aboutsummaryrefslogtreecommitdiff
path: root/src/lib/test
diff options
context:
space:
mode:
Diffstat (limited to 'src/lib/test')
-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
9 files changed, 913 insertions, 0 deletions
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 2024-11-22 20:35:52 +0000