++i) {
268 *value++ *= scale;
269 }
270}
271
272void nnMatrixTranspose(const nnMatrix* in, nnMatrix* out) {
273 assert(in);
274 assert(out);
275 assert(in != out);
276 assert(in->rows == out->cols);
277 assert(in->cols == out->rows);
278
279 for (int i = 0; i < in->rows; ++i) {
280 for (int j = 0; j < in->cols; ++j) {
281 nnMatrixSet(out, j, i, nnMatrixAt(in, i, j));
282 }
283 }
284}
285
286void nnMatrixGt(const nnMatrix* in, R threshold, nnMatrix* out) {
287 assert(in);
288 assert(out);
289 assert(in->rows == out->rows);
290 assert(in->cols == out->cols);
291
292 const R* in_value = in->values;
293 R* out_value = out->values;
294
295 for (int i = 0; i < in->rows * in->cols; ++i) {
296 *out_value++ = (*in_value++) > threshold ? 1 : 0;
297 }
298}
diff --git a/src/lib/src/neuralnet.c b/src/lib/src/neuralnet.c
new file mode 100644
index 0000000..cac611a
--- /dev/null
+++ b/src/lib/src/neuralnet.c
@@ -0,0 +1,228 @@
1#include <neuralnet/neuralnet.h>
2
3#include <neuralnet/matrix.h>
4#include "activation.h"
5#include "neuralnet_impl.h"
6
7#include <assert.h>
8#include <stdlib.h>
9
10nnNeuralNetwork* nnMakeNet(int num_layers, const int* layer_sizes, const nnActivation* activations) {
11 assert(num_layers > 0);
12 assert(layer_sizes);
13 assert(activations);
14
15 nnNeuralNetwork* net = calloc(1, sizeof(nnNeuralNetwork));
16 if (net == 0) {
17 return 0;
18 }
19
20 net->num_layers = num_layers;
21
22 net->weights = calloc(num_layers, sizeof(nnMatrix));
23 net->biases = calloc(num_layers, sizeof(nnMatrix));
24 net->activations = calloc(num_layers, sizeof(nnActivation));
25 if ( (net->weights == 0) || (net->biases == 0) || (net->activations == 0) ) {
26 nnDeleteNet(&net);
27 return 0;
28 }
29
30 for (int l = 0; l < num_layers; ++l) {
31 // layer_sizes = { input layer size, first hidden layer size, ...}
32 const int layer_input_size = layer_sizes[l];
33 const int layer_output_size = layer_sizes[l+1];
34
35 // We store the transpose of the weight matrix as written in textbooks.
36 // Our vectors are row vectors and the matrices row-major.
37 const int rows = layer_input_size;
38 const int cols = layer_output_size;
39
40 net->weights[l] = nnMatrixMake(rows, cols);
41 net->biases[l] = nnMatrixMake(1, cols);
42 net->activations[l] = activations[l];
43 }
44
45 return net;
46}
47
48void nnDeleteNet(nnNeuralNetwork** net) {
49 if ( (!net) || (!(*net)) ) {
50 return;
51 }
52 if ((*net)->weights != 0) {
53 for (int l = 0; l < (*net)->num_layers; ++l) {
54 nnMatrixDel(&(*net)->weights[l]);
55 }
56 free((*net)->weights);
57 (*net)->weights = 0;
58 }
59 if ((*net)->biases != 0) {
60 for (int l = 0; l < (*net)->num_layers; ++l) {
61 nnMatrixDel(&(*net)->biases[l]);
62 }
63 free((*net)->biases);
64 (*net)->biases = 0;
65 }
66 if ((*net)->activations) {
67 free((*net)->activations);
68 (*net)->activations = 0;
69 }
70 free(*net);
71 *net = 0;
72}
73
74void nnSetWeights(nnNeuralNetwork* net, const R* weights) {
75 assert(net);
76 assert(weights);
77
78 for (int l = 0; l < net->num_layers; ++l) {
79 nnMatrix* layer_weights = &net->weights[l];
80 R* layer_values = layer_weights->values;
81
82 for (int j = 0; j < layer_weights->rows * layer_weights->cols; ++j) {
83 *layer_values++ = *weights++;
84 }
85 }
86}
87
88void nnSetBiases(nnNeuralNetwork* net, const R* biases) {
89 assert(net);
90 assert(biases);
91
92 for (int l = 0; l < net->num_layers; ++l) {
93 nnMatrix* layer_biases = &net->biases[l];
94 R* layer_values = layer_biases->values;
95
96 for (int j = 0; j < layer_biases->rows * layer_biases->cols; ++j) {
97 *layer_values++ = *biases++;
98 }
99 }
100}
101
102void nnQuery(const nnNeuralNetwork* net, nnQueryObject* query, const nnMatrix* input) {
103 assert(net);
104 assert(query);
105 assert(input);
106 assert(net->num_layers == query->num_layers);
107 assert(input->rows <= query->network_outputs->rows);
108 assert(input->cols == nnNetInputSize(net));
109
110 for (int i = 0; i < input->rows; ++i) {
111 // Not mutating the input, but we need the cast to borrow.
112 nnMatrix input_vector = nnMatrixBorrowRows((nnMatrix*)input, i, 1);
113
114 for (int l = 0; l < net->num_layers; ++l) {
115 const nnMatrix* layer_weights = &net->weights[l];
116 const nnMatrix* layer_biases = &net->biases[l];
117 // Y^T = (W*X)^T = X^T*W^T
118 //
119 // TODO: If we had a row-row matrix multiplication, we could compute:
120 // Y^T = W ** X^T
121 // The row-row multiplication could be more cache-friendly. We just need
122 // to store W as is, without transposing.
123 // We could also rewrite the original Mul function to go row x row,
124 // decomposing the multiplication. Preserving the original meaning of Mul
125 // makes everything clearer.
126 nnMatrix output_vector = nnMatrixBorrowRows(&query->layer_outputs[l], i, 1);
127 nnMatrixMul(&input_vector, layer_weights, &output_vector);
128 nnMatrixAddRow(&output_vector, layer_biases, &output_vector);
129
130 switch (net->activations[l]) {
131 case nnIdentity:
132 break; // Nothing to do for the identity function.
133 case nnSigmoid:
134 sigmoid_array(output_vector.values, output_vector.values, output_vector.cols);
135 break;
136 case nnRelu:
137 relu_array(output_vector.values, output_vector.values, output_vector.cols);
138 break;
139 default:
140 assert(0);
141 }
142
143 input_vector = output_vector; // Borrow.
144 }
145 }
146}
147
148void nnQueryArray(const nnNeuralNetwork* net, nnQueryObject* query, const R* input, R* output) {
149 assert(net);
150 assert(query);
151 assert(input);
152 assert(output);
153 assert(net->num_layers > 0);
154
155 nnMatrix input_vector = nnMatrixMake(net->weights[0].cols, 1);
156 nnMatrixInit(&input_vector, input);
157 nnQuery(net, query, &input_vector);
158 nnMatrixRowToArray(query->network_outputs, 0, output);
159}
160
161nnQueryObject* nnMakeQueryObject(const nnNeuralNetwork* net, int num_inputs) {
162 assert(net);
163 assert(num_inputs > 0);
164 assert(net->num_layers > 0);
165
166 nnQueryObject* query = calloc(1, sizeof(nnQueryObject));
167 if (!query) {
168 return 0;
169 }
170
171 query->num_layers = net->num_layers;
172
173 // Allocate the intermediate layer output matrices.
174 query->layer_outputs = calloc(net->num_layers, sizeof(nnMatrix));
175 if (!query->layer_outputs) {
176 free(query);
177 return 0;
178 }
179 for (int l = 0; l < net->num_layers; ++l) {
180 const nnMatrix* layer_weights = &net->weights[l];
181 const int layer_output_size = nnLayerOutputSize(layer_weights);
182 query->layer_outputs[l] = nnMatrixMake(num_inputs, layer_output_size);
183 }
184 query->network_outputs = &query->layer_outputs[net->num_layers - 1];
185
186 return query;
187}
188
189void nnDeleteQueryObject(nnQueryObject** query) {
190 if ( (!query) || (!(*query)) ) {
191 return;
192 }
193 if ((*query)->layer_outputs != 0) {
194 for (int l = 0; l < (*query)->num_layers; ++l) {
195 nnMatrixDel(&(*query)->layer_outputs[l]);
196 }
197 }
198 free((*query)->layer_outputs);
199 free(*query);
200 *query = 0;
201}
202
203const nnMatrix* nnNetOutputs(const nnQueryObject* query) {
204 assert(query);
205 return query->network_outputs;
206}
207
208int nnNetInputSize(const nnNeuralNetwork* net) {
209 assert(net);
210 assert(net->num_layers > 0);
211 return net->weights[0].rows;
212}
213
214int nnNetOutputSize(const nnNeuralNetwork* net) {
215 assert(net);
216 assert(net->num_layers > 0);
217 return net->weights[net->num_layers - 1].cols;
218}
219
220int nnLayerInputSize(const nnMatrix* weights) {
221 assert(weights);
222 return weights->rows;
223}
224
225int nnLayerOutputSize(const nnMatrix* weights) {
226 assert(weights);
227 return weights->cols;
228}
diff --git a/src/lib/src/neuralnet_impl.h b/src/lib/src/neuralnet_impl.h
new file mode 100644
index 0000000..26107b5
--- /dev/null
+++ b/src/lib/src/neuralnet_impl.h
@@ -0,0 +1,36 @@
1#pragma once
2
3#include <neuralnet/matrix.h>
4
5/// Neural network object.
6///
7/// We store the transposes of the weight matrices so that we can do forward
8/// passes with a minimal amount of work. That is, if in paper we write:
9///
10/// [w11 w21]
11/// [w12 w22]
12///
13/// then the weight matrix in memory is stored as the following array:
14///
15/// w11 w12 w21 w22
16typedef struct nnNeuralNetwork {
17 int num_layers; // Number of non-input layers (hidden + output).
18 nnMatrix* weights; // One matrix per non-input layer.
19 nnMatrix* biases; // One vector per non-input layer.
20 nnActivation* activations; // One per non-input layer.
21} nnNeuralNetwork;
22
23/// A query object that holds all the memory necessary to query a network.
24///
25/// |layer_outputs| is an array of matrices of intermediate layer outputs. There
26/// is one matrix per intermediate layer. Each matrix holds the layer's output,
27/// with one row per input, and as many columns as the layer's output size (the
28/// output vector is transposed.)
29///
30/// |network_outputs| points to the last output matrix in |layer_outputs| for
31/// convenience.
32typedef struct nnQueryObject {
33 int num_layers;
34 nnMatrix* layer_outputs; // Output matrices, one output per layer.
35 nnMatrix* network_outputs; // Points to the last output matrix.
36} nnTrainingQueryObject;
diff --git a/src/lib/src/train.c b/src/lib/src/train.c
new file mode 100644
index 0000000..027de66
--- /dev/null
+++ b/src/lib/src/train.c
@@ -0,0 +1,346 @@
1#include <neuralnet/train.h>
2
3#include <neuralnet/matrix.h>
4#include "neuralnet_impl.h"
5
6#include <random/mt19937-64.h>
7#include <random/normal.h>
8
9#include <assert.h>
10#include <math.h>
11#include <stdlib.h>
12
13#include <stdio.h>
14#define LOGD printf
15
16// If debug mode is requested, we will show progress every this many iterations.
17static const int PROGRESS_THRESHOLD = 5; // %
18
19/// Computes the total MSE from the output error matrix.
20R ComputeMSE(const nnMatrix* errors) {
21 R sum_sq = 0;
22 const int N = errors->rows * errors->cols;
23 const R* value = errors->values;
24 for (int i = 0; i < N; ++i) {
25 sum_sq += *value * *value;
26 value++;
27 }
28 return sum_sq / (R)N;
29}
30
31/// Holds the bits required to compute a sigmoid gradient.
32typedef struct nnSigmoidGradientElements {
33 nnMatrix ones; // A vector of just ones, same size as the layer.
34} nnSigmoidGradientElements;
35
36/// Holds the various elements required to compute gradients. These depend on
37/// what activation function are used, so they'll potentially be different for
38/// each layer. A data type is defined for these because we allocate all the
39/// required memory up front before entering the training loop.
40typedef struct nnGradientElements {
41 nnActivation type;
42 // Gradient vector, same size as the layer.
43 // This will contain the gradient expression except for the output value of
44 // the previous layer.
45 nnMatrix gradient;
46 union {
47 nnSigmoidGradientElements sigmoid;
48 };
49} nnGradientElements;
50
51// Initialize the network's weights randomly and set their biases to 0.
52void nnInitNet(nnNeuralNetwork* net, uint64_t seed, const nnWeightInitStrategy strategy) {
53 assert(net);
54
55 mt19937_64 rng = mt19937_64_make();
56 mt19937_64_init(&rng, seed);
57
58 for (int l = 0; l < net->num_layers; ++l) {
59 nnMatrix* weights = &net->weights[l];
60 nnMatrix* biases = &net->biases[l];
61
62 const R layer_size = (R)nnLayerInputSize(weights);
63 const R scale = 1. / layer_size;
64 const R stdev = 1. / sqrt((R)layer_size);
65 const R sigma = stdev * stdev;
66
67 R* value = weights->values;
68 for (int k = 0; k < weights->rows * weights->cols; ++k) {
69 switch (strategy) {
70 case nnWeightInit01: {
71 const R x01 = mt19937_64_gen_real3(&rng); // (0, +1) interval.
72 *value++ = scale * x01;
73 break;
74 }
75 case nnWeightInit11: {
76 const R x11 = mt19937_64_gen_real4(&rng); // (-1, +1) interval.
77 *value++ = scale * x11;
78 break;
79 }
80 case nnWeightInitNormal:
81 // Using initialization with a normal distribution of standard
82 // deviation 1 / sqrt(num_layer_weights) to prevent saturation when
83 // multiplying inputs.
84 const R u01 = mt19937_64_gen_real3(&rng); // (0, +1) interval.
85 const R v01 = mt19937_64_gen_real3(&rng); // (0, +1) interval.
86 R z0, z1;
87 normal2(u01, v01, &z0, &z1);
88 z0 = normal_transform(z0, /*mu=*/0, sigma);
89 z1 = normal_transform(z1, /*mu=*/0, sigma);
90 *value++ = z0;
91 if (k < weights->rows * weights->cols - 1) {
92 *value++ = z1;
93 ++k;
94 }
95 break;
96 default:
97 assert(false);
98 }
99 }
100
101 // Initialize biases.
102 // 0 is used so that functions originally go through the origin.
103 value = biases->values;
104 for (int k = 0; k < biases->rows * biases->cols; ++k, ++value) {
105 *value = 0;
106 }
107 }
108}
109
110// |inputs| has one row vector per sample.
111// |targets| has one row vector per sample.
112//
113// For now, each iteration trains with one sample (row) at a time.
114void nnTrain(
115 nnNeuralNetwork* net,
116 const nnMatrix* inputs,
117 const nnMatrix* targets,
118 const nnTrainingParams* params) {
119 assert(net);
120 assert(inputs);
121 assert(targets);
122 assert(params);
123 assert(nnNetOutputSize(net) == targets->cols);
124 assert(net->num_layers > 0);
125
126 // Allocate error vectors to hold the backpropagated error values.
127 // For now, these are one row vector per layer, meaning that we will train
128 // with one sample at a time.
129 nnMatrix* errors = calloc(net->num_layers, sizeof(nnMatrix));
130
131 // Allocate the weight transpose matrices up front for backpropagation.
132 nnMatrix* weights_T = calloc(net->num_layers, sizeof(nnMatrix));
133
134 // Allocate the weight delta matrices.
135 nnMatrix* weight_deltas = calloc(net->num_layers, sizeof(nnMatrix));
136
137 // Allocate the data structures required to compute gradients.
138 // This depends on each layer's activation type.
139 nnGradientElements* gradient_elems = calloc(net->num_layers, sizeof(nnGradientElements));
140
141 // Allocate the output transpose vectors for weight delta calculation.
142 // This is one column vector per layer.
143 nnMatrix* outputs_T = calloc(net->num_layers, sizeof(nnMatrix));
144
145 assert(errors != 0);
146 assert(weights_T != 0);
147 assert(weight_deltas != 0);
148 assert(gradient_elems);
149 assert(outputs_T);
150
151 for (int l = 0; l < net->num_layers; ++l) {
152 const nnMatrix* layer_weights = &net->weights[l];
153 const int layer_output_size = net->weights[l].cols;
154 const nnActivation activation = net->activations[l];
155
156 errors[l] = nnMatrixMake(1, layer_weights->cols);
157
158 weights_T[l] = nnMatrixMake(layer_weights->cols, layer_weights->rows);
159 nnMatrixTranspose(layer_weights, &weights_T[l]);
160
161 weight_deltas[l] = nnMatrixMake(layer_weights->rows, layer_weights->cols);
162
163 outputs_T[l] = nnMatrixMake(layer_output_size, 1);
164
165 // Allocate the gradient elements and vectors for weight delta calculation.
166 nnGradientElements* elems = &gradient_elems[l];
167 elems->type = activation;
168 switch (activation) {
169 case nnIdentity:
170 break; // Gradient vector will be borrowed, no need to allocate.
171
172 case nnSigmoid:
173 elems->gradient = nnMatrixMake(1, layer_output_size);
174 // Allocate the 1s vectors.
175 elems->sigmoid.ones = nnMatrixMake(1, layer_output_size);
176 nnMatrixInitConstant(&elems->sigmoid.ones, 1);
177 break;
178
179 case nnRelu:
180 elems->gradient = nnMatrixMake(1, layer_output_size);
181 break;
182 }
183 }
184
185 // Construct the query object with a size of 1 since we are training with one
186 // sample at a time.
187 nnQueryObject* query = nnMakeQueryObject(net, 1);
188
189 // Network outputs are given by the query object. Every network query updates
190 // the outputs.
191 const nnMatrix* const training_outputs = query->network_outputs;
192
193 // A vector to store the training input transposed.
194 nnMatrix training_inputs_T = nnMatrixMake(inputs->cols, 1);
195
196 // If debug mode is requested, we will show progress every Nth iteration.
197 const int progress_frame =
198 (params->max_iterations < PROGRESS_THRESHOLD)
199 ? 1
200 : (params->max_iterations * PROGRESS_THRESHOLD / 100);
201
202 // --- TRAIN
203
204 nnInitNet(net, params->seed, params->weight_init);
205
206 for (int iter = 0; iter < params->max_iterations; ++iter) {
207
208 // For now, we train with one sample at a time.
209 for (int sample = 0; sample < inputs->rows; ++sample) {
210 // Slice the input and target matrices with the batch size.
211 // We are not mutating the inputs, but we need the cast to borrow.
212 nnMatrix training_inputs = nnMatrixBorrowRows((nnMatrix*)inputs, sample, 1);
213 nnMatrix training_targets = nnMatrixBorrowRows((nnMatrix*)targets, sample, 1);
214
215 // Will need the input transposed for backpropagation.
216 // Assuming one training input per iteration for now.
217 nnMatrixTranspose(&training_inputs, &training_inputs_T);
218
219 // Run a forward pass and compute the output layer error.
220 // We don't square the error here; instead, we just compute t-o, which is
221 // part of the derivative, -2(t-o). Also, we compute o-t instead to
222 // remove that outer negative sign.
223 nnQuery(net, query, &training_inputs);
224 //nnMatrixSub(&training_targets, training_outputs, &errors[net->num_layers - 1]);
225 nnMatrixSub(training_outputs, &training_targets, &errors[net->num_layers - 1]);
226
227 // Update outputs_T, which we need during weight updates.
228 for (int l = 0; l < net->num_layers; ++l) {
229 nnMatrixTranspose(&query->layer_outputs[l], &outputs_T[l]);
230 }
231
232 // Update weights and biases for each internal layer, backpropagating
233 // errors along the way.
234 for (int l = net->num_layers - 1; l >= 0; --l) {
235 const nnMatrix* layer_output = &query->layer_outputs[l];
236 nnMatrix* layer_weights = &net->weights[l];
237 nnMatrix* layer_biases = &net->biases[l];
238 nnGradientElements* elems = &gradient_elems[l];
239 nnMatrix* gradient = &elems->gradient;
240 const nnActivation activation = net->activations[l];
241
242 // Compute the gradient (the part of the expression that does not
243 // contain the output of the previous layer).
244 //
245 // Identity: G = error_k
246 // Sigmoid: G = error_k * output_k * (1 - output_k).
247 // Relu: G = error_k * (output_k > 0 ? 1 : 0)
248 switch (activation) {
249 case nnIdentity:
250 // TODO: Just copy the pointer?
251 *gradient = nnMatrixBorrow(&errors[l]);
252 break;
253 case nnSigmoid:
254 nnMatrixSub(&elems->sigmoid.ones, layer_output, gradient);
255 nnMatrixMulPairs(layer_output, gradient, gradient);
256 nnMatrixMulPairs(&errors[l], gradient, gradient);
257 break;
258 case nnRelu:
259 nnMatrixGt(layer_output, 0, gradient);
260 nnMatrixMulPairs(&errors[l], gradient, gradient);
261 break;
262 }
263
264 // Outer product to compute the weight deltas.
265 const nnMatrix* output_T = (l == 0) ? &training_inputs_T : &outputs_T[l-1];
266 nnMatrixMul(output_T, gradient, &weight_deltas[l]);
267
268 // Backpropagate the error before updating weights.
269 if (l > 0) {
270 nnMatrixMul(gradient, &weights_T[l], &errors[l-1]);
271 }
272
273 // Update weights.
274 nnMatrixScale(&weight_deltas[l], params->learning_rate);
275 // The gradient has a negative sign from -(t - o), but we have computed
276 // e = o - t instead, so we can subtract directly.
277 //nnMatrixAdd(layer_weights, &weight_deltas[l], layer_weights);
278 nnMatrixSub(layer_weights, &weight_deltas[l], layer_weights);
279
280 // Update weight transpose matrix for the next training iteration.
281 nnMatrixTranspose(layer_weights, &weights_T[l]);
282
283 // Update biases.
284 // This is the same formula as for weights, except that the o_j term is
285 // just 1. We can simply re-use the gradient that we have already
286 // computed for the weight update.
287 //nnMatrixMulAdd(layer_biases, gradient, params->learning_rate, layer_biases);
288 nnMatrixMulSub(layer_biases, gradient, params->learning_rate, layer_biases);
289 }
290
291 // TODO: Add this under a verbose debugging mode.
292 // if (params->debug) {
293 // LOGD("Iter: %d, Sample: %d, Error: %f\n", iter, sample, ComputeMSE(&errors[net->num_layers - 1]));
294 // LOGD("TGT: ");
295 // for (int i = 0; i < training_targets.cols; ++i) {
296 // printf("%.3f ", training_targets.values[i]);
297 // }
298 // printf("\n");
299 // LOGD("OUT: ");
300 // for (int i = 0; i < training_outputs->cols; ++i) {
301 // printf("%.3f ", training_outputs->values[i]);
302 // }
303 // printf("\n");
304 // }
305 }
306
307 if (params->debug && ((iter % progress_frame) == 0)) {
308 LOGD("Iter: %d/%d, Error: %f\n",
309 iter, params->max_iterations, ComputeMSE(&errors[net->num_layers - 1]));
310 }
311 }
312
313 // Print the final error.
314 if (params->debug) {
315 LOGD("Iter: %d/%d, Error: %f\n",
316 params->max_iterations, params->max_iterations, ComputeMSE(&errors[net->num_layers - 1]));
317 }
318
319 for (int l = 0; l < net->num_layers; ++l) {
320 nnMatrixDel(&errors[l]);
321 nnMatrixDel(&outputs_T[l]);
322 nnMatrixDel(&weights_T[l]);
323 nnMatrixDel(&weight_deltas[l]);
324
325 nnGradientElements* elems = &gradient_elems[l];
326 switch (elems->type) {
327 case nnIdentity:
328 break; // Gradient vector is borrowed, no need to deallocate.
329
330 case nnSigmoid:
331 nnMatrixDel(&elems->gradient);
332 nnMatrixDel(&elems->sigmoid.ones);
333 break;
334
335 case nnRelu:
336 nnMatrixDel(&elems->gradient);
337 break;
338 }
339 }
340 nnMatrixDel(&training_inputs_T);
341 free(errors);
342 free(outputs_T);
343 free(weights_T);
344 free(weight_deltas);
345 free(gradient_elems);
346}
diff --git a/src/lib/test/matrix_test.c b/src/lib/test/matrix_test.c
new file mode 100644
index 0000000..8191c97
--- /dev/null
+++ b/src/lib/test/matrix_test.c
@@ -0,0 +1,350 @@
1#include <neuralnet/matrix.h>
2
3#include "test.h"
4#include "test_util.h"
5
6#include <assert.h>
7#include <stdlib.h>
8
9// static void PrintMatrix(const nnMatrix* matrix) {
10// assert(matrix);
11
12// for (int i = 0; i < matrix->rows; ++i) {
13// for (int j = 0; j < matrix->cols; ++j) {
14// printf("%f ", nnMatrixAt(matrix, i, j));
15// }
16// printf("\n");
17// }
18// }
19
20TEST_CASE(nnMatrixMake_1x1) {
21 nnMatrix A = nnMatrixMake(1, 1);
22 TEST_EQUAL(A.rows, 1);
23 TEST_EQUAL(A.cols, 1);
24}
25
26TEST_CASE(nnMatrixMake_3x1) {
27 nnMatrix A = nnMatrixMake(3, 1);
28 TEST_EQUAL(A.rows, 3);
29 TEST_EQUAL(A.cols, 1);
30}
31
32TEST_CASE(nnMatrixInit_3x1) {
33 nnMatrix A = nnMatrixMake(3, 1);
34 nnMatrixInit(&A, (R[]) { 1, 2, 3 });
35 TEST_EQUAL(A.values[0], 1);
36 TEST_EQUAL(A.values[1], 2);
37 TEST_EQUAL(A.values[2], 3);
38}
39
40TEST_CASE(nnMatrixCopyCol_test) {
41 nnMatrix A = nnMatrixMake(3, 2);
42 nnMatrix B = nnMatrixMake(3, 1);
43
44 nnMatrixInit(&A, (R[]) {
45 1, 2,
46 3, 4,
47 5, 6,
48 });
49
50 nnMatrixCopyCol(&A, &B, 1, 0);
51
52 TEST_EQUAL(nnMatrixAt(&B, 0, 0), 2);
53 TEST_EQUAL(nnMatrixAt(&B, 1, 0), 4);
54 TEST_EQUAL(nnMatrixAt(&B, 2, 0), 6);
55
56 nnMatrixDel(&A);
57 nnMatrixDel(&B);
58}
59
60TEST_CASE(nnMatrixMul_square_3x3) {
61 nnMatrix A = nnMatrixMake(3, 3);
62 nnMatrix B = nnMatrixMake(3, 3);
63 nnMatrix O = nnMatrixMake(3, 3);
64
65 nnMatrixInit(&A, (const R[]){
66 1, 2, 3,
67 4, 5, 6,
68 7, 8, 9,
69 });
70 nnMatrixInit(&B, (const R[]){
71 2, 4, 3,
72 6, 8, 5,
73 1, 7, 9,
74 });
75 nnMatrixMul(&A, &B, &O);
76
77 const R expected[3][3] = {
78 { 17, 41, 40 },
79 { 44, 98, 91 },
80 { 71, 155, 142 },
81 };
82 for (int i = 0; i < O.rows; ++i) {
83 for (int j = 0; j < O.cols; ++j) {
84 TEST_TRUE(double_eq(nnMatrixAt(&O, i, j), expected[i][j], EPS));
85 }
86 }
87
88 nnMatrixDel(&A);
89 nnMatrixDel(&B);
90 nnMatrixDel(&O);
91}
92
93TEST_CASE(nnMatrixMul_non_square_2x3_3x1) {
94 nnMatrix A = nnMatrixMake(2, 3);
95 nnMatrix B = nnMatrixMake(3, 1);
96 nnMatrix O = nnMatrixMake(2, 1);
97
98 nnMatrixInit(&A, (const R[]){
99 1, 2, 3,
100 4, 5, 6,
101 });
102 nnMatrixInit(&B, (const R[]){
103 2,
104 6,
105 1,
106 });
107 nnMatrixMul(&A, &B, &O);
108
109 const R expected[2][1] = {
110 { 17 },
111 { 44 },
112 };
113 for (int i = 0; i < O.rows; ++i) {
114 for (int j = 0; j < O.cols; ++j) {
115 TEST_TRUE(double_eq(nnMatrixAt(&O, i, j), expected[i][j], EPS));
116 }
117 }
118
119 nnMatrixDel(&A);
120 nnMatrixDel(&B);
121 nnMatrixDel(&O);
122}
123
124TEST_CASE(nnMatrixMulAdd_test) {
125 nnMatrix A = nnMatrixMake(2, 3);
126 nnMatrix B = nnMatrixMake(2, 3);
127 nnMatrix O = nnMatrixMake(2, 3);
128 const R scale = 2;
129
130 nnMatrixInit(&A, (const R[]){
131 1, 2, 3,
132 4, 5, 6,
133 });
134 nnMatrixInit(&B, (const R[]){
135 2, 3, 1,
136 7, 4, 3
137 });
138 nnMatrixMulAdd(&A, &B, scale, &O); // O = A + B * scale
139
140 const R expected[2][3] = {
141 { 5, 8, 5 },
142 { 18, 13, 12 },
143 };
144 for (int i = 0; i < O.rows; ++i) {
145 for (int j = 0; j < O.cols; ++j) {
146 TEST_TRUE(double_eq(nnMatrixAt(&O, i, j), expected[i][j], EPS));
147 }
148 }
149
150 nnMatrixDel(&A);
151 nnMatrixDel(&B);
152 nnMatrixDel(&O);
153}
154
155TEST_CASE(nnMatrixMulSub_test) {
156 nnMatrix A = nnMatrixMake(2, 3);
157 nnMatrix B = nnMatrixMake(2, 3);
158 nnMatrix O = nnMatrixMake(2, 3);
159 const R scale = 2;
160
161 nnMatrixInit(&A, (const R[]){
162 1, 2, 3,
163 4, 5, 6,
164 });
165 nnMatrixInit(&B, (const R[]){
166 2, 3, 1,
167 7, 4, 3
168 });
169 nnMatrixMulSub(&A, &B, scale, &O); // O = A - B * scale
170
171 const R expected[2][3] = {
172 { -3, -4, 1 },
173 { -10, -3, 0 },
174 };
175 for (int i = 0; i < O.rows; ++i) {
176 for (int j = 0; j < O.cols; ++j) {
177 TEST_TRUE(double_eq(nnMatrixAt(&O, i, j), expected[i][j], EPS));
178 }
179 }
180
181 nnMatrixDel(&A);
182 nnMatrixDel(&B);
183 nnMatrixDel(&O);
184}
185
186TEST_CASE(nnMatrixMulPairs_2x3) {
187 nnMatrix A = nnMatrixMake(2, 3);
188 nnMatrix B = nnMatrixMake(2, 3);
189 nnMatrix O = nnMatrixMake(2, 3);
190
191 nnMatrixInit(&A, (const R[]){
192 1, 2, 3,
193 4, 5, 6,
194 });
195 nnMatrixInit(&B, (const R[]){
196 2, 3, 1,
197 7, 4, 3
198 });
199 nnMatrixMulPairs(&A, &B, &O);
200
201 const R expected[2][3] = {
202 { 2, 6, 3 },
203 { 28, 20, 18 },
204 };
205 for (int i = 0; i < O.rows; ++i) {
206 for (int j = 0; j < O.cols; ++j) {
207 TEST_TRUE(double_eq(nnMatrixAt(&O, i, j), expected[i][j], EPS));
208 }
209 }
210
211 nnMatrixDel(&A);
212 nnMatrixDel(&B);
213 nnMatrixDel(&O);
214}
215
216TEST_CASE(nnMatrixAdd_square_2x2) {
217 nnMatrix A = nnMatrixMake(2, 2);
218 nnMatrix B = nnMatrixMake(2, 2);
219 nnMatrix C = nnMatrixMake(2, 2);
220
221 nnMatrixInit(&A, (R[]) {
222 1, 2,
223 3, 4,
224 });
225 nnMatrixInit(&B, (R[]) {
226 2, 1,
227 5, 3,
228 });
229
230 nnMatrixAdd(&A, &B, &C);
231
232 TEST_TRUE(double_eq(nnMatrixAt(&C, 0, 0), 3, EPS));
233 TEST_TRUE(double_eq(nnMatrixAt(&C, 0, 1), 3, EPS));
234 TEST_TRUE(double_eq(nnMatrixAt(&C, 1, 0), 8, EPS));
235 TEST_TRUE(double_eq(nnMatrixAt(&C, 1, 1), 7, EPS));
236
237 nnMatrixDel(&A);
238 nnMatrixDel(&B);
239 nnMatrixDel(&C);
240}
241
242TEST_CASE(nnMatrixSub_square_2x2) {
243 nnMatrix A = nnMatrixMake(2, 2);
244 nnMatrix B = nnMatrixMake(2, 2);
245 nnMatrix C = nnMatrixMake(2, 2);
246
247 nnMatrixInit(&A, (R[]) {
248 1, 2,
249 3, 4,
250 });
251 nnMatrixInit(&B, (R[]) {
252 2, 1,
253 5, 3,
254 });
255
256 nnMatrixSub(&A, &B, &C);
257
258 TEST_TRUE(double_eq(nnMatrixAt(&C, 0, 0), -1, EPS));
259 TEST_TRUE(double_eq(nnMatrixAt(&C, 0, 1), +1, EPS));
260 TEST_TRUE(double_eq(nnMatrixAt(&C, 1, 0), -2, EPS));
261 TEST_TRUE(double_eq(nnMatrixAt(&C, 1, 1), +1, EPS));
262
263 nnMatrixDel(&A);
264 nnMatrixDel(&B);
265 nnMatrixDel(&C);
266}
267
268TEST_CASE(nnMatrixAddRow_test) {
269 nnMatrix A = nnMatrixMake(2, 3);
270 nnMatrix B = nnMatrixMake(1, 3);
271 nnMatrix C = nnMatrixMake(2, 3);
272
273 nnMatrixInit(&A, (R[]) {
274 1, 2, 3,
275 4, 5, 6,
276 });
277 nnMatrixInit(&B, (R[]) {
278 2, 1, 3,
279 });
280
281 nnMatrixAddRow(&A, &B, &C);
282
283 TEST_TRUE(double_eq(nnMatrixAt(&C, 0, 0), 3, EPS));
284 TEST_TRUE(double_eq(nnMatrixAt(&C, 0, 1), 3, EPS));
285 TEST_TRUE(double_eq(nnMatrixAt(&C, 0, 2), 6, EPS));
286 TEST_TRUE(double_eq(nnMatrixAt(&C, 1, 0), 6, EPS));
287 TEST_TRUE(double_eq(nnMatrixAt(&C, 1, 1), 6, EPS));
288 TEST_TRUE(double_eq(nnMatrixAt(&C, 1, 2), 9, EPS));
289
290 nnMatrixDel(&A);
291 nnMatrixDel(&B);
292 nnMatrixDel(&C);
293}
294
295TEST_CASE(nnMatrixTranspose_square_2x2) {
296 nnMatrix A = nnMatrixMake(2, 2);
297 nnMatrix B = nnMatrixMake(2, 2);
298
299 nnMatrixInit(&A, (R[]) {
300 1, 2,
301 3, 4
302 });
303
304 nnMatrixTranspose(&A, &B);
305 TEST_TRUE(double_eq(nnMatrixAt(&B, 0, 0), 1, EPS));
306 TEST_TRUE(double_eq(nnMatrixAt(&B, 0, 1), 3, EPS));
307 TEST_TRUE(double_eq(nnMatrixAt(&B, 1, 0), 2, EPS));
308 TEST_TRUE(double_eq(nnMatrixAt(&B, 1, 1), 4, EPS));
309
310 nnMatrixDel(&A);
311 nnMatrixDel(&B);
312}
313
314TEST_CASE(nnMatrixTranspose_non_square_2x1) {
315 nnMatrix A = nnMatrixMake(2, 1);
316 nnMatrix B = nnMatrixMake(1, 2);
317
318 nnMatrixInit(&A, (R[]) {
319 1,
320 3,
321 });
322
323 nnMatrixTranspose(&A, &B);
324 TEST_TRUE(double_eq(nnMatrixAt(&B, 0, 0), 1, EPS));
325 TEST_TRUE(double_eq(nnMatrixAt(&B, 0, 1), 3, EPS));
326
327 nnMatrixDel(&A);
328 nnMatrixDel(&B);
329}
330
331TEST_CASE(nnMatrixGt_test) {
332 nnMatrix A = nnMatrixMake(2, 3);
333 nnMatrix B = nnMatrixMake(2, 3);
334
335 nnMatrixInit(&A, (R[]) {
336 -3, 2, 0,
337 4, -1, 5
338 });
339
340 nnMatrixGt(&A, 0, &B);
341 TEST_TRUE(double_eq(nnMatrixAt(&B, 0, 0), 0, EPS));
342 TEST_TRUE(double_eq(nnMatrixAt(&B, 0, 1), 1, EPS));
343 TEST_TRUE(double_eq(nnMatrixAt(&B, 0, 2), 0, EPS));
344 TEST_TRUE(double_eq(nnMatrixAt(&B, 1, 0), 1, EPS));
345 TEST_TRUE(double_eq(nnMatrixAt(&B, 1, 1), 0, EPS));
346 TEST_TRUE(double_eq(nnMatrixAt(&B, 1, 2), 1, EPS));
347
348 nnMatrixDel(&A);
349 nnMatrixDel(&B);
350}
diff --git a/src/lib/test/neuralnet_test.c b/src/lib/test/neuralnet_test.c
new file mode 100644
index 0000000..14d9438
--- /dev/null
+++ b/src/lib/test/neuralnet_test.c
@@ -0,0 +1,92 @@
1#include <neuralnet/neuralnet.h>
2
3#include <neuralnet/matrix.h>
4#include "activation.h"
5#include "neuralnet_impl.h"
6
7#include "test.h"
8#include "test_util.h"
9
10#include <assert.h>
11
12TEST_CASE(neuralnet_perceptron_test) {
13 const int num_layers = 1;
14 const int layer_sizes[] = { 1, 1 };
15 const nnActivation layer_activations[] = { nnSigmoid };
16 const R weights[] = { 0.3 };
17
18 nnNeuralNetwork* net = nnMakeNet(num_layers, layer_sizes, layer_activations);
19 assert(net);
20 nnSetWeights(net, weights);
21
22 nnQueryObject* query = nnMakeQueryObject(net, /*num_inputs=*/1);
23
24 const R input[] = { 0.9 };
25 R output[1];
26 nnQueryArray(net, query, input, output);
27
28 const R expected_output = sigmoid(input[0] * weights[0]);
29 printf("\nOutput: %f, Expected: %f\n", output[0], expected_output);
30 TEST_TRUE(double_eq(output[0], expected_output, EPS));
31
32 nnDeleteQueryObject(&query);
33 nnDeleteNet(&net);
34}
35
36TEST_CASE(neuralnet_xor_test) {
37 const int num_layers = 2;
38 const int layer_sizes[] = { 2, 2, 1 };
39 const nnActivation layer_activations[] = { nnRelu, nnIdentity };
40 const R weights[] = {
41 1, 1, 1, 1, // First (hidden) layer.
42 1, -2 // Second (output) layer.
43 };
44 const R biases[] = {
45 0, -1, // First (hidden) layer.
46 0 // Second (output) layer.
47 };
48
49 nnNeuralNetwork* net = nnMakeNet(num_layers, layer_sizes, layer_activations);
50 assert(net);
51 nnSetWeights(net, weights);
52 nnSetBiases(net, biases);
53
54 // First layer weights.
55 TEST_EQUAL(nnMatrixAt(&net->weights[0], 0, 0), 1);
56 TEST_EQUAL(nnMatrixAt(&net->weights[0], 0, 1), 1);
57 TEST_EQUAL(nnMatrixAt(&net->weights[0], 0, 2), 1);
58 TEST_EQUAL(nnMatrixAt(&net->weights[0], 0, 3), 1);
59 // Second layer weights.
60 TEST_EQUAL(nnMatrixAt(&net->weights[1], 0, 0), 1);
61 TEST_EQUAL(nnMatrixAt(&net->weights[1], 0, 1), -2);
62 // First layer biases.
63 TEST_EQUAL(nnMatrixAt(&net->biases[0], 0, 0), 0);
64 TEST_EQUAL(nnMatrixAt(&net->biases[0], 0, 1), -1);
65 // Second layer biases.
66 TEST_EQUAL(nnMatrixAt(&net->biases[1], 0, 0), 0);
67
68 // Test.
69
70 #define M 4
71
72 nnQueryObject* query = nnMakeQueryObject(net, /*num_inputs=*/M);
73
74 const R test_inputs[M][2] = { { 0., 0. }, { 1., 0. }, { 0., 1. }, { 1., 1. } };
75 nnMatrix test_inputs_matrix = nnMatrixMake(M, 2);
76 nnMatrixInit(&test_inputs_matrix, (const R*)test_inputs);
77 nnQuery(net, query, &test_inputs_matrix);
78
79 const R expected_outputs[M] = { 0., 1., 1., 0. };
80 for (int i = 0; i < M; ++i) {
81 const R test_output = nnMatrixAt(nnNetOutputs(query), i, 0);
82 printf("\nInput: (%f, %f), Output: %f, Expected: %f\n",
83 test_inputs[i][0], test_inputs[i][1], test_output, expected_outputs[i]);
84 }
85 for (int i = 0; i < M; ++i) {
86 const R test_output = nnMatrixAt(nnNetOutputs(query), i, 0);
87 TEST_TRUE(double_eq(test_output, expected_outputs[i], OUTPUT_EPS));
88 }
89
90 nnDeleteQueryObject(&query);
91 nnDeleteNet(&net);
92}
diff --git a/src/lib/test/test.h b/src/lib/test/test.h
new file mode 100644
index 0000000..fd8dc22
--- /dev/null
+++ b/src/lib/test/test.h
@@ -0,0 +1,185 @@
1// SPDX-License-Identifier: MIT
2#pragma once
3
4#ifdef UNIT_TEST
5
6#include <stdbool.h>
7#include <stdio.h>
8#include <stdlib.h>
9#include <string.h>
10
11#if defined(__DragonFly__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || \
12 defined(__NetBSD__) || defined(__OpenBSD__)
13#define USE_SYSCTL_FOR_ARGS 1
14// clang-format off
15#include <sys/types.h>
16#include <sys/sysctl.h>
17// clang-format on
18#include <unistd.h> // getpid
19#endif
20
21struct test_file_metadata;
22
23struct test_failure {
24 bool present;
25 const char *message;
26 const char *file;
27 int line;
28};
29
30struct test_case_metadata {
31 void (*fn)(struct test_case_metadata *, struct test_file_metadata *);
32 struct test_failure failure;
33 const char *name;
34 struct test_case_metadata *next;
35};
36
37struct test_file_metadata {
38 bool registered;
39 const char *name;
40 struct test_file_metadata *next;
41 struct test_case_metadata *tests;
42};
43
44struct test_file_metadata __attribute__((weak)) * test_file_head;
45
46#define SET_FAILURE(_message) \
47 metadata->failure = (struct test_failure) { \
48 .message = _message, .file = __FILE__, .line = __LINE__, .present = true, \
49 }
50
51#define TEST_EQUAL(a, b) \
52 do { \
53 if ((a) != (b)) { \
54 SET_FAILURE(#a " != " #b); \
55 return; \
56 } \
57 } while (0)
58
59#define TEST_TRUE(a) \
60 do { \
61 if (!(a)) { \
62 SET_FAILURE(#a " is not true"); \
63 return; \
64 } \
65 } while (0)
66
67#define TEST_STREQUAL(a, b) \
68 do { \
69 if (strcmp(a, b) != 0) { \
70 SET_FAILURE(#a " != " #b); \
71 return; \
72 } \
73 } while (0)
74
75#define TEST_CASE(_name) \
76 static void __test_h_##_name(struct test_case_metadata *, \
77 struct test_file_metadata *); \
78 static struct test_file_metadata __test_h_file; \
79 static struct test_case_metadata __test_h_meta_##_name = { \
80 .name = #_name, \
81 .fn = __test_h_##_name, \
82 }; \
83 static void __attribute__((constructor(101))) __test_h_##_name##_register(void) { \
84 __test_h_meta_##_name.next = __test_h_file.tests; \
85 __test_h_file.tests = &__test_h_meta_##_name; \
86 if (!__test_h_file.registered) { \
87 __test_h_file.name = __FILE__; \
88 __test_h_file.next = test_file_head; \
89 test_file_head = &__test_h_file; \
90 __test_h_file.registered = true; \
91 } \
92 } \
93 static void __test_h_##_name( \
94 struct test_case_metadata *metadata __attribute__((unused)), \
95 struct test_file_metadata *file_metadata __attribute__((unused)))
96
97extern void __attribute__((weak)) (*test_h_unittest_setup)(void);
98/// Run defined tests, return true if all tests succeeds
99/// @param[out] tests_run if not NULL, set to whether tests were run
100static inline void __attribute__((constructor(102))) run_tests(void) {
101 bool should_run = false;
102#ifdef USE_SYSCTL_FOR_ARGS
103 int mib[] = {
104 CTL_KERN,
105#if defined(__NetBSD__) || defined(__OpenBSD__)
106 KERN_PROC_ARGS,
107 getpid(),
108 KERN_PROC_ARGV,
109#else
110 KERN_PROC,
111 KERN_PROC_ARGS,
112 getpid(),
113#endif
114 };
115 char *arg = NULL;
116 size_t arglen;
117 sysctl(mib, sizeof(mib) / sizeof(mib[0]), NULL, &arglen, NULL, 0);
118 arg = malloc(arglen);
119 sysctl(mib, sizeof(mib) / sizeof(mib[0]), arg, &arglen, NULL, 0);
120#else
121 FILE *cmdlinef = fopen("/proc/self/cmdline", "r");
122 char *arg = NULL;
123 int arglen;
124 fscanf(cmdlinef, "%ms%n", &arg, &arglen);
125 fclose(cmdlinef);
126#endif
127 for (char *pos = arg; pos < arg + arglen; pos += strlen(pos) + 1) {
128 if (strcmp(pos, "--unittest") == 0) {
129 should_run = true;
130 break;
131 }
132 }
133 free(arg);
134
135 if (!should_run) {
136 return;
137 }
138
139 if (&test_h_unittest_setup) {
140 test_h_unittest_setup();
141 }
142
143 struct test_file_metadata *i = test_file_head;
144 int failed = 0, success = 0;
145 while (i) {
146 fprintf(stderr, "Running tests from %s:\n", i->name);
147 struct test_case_metadata *j = i->tests;
148 while (j) {
149 fprintf(stderr, "\t%s ... ", j->name);
150 j->failure.present = false;
151 j->fn(j, i);
152 if (j->failure.present) {
153 fprintf(stderr, "failed (%s at %s:%d)\n", j->failure.message,
154 j->failure.file, j->failure.line);
155 failed++;
156 } else {
157 fprintf(stderr, "passed\n");
158 success++;
159 }
160 j = j->next;
161 }
162 fprintf(stderr, "\n");
163 i = i->next;
164 }
165 int total = failed + success;
166 fprintf(stderr, "Test results: passed %d/%d, failed %d/%d\n", success, total,
167 failed, total);
168 exit(failed == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
169}
170
171#else
172
173#include <stdbool.h>
174
175#define TEST_CASE(name) static void __attribute__((unused)) __test_h_##name(void)
176
177#define TEST_EQUAL(a, b) \
178 (void)(a); \
179 (void)(b)
180#define TEST_TRUE(a) (void)(a)
181#define TEST_STREQUAL(a, b) \
182 (void)(a); \
183 (void)(b)
184
185#endif
diff --git a/src/lib/test/test_main.c b/src/lib/test/test_main.c
new file mode 100644
index 0000000..4cce7f6
--- /dev/null
+++ b/src/lib/test/test_main.c
@@ -0,0 +1,3 @@
1int main() {
2 return 0;
3}
diff --git a/src/lib/test/test_util.h b/src/lib/test/test_util.h
new file mode 100644
index 0000000..8abb99a
--- /dev/null
+++ b/src/lib/test/test_util.h
@@ -0,0 +1,22 @@
1#pragma once
2
3#include <neuralnet/types.h>
4
5#include <math.h>
6
7// General epsilon for comparing values.
8static const R EPS = 1e-10;
9
10// Epsilon for comparing network weights after training.
11static const R WEIGHT_EPS = 0.01;
12
13// Epsilon for comparing network outputs after training.
14static const R OUTPUT_EPS = 0.01;
15
16static inline bool double_eq(double a, double b, double eps) {
17 return fabs(a - b) <= eps;
18}
19
20static inline R lerp(R a, R b, R t) {
21 return a + t*(b-a);
22}
diff --git a/src/lib/test/train_linear_perceptron_non_origin_test.c b/src/lib/test/train_linear_perceptron_non_origin_test.c
new file mode 100644
index 0000000..5a320ac
--- /dev/null
+++ b/src/lib/test/train_linear_perceptron_non_origin_test.c
@@ -0,0 +1,67 @@
1#include <neuralnet/train.h>
2
3#include <neuralnet/matrix.h>
4#include <neuralnet/neuralnet.h>
5#include "activation.h"
6#include "neuralnet_impl.h"
7
8#include "test.h"
9#include "test_util.h"
10
11#include <assert.h>
12
13TEST_CASE(neuralnet_train_linear_perceptron_non_origin_test) {
14 const int num_layers = 1;
15 const int layer_sizes[] = { 1, 1 };
16 const nnActivation layer_activations[] = { nnIdentity };
17
18 nnNeuralNetwork* net = nnMakeNet(num_layers, layer_sizes, layer_activations);
19 assert(net);
20
21 // Train.
22
23 // Try to learn the Y = 2X + 1 line.
24 #define N 2
25 const R inputs[N] = { 0., 1. };
26 const R targets[N] = { 1., 3. };
27
28 nnMatrix inputs_matrix = nnMatrixMake(N, 1);
29 nnMatrix targets_matrix = nnMatrixMake(N, 1);
30 nnMatrixInit(&inputs_matrix, inputs);
31 nnMatrixInit(&targets_matrix, targets);
32
33 nnTrainingParams params = {
34 .learning_rate = 0.7,
35 .max_iterations = 20,
36 .seed = 0,
37 .weight_init = nnWeightInit01,
38 .debug = false,
39 };
40
41 nnTrain(net, &inputs_matrix, &targets_matrix, &params);
42
43 const R weight = nnMatrixAt(&net->weights[0], 0, 0);
44 const R expected_weight = 2.0;
45 printf("\nTrained network weight: %f, Expected: %f\n", weight, expected_weight);
46 TEST_TRUE(double_eq(weight, expected_weight, WEIGHT_EPS));
47
48 const R bias = nnMatrixAt(&net->biases[0], 0, 0);
49 const R expected_bias = 1.0;
50 printf("Trained network bias: %f, Expected: %f\n", bias, expected_bias);
51 TEST_TRUE(double_eq(bias, expected_bias, WEIGHT_EPS));
52
53 // Test.
54
55 nnQueryObject* query = nnMakeQueryObject(net, /*num_inputs=*/1);
56
57 const R test_input[] = { 2.3 };
58 R test_output[1];
59 nnQueryArray(net, query, test_input, test_output);
60
61 const R expected_output = test_input[0] * expected_weight + expected_bias;
62 printf("Output: %f, Expected: %f\n", test_output[0], expected_output);
63 TEST_TRUE(double_eq(test_output[0], expected_output, OUTPUT_EPS));
64
65 nnDeleteQueryObject(&query);
66 nnDeleteNet(&net);
67}
diff --git a/src/lib/test/train_linear_perceptron_test.c b/src/lib/test/train_linear_perceptron_test.c
new file mode 100644
index 0000000..2b1336d
--- /dev/null
+++ b/src/lib/test/train_linear_perceptron_test.c
@@ -0,0 +1,62 @@
1#include <neuralnet/train.h>
2
3#include <neuralnet/matrix.h>
4#include <neuralnet/neuralnet.h>
5#include "activation.h"
6#include "neuralnet_impl.h"
7
8#include "test.h"
9#include "test_util.h"
10
11#include <assert.h>
12
13TEST_CASE(neuralnet_train_linear_perceptron_test) {
14 const int num_layers = 1;
15 const int layer_sizes[] = { 1, 1 };
16 const nnActivation layer_activations[] = { nnIdentity };
17
18 nnNeuralNetwork* net = nnMakeNet(num_layers, layer_sizes, layer_activations);
19 assert(net);
20
21 // Train.
22
23 // Try to learn the Y=X line.
24 #define N 2
25 const R inputs[N] = { 0., 1. };
26 const R targets[N] = { 0., 1. };
27
28 nnMatrix inputs_matrix = nnMatrixMake(N, 1);
29 nnMatrix targets_matrix = nnMatrixMake(N, 1);
30 nnMatrixInit(&inputs_matrix, inputs);
31 nnMatrixInit(&targets_matrix, targets);
32
33 nnTrainingParams params = {
34 .learning_rate = 0.7,
35 .max_iterations = 10,
36 .seed = 0,
37 .weight_init = nnWeightInit01,
38 .debug = false,
39 };
40
41 nnTrain(net, &inputs_matrix, &targets_matrix, &params);
42
43 const R weight = nnMatrixAt(&net->weights[0], 0, 0);
44 const R expected_weight = 1.0;
45 printf("\nTrained network weight: %f, Expected: %f\n", weight, expected_weight);
46 TEST_TRUE(double_eq(weight, expected_weight, WEIGHT_EPS));
47
48 // Test.
49
50 nnQueryObject* query = nnMakeQueryObject(net, /*num_inputs=*/1);
51
52 const R test_input[] = { 2.3 };
53 R test_output[1];
54 nnQueryArray(net, query, test_input, test_output);
55
56 const R expected_output = test_input[0];
57 printf("Output: %f, Expected: %f\n", test_output[0], expected_output);
58 TEST_TRUE(double_eq(test_output[0], expected_output, OUTPUT_EPS));
59
60 nnDeleteQueryObject(&query);
61 nnDeleteNet(&net);
62}
diff --git a/src/lib/test/train_sigmoid_test.c b/src/lib/test/train_sigmoid_test.c
new file mode 100644
index 0000000..588e7ca
--- /dev/null
+++ b/src/lib/test/train_sigmoid_test.c
@@ -0,0 +1,66 @@
1#include <neuralnet/train.h>
2
3#include <neuralnet/matrix.h>
4#include <neuralnet/neuralnet.h>
5#include "activation.h"
6#include "neuralnet_impl.h"
7
8#include "test.h"
9#include "test_util.h"
10
11#include <assert.h>
12
13TEST_CASE(neuralnet_train_sigmoid_test) {
14 const int num_layers = 1;
15 const int layer_sizes[] = { 1, 1 };
16 const nnActivation layer_activations[] = { nnSigmoid };
17
18 nnNeuralNetwork* net = nnMakeNet(num_layers, layer_sizes, layer_activations);
19 assert(net);
20
21 // Train.
22
23 // Try to learn the sigmoid function.
24 #define N 3
25 R inputs[N];
26 R targets[N];
27 for (int i = 0; i < N; ++i) {
28 inputs[i] = lerp(-1, +1, (R)i / (R)(N-1));
29 targets[i] = sigmoid(inputs[i]);
30 }
31
32 nnMatrix inputs_matrix = nnMatrixMake(N, 1);
33 nnMatrix targets_matrix = nnMatrixMake(N, 1);
34 nnMatrixInit(&inputs_matrix, inputs);
35 nnMatrixInit(&targets_matrix, targets);
36
37 nnTrainingParams params = {
38 .learning_rate = 0.9,
39 .max_iterations = 100,
40 .seed = 0,
41 .weight_init = nnWeightInit01,
42 .debug = false,
43 };
44
45 nnTrain(net, &inputs_matrix, &targets_matrix, &params);
46
47 const R weight = nnMatrixAt(&net->weights[0], 0, 0);
48 const R expected_weight = 1.0;
49 printf("\nTrained network weight: %f, Expected: %f\n", weight, expected_weight);
50 TEST_TRUE(double_eq(weight, expected_weight, WEIGHT_EPS));
51
52 // Test.
53
54 nnQueryObject* query = nnMakeQueryObject(net, /*num_inputs=*/1);
55
56 const R test_input[] = { 0.3 };
57 R test_output[1];
58 nnQueryArray(net, query, test_input, test_output);
59
60 const R expected_output = 0.574442516811659; // sigmoid(0.3)
61 printf("Output: %f, Expected: %f\n", test_output[0], expected_output);
62 TEST_TRUE(double_eq(test_output[0], expected_output, OUTPUT_EPS));
63
64 nnDeleteQueryObject(&query);
65 nnDeleteNet(&net);
66}
diff --git a/src/lib/test/train_xor_test.c b/src/lib/test/train_xor_test.c
new file mode 100644
index 0000000..6ddc6e0
--- /dev/null
+++ b/src/lib/test/train_xor_test.c
@@ -0,0 +1,66 @@
1#include <neuralnet/train.h>
2
3#include <neuralnet/matrix.h>
4#include <neuralnet/neuralnet.h>
5#include "activation.h"
6#include "neuralnet_impl.h"
7
8#include "test.h"
9#include "test_util.h"
10
11#include <assert.h>
12
13TEST_CASE(neuralnet_train_xor_test) {
14 const int num_layers = 2;
15 const int layer_sizes[] = { 2, 2, 1 };
16 const nnActivation layer_activations[] = { nnRelu, nnIdentity };
17
18 nnNeuralNetwork* net = nnMakeNet(num_layers, layer_sizes, layer_activations);
19 assert(net);
20
21 // Train.
22
23 #define N 4
24 const R inputs[N][2] = { { 0., 0. }, { 0., 1. }, { 1., 0. }, { 1., 1. } };
25 const R targets[N] = { 0., 1., 1., 0. };
26
27 nnMatrix inputs_matrix = nnMatrixMake(N, 2);
28 nnMatrix targets_matrix = nnMatrixMake(N, 1);
29 nnMatrixInit(&inputs_matrix, (const R*)inputs);
30 nnMatrixInit(&targets_matrix, targets);
31
32 nnTrainingParams params = {
33 .learning_rate = 0.1,
34 .max_iterations = 500,
35 .seed = 0,
36 .weight_init = nnWeightInit01,
37 .debug = false,
38 };
39
40 nnTrain(net, &inputs_matrix, &targets_matrix, &params);
41
42 // Test.
43
44 #define M 4
45
46 nnQueryObject* query = nnMakeQueryObject(net, /*num_inputs=*/M);
47
48 const R test_inputs[M][2] = { { 0., 0. }, { 1., 0. }, { 0., 1. }, { 1., 1. } };
49 nnMatrix test_inputs_matrix = nnMatrixMake(M, 2);
50 nnMatrixInit(&test_inputs_matrix, (const R*)test_inputs);
51 nnQuery(net, query, &test_inputs_matrix);
52
53 const R expected_outputs[M] = { 0., 1., 1., 0. };
54 for (int i = 0; i < M; ++i) {
55 const R test_output = nnMatrixAt(nnNetOutputs(query), i, 0);
56 printf("\nInput: (%f, %f), Output: %f, Expected: %f\n",
57 test_inputs[i][0], test_inputs[i][1], test_output, expected_outputs[i]);
58 }
59 for (int i = 0; i < M; ++i) {
60 const R test_output = nnMatrixAt(nnNetOutputs(query), i, 0);
61 TEST_TRUE(double_eq(test_output, expected_outputs[i], OUTPUT_EPS));
62 }
63
64 nnDeleteQueryObject(&query);
65 nnDeleteNet(&net);
66}
generated by cgit v1.2.3 (git 2.25.1) at 2024-12-26 07:26:10 +0000