1#include "ggml-opt.h"
2
3#include "ggml.h"
4#include "ggml-alloc.h"
5#include "ggml-backend.h"
6#include "ggml-impl.h"
7
8#include <algorithm>
9#include <cmath>
10#include <cstdint>
11#include <cinttypes>
12#include <map>
13#include <random>
14#include <vector>
15
16struct ggml_opt_dataset {
17 struct ggml_context * ctx = nullptr;
18 ggml_backend_buffer_t buf = nullptr;
19 struct ggml_tensor * data = nullptr;
20 struct ggml_tensor * labels = nullptr;
21
22 int64_t ndata = -1;
23 int64_t ndata_shard = -1;
24 size_t nbs_data = -1;
25 size_t nbs_labels = -1;
26
27 std::vector<int64_t> permutation;
28};
29
30struct ggml_opt_context {
31 ggml_backend_sched_t backend_sched = nullptr;
32 ggml_cgraph * allocated_graph = nullptr;
33 ggml_cgraph * allocated_graph_copy = nullptr;
34 struct ggml_context * ctx_static = nullptr;
35 struct ggml_context * ctx_cpu = nullptr;
36 struct ggml_context * ctx_compute = nullptr;
37 struct ggml_context * ctx_copy = nullptr;
38 ggml_backend_buffer_t buf_static = nullptr;
39 ggml_backend_buffer_t buf_cpu = nullptr;
40 std::mt19937 rng;
41 enum ggml_opt_loss_type loss_type;
42 enum ggml_opt_build_type build_type;
43 enum ggml_opt_build_type build_type_alloc;
44
45 struct ggml_tensor * inputs = nullptr;
46 struct ggml_tensor * outputs = nullptr;
47 struct ggml_tensor * labels = nullptr;
48
49 struct ggml_tensor * loss = nullptr;
50 struct ggml_tensor * pred = nullptr;
51 struct ggml_tensor * ncorrect = nullptr;
52
53 struct ggml_cgraph * gf = nullptr;
54 struct ggml_cgraph * gb_grad = nullptr;
55 struct ggml_cgraph * gb_opt = nullptr;
56 bool static_graphs = false;
57 bool eval_ready = false;
58 std::vector<struct ggml_tensor *> grad_accs;
59 std::vector<struct ggml_tensor *> grad_m;
60 std::vector<struct ggml_tensor *> grad_v;
61
62 int64_t iter = 1;
63 int32_t opt_period = 1;
64 int32_t opt_i = 0;
65 bool loss_per_datapoint = false;
66
67 ggml_opt_get_optimizer_params get_opt_pars = nullptr;
68 void * get_opt_pars_ud = nullptr;
69 struct ggml_tensor * opt_step_params = nullptr; // Stores output of get_opt_pars.
70
71 enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
72};
73
74struct ggml_opt_result {
75 int64_t ndata = 0;
76 std::vector<float> loss;
77 std::vector<int32_t> pred;
78 int64_t ncorrect = 0;
79
80 int64_t opt_period = -1;
81 bool loss_per_datapoint = false;
82};
83
84// ====== Dataset ======
85
86ggml_opt_dataset_t ggml_opt_dataset_init(
87 enum ggml_type type_data,
88 enum ggml_type type_label,
89 int64_t ne_datapoint,
90 int64_t ne_label,
91 int64_t ndata,
92 int64_t ndata_shard) {
93 GGML_ASSERT(ne_datapoint > 0);
94 GGML_ASSERT(ne_label >= 0);
95 GGML_ASSERT(ndata > 0);
96 GGML_ASSERT(ndata_shard > 0);
97
98 ggml_opt_dataset_t result = new ggml_opt_dataset;
99 result->ndata = ndata;
100 result->ndata_shard = ndata_shard;
101
102 {
103 struct ggml_init_params params = {
104 /*.mem_size =*/ 2*ggml_tensor_overhead(),
105 /*.mem_buffer =*/ nullptr,
106 /*.no_alloc =*/ true,
107 };
108 result->ctx = ggml_init(params);
109 }
110
111 result->data = ggml_new_tensor_2d(result->ctx, type_data, ne_datapoint, ndata);
112 result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
113
114 if (ne_label > 0) {
115 result->labels = ggml_new_tensor_2d(result->ctx, type_label, ne_label, ndata);
116 result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
117 } else {
118 result->labels = nullptr;
119 result->nbs_labels = 0;
120 }
121
122 result->buf = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx, ggml_backend_cpu_buffer_type());
123
124 const int64_t nshards = ndata/ndata_shard;
125 result->permutation.resize(nshards);
126 for (int64_t i = 0; i < nshards; ++i) {
127 result->permutation[i] = i;
128 }
129 return result;
130}
131
132void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
133 ggml_backend_buffer_free(dataset->buf);
134 ggml_free(dataset->ctx);
135 delete dataset;
136}
137
138int64_t ggml_opt_dataset_ndata(ggml_opt_dataset_t dataset) {
139 return dataset->ndata;
140}
141
142struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
143 return dataset->data;
144}
145
146struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset) {
147 return dataset->labels;
148}
149
150void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata) {
151 GGML_ASSERT(idata <= dataset->ndata);
152
153 if (idata < 0) {
154 std::shuffle(dataset->permutation.begin(), dataset->permutation.end(), opt_ctx->rng);
155 return;
156 }
157
158 GGML_ASSERT(idata % dataset->ndata_shard == 0);
159 const int64_t ishard_max = idata / dataset->ndata_shard;
160 std::shuffle(dataset->permutation.begin(), dataset->permutation.begin() + ishard_max, opt_ctx->rng);
161}
162
163void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor * data_batch, struct ggml_tensor * labels_batch, int64_t ibatch) {
164 GGML_ASSERT( data_batch && ggml_is_contiguous(data_batch));
165 GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
166 GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
167 GGML_ASSERT( data_batch->type == dataset->data->type);
168 GGML_ASSERT(!labels_batch || labels_batch->type == dataset->labels->type);
169
170 const size_t nb_data_batch = ggml_nbytes(data_batch);
171 GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
172 const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
173
174 if (labels_batch) {
175 const size_t nb_labels_batch = ggml_nbytes(labels_batch);
176 GGML_ASSERT(nb_labels_batch == shards_per_batch*dataset->nbs_labels);
177 }
178
179 GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
180
181 for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
182 const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
183
184 const char * ptr_data = (const char *) dataset->data->data + ishard*dataset->nbs_data;
185 ggml_backend_tensor_set(data_batch, ptr_data, ishard_batch*dataset->nbs_data, dataset->nbs_data);
186
187 if (!labels_batch) {
188 continue;
189 }
190
191 const char * ptr_labels = (const char *) dataset->labels->data + ishard*dataset->nbs_labels;
192 ggml_backend_tensor_set(labels_batch, ptr_labels, ishard_batch*dataset->nbs_labels, dataset->nbs_labels);
193 }
194}
195
196void ggml_opt_dataset_get_batch_host(ggml_opt_dataset_t dataset, void * data_batch, size_t nb_data_batch, void * labels_batch, int64_t ibatch) {
197 GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
198 GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
199
200 const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
201
202 GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
203
204 for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
205 const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
206
207 const char * ptr_data = (const char *) dataset->data->data + ishard *dataset->nbs_data;
208 char * ptr_data_batch = (char *) data_batch + ishard_batch*dataset->nbs_data;
209 memcpy(ptr_data_batch, ptr_data, dataset->nbs_data);
210
211 if (!labels_batch) {
212 continue;
213 }
214
215 const char * ptr_labels = (const char *) dataset->labels->data + ishard *dataset->nbs_labels;
216 char * ptr_labels_batch = (char *) labels_batch + ishard_batch*dataset->nbs_labels;
217 memcpy(ptr_labels_batch, ptr_labels, dataset->nbs_labels);
218 }
219}
220
221// ====== Model / Context ======
222
223struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
224 GGML_UNUSED(userdata);
225
226 ggml_opt_optimizer_params result;
227
228 result.adamw.alpha = 0.001f;
229 result.adamw.beta1 = 0.9f;
230 result.adamw.beta2 = 0.999f;
231 result.adamw.eps = 1e-8f;
232 result.adamw.wd = 0.0f;
233
234 result.sgd.alpha = 1e-3f;
235 result.sgd.wd = 0.0f;
236
237 return result;
238}
239
240
241struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) {
242 return *((struct ggml_opt_optimizer_params *) userdata);
243}
244
245struct ggml_opt_params ggml_opt_default_params(
246 ggml_backend_sched_t backend_sched,
247 enum ggml_opt_loss_type loss_type) {
248 return {
249 /*backend_sched =*/ backend_sched,
250 /*ctx_compute =*/ nullptr,
251 /*inputs =*/ nullptr,
252 /*logits =*/ nullptr,
253 /*loss_type =*/ loss_type,
254 /*build_type =*/ GGML_OPT_BUILD_TYPE_OPT,
255 /*opt_period =*/ 1,
256 /*get_opt_pars =*/ ggml_opt_get_default_optimizer_params,
257 /*get_opt_pars_ud =*/ nullptr,
258 /*optimizer =*/ GGML_OPT_OPTIMIZER_TYPE_ADAMW,
259 };
260}
261
262static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_map, ggml_context * ctx, ggml_tensor * tensor) {
263 if (!tensor) {
264 return nullptr;
265 }
266
267 if (tensor_map.find(tensor) != tensor_map.end()) {
268 return tensor_map[tensor];
269 }
270
271 ggml_tensor * new_tensor = ggml_dup_tensor(ctx, tensor);
272 tensor_map[tensor] = new_tensor;
273
274 new_tensor->op = tensor->op;
275 for (int i = 0; i < GGML_MAX_DIMS; i++) {
276 new_tensor->nb[i] = tensor->nb[i];
277 }
278 new_tensor->flags = tensor->flags;
279 memcpy(new_tensor->op_params, tensor->op_params, sizeof(tensor->op_params));
280 strcpy(new_tensor->name, tensor->name);
281 new_tensor->data = tensor->data;
282 new_tensor->buffer = tensor->buffer;
283 new_tensor->extra = tensor->extra;
284 new_tensor->view_offs = tensor->view_offs;
285 new_tensor->view_src = map_tensor(tensor_map, ctx, tensor->view_src);
286 for (int i = 0; i < GGML_MAX_SRC; i++) {
287 new_tensor->src[i] = map_tensor(tensor_map, ctx, tensor->src[i]);
288 }
289
290 return new_tensor;
291}
292
293static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
294 std::map<ggml_tensor *, ggml_tensor *> tensor_map;
295
296 ggml_cgraph * dst = ggml_new_graph_custom(ctx, src->size, /*grads =*/ true);
297
298 for (int i = 0; i < src->n_leafs; i++) {
299 ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->leafs[i]));
300 }
301 GGML_ASSERT(dst->n_leafs == src->n_leafs);
302 for (int i = 0; i < src->n_nodes; i++) {
303 ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->nodes[i]));
304 }
305 GGML_ASSERT(dst->n_nodes == src->n_nodes);
306 for (int i = 0; i < src->n_nodes; ++i) {
307 const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
308 const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
309
310 GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
311 GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
312 GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
313 GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
314
315 dst->grads[igrad_dst] = src->grads[igrad_src];
316 dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
317 }
318
319 return dst;
320}
321
322static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
323 GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc");
324 GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically");
325
326 const enum ggml_opt_optimizer_type optimizer = opt_ctx->optimizer;
327
328 const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD &&
329 !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1);
330
331 const bool need_momenta = opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT &&
332 opt_ctx->optimizer == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
333
334 ggml_set_input(opt_ctx->inputs);
335 ggml_set_output(opt_ctx->outputs);
336
337 int n_param = 0;
338 for (int i = 0; i < opt_ctx->gf->n_nodes; ++i) {
339 const struct ggml_tensor * node = opt_ctx->gf->nodes[i];
340 if (node->flags & GGML_TENSOR_FLAG_PARAM) {
341 n_param++;
342 }
343 GGML_ASSERT(!(node->flags & GGML_TENSOR_FLAG_LOSS) && "support for extra loss terms not implemented");
344 }
345
346 if (!opt_ctx->ctx_static) {
347 // The static context is used for:
348 // - gradients (1 per loss, 1 tensor per param if using gradient accumulation)
349 // - optimizer momenta (2 tensors per param)
350 // - labels (if using static graphs)
351 // - loss (if using static graphs, up to 5 tensors)
352 // - pred (if using static graphs)
353 // - ncorrect (if using static graphs, 2 tensors).
354 constexpr size_t n_loss = 1;
355 const size_t tensors_per_param = (accumulate ? 1 : 0) + (need_momenta ? 2 : 0);
356 const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0;
357 const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead();
358 struct ggml_init_params params = {
359 /*.mem_size =*/ size_meta,
360 /*.mem_buffer =*/ nullptr,
361 /*.no_alloc =*/ true,
362 };
363 opt_ctx->ctx_static = ggml_init(params);
364 }
365 GGML_ASSERT(opt_ctx->build_type <= opt_ctx->build_type_alloc);
366
367 {
368 // The cpu context is allocated statically if using static graphs, dynamically otherwise.
369 // It is used for:
370 // - optimizer parameters (1 shared for all optimizer invocations)
371 const size_t size_meta = 1 * ggml_tensor_overhead();
372 struct ggml_init_params params = {
373 /*.mem_size =*/ size_meta,
374 /*.mem_buffer =*/ nullptr,
375 /*.no_alloc =*/ true,
376 };
377 ggml_free(opt_ctx->ctx_cpu);
378 opt_ctx->ctx_cpu = ggml_init(params);
379
380 ggml_backend_buffer_free(opt_ctx->buf_cpu);
381 opt_ctx->buf_cpu = nullptr;
382 }
383
384 struct ggml_context * ctx_results = opt_ctx->static_graphs ? opt_ctx->ctx_static : opt_ctx->ctx_compute;
385
386 switch (opt_ctx->loss_type) {
387 case GGML_OPT_LOSS_TYPE_MEAN: {
388 opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
389 ggml_set_name(opt_ctx->loss, "loss_sum");
390 const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
391 opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
392 ggml_set_name(opt_ctx->loss, "loss_mean");
393 opt_ctx->loss_per_datapoint = true;
394 break;
395 }
396 case GGML_OPT_LOSS_TYPE_SUM: {
397 opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
398 ggml_set_name(opt_ctx->loss, "loss_sum");
399 opt_ctx->loss_per_datapoint = false;
400 break;
401 }
402 case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
403 opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
404 ggml_set_input(opt_ctx->labels);
405 ggml_set_name(opt_ctx->labels, "labels");
406 opt_ctx->loss = ggml_cross_entropy_loss(ctx_results, opt_ctx->outputs, opt_ctx->labels);
407 ggml_set_name(opt_ctx->loss, "loss_cross_entropy");
408 if (opt_ctx->opt_period > 1) {
409 opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, 1.0f / opt_ctx->opt_period);
410 ggml_set_name(opt_ctx->loss, "loss_cross_entropy_scaled");
411 }
412 opt_ctx->loss_per_datapoint = true;
413 break;
414 }
415 case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
416 opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
417 ggml_set_input(opt_ctx->labels);
418 ggml_set_name(opt_ctx->labels, "labels");
419 opt_ctx->loss = ggml_sub(ctx_results, opt_ctx->outputs, opt_ctx->labels);
420 ggml_set_name(opt_ctx->loss, "loss_error");
421 opt_ctx->loss = ggml_sqr(ctx_results, opt_ctx->loss);
422 ggml_set_name(opt_ctx->loss, "loss_squared_error");
423 opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->loss);
424 ggml_set_name(opt_ctx->loss, "loss_sum_squared_error");
425 const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
426 opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
427 ggml_set_name(opt_ctx->loss, "loss_mean_squared_error");
428 opt_ctx->loss_per_datapoint = true;
429 break;
430 }
431 }
432 ggml_set_output(opt_ctx->loss);
433 ggml_set_loss(opt_ctx->loss);
434 ggml_build_forward_expand(opt_ctx->gf, opt_ctx->loss);
435
436 if (opt_ctx->loss_type == GGML_OPT_LOSS_TYPE_CROSS_ENTROPY) {
437 opt_ctx->pred = ggml_argmax(ctx_results, opt_ctx->outputs);
438 ggml_set_name(opt_ctx->pred, "pred");
439 ggml_set_output(opt_ctx->pred);
440 ggml_build_forward_expand(opt_ctx->gf, opt_ctx->pred);
441
442 opt_ctx->ncorrect = ggml_count_equal(ctx_results, opt_ctx->pred, ggml_argmax(ctx_results, opt_ctx->labels));
443 ggml_set_name(opt_ctx->ncorrect, "ncorrect");
444 ggml_set_output(opt_ctx->ncorrect);
445 ggml_build_forward_expand(opt_ctx->gf, opt_ctx->ncorrect);
446 }
447
448 if (opt_ctx->buf_static) {
449 if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
450 return;
451 }
452 } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_FORWARD) {
453 opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
454 opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
455 return;
456 }
457
458 if (opt_ctx->grad_accs.empty()) {
459 GGML_ASSERT(opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD);
460
461 const int n_nodes = opt_ctx->gf->n_nodes;
462 opt_ctx->grad_accs.resize(n_nodes);
463 for (int i = 0; i < n_nodes; ++i) {
464 ggml_tensor * node = opt_ctx->gf->nodes[i];
465 if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
466 opt_ctx->grad_accs[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
467 } else {
468 opt_ctx->grad_accs[i] = nullptr;
469 }
470 }
471
472 if (need_momenta && opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
473 opt_ctx->grad_m.resize(n_nodes);
474 opt_ctx->grad_v.resize(n_nodes);
475 for (int i = 0; i < n_nodes; ++i) {
476 ggml_tensor * node = opt_ctx->gf->nodes[i];
477 if (node->flags & GGML_TENSOR_FLAG_PARAM) {
478 opt_ctx->grad_m[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
479 opt_ctx->grad_v[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
480 } else {
481 opt_ctx->grad_m[i] = nullptr;
482 opt_ctx->grad_v[i] = nullptr;
483 }
484 }
485 }
486 }
487
488 // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
489 opt_ctx->gb_grad = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gf, /*force_grads =*/ true);
490 ggml_build_backward_expand(opt_ctx->ctx_compute, opt_ctx->gb_grad, opt_ctx->grad_accs.data());
491
492 if (opt_ctx->buf_static) {
493 if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_GRAD) {
494 return;
495 }
496 } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_GRAD) {
497 opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
498 ggml_graph_reset(opt_ctx->gb_grad);
499 }
500
501 GGML_ASSERT(opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT);
502
503 // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
504 opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true);
505
506 opt_ctx->opt_step_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, need_momenta ? 7 : 2);
507 ggml_tensor * adamw_params = opt_ctx->opt_step_params;
508 ggml_set_input(adamw_params);
509 const char * optimizer_name = ggml_opt_optimizer_name(opt_ctx->optimizer);
510 ggml_format_name(adamw_params, "%s_params", optimizer_name);
511 for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) {
512 struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i];
513 struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node);
514
515 if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
516 struct ggml_tensor * m = nullptr;
517 struct ggml_tensor * v = nullptr;
518 if (need_momenta) {
519 m = opt_ctx->grad_m[i];
520 v = opt_ctx->grad_v[i];
521 ggml_format_name(m, "AdamW m for %s", node->name);
522 ggml_format_name(v, "AdamW v for %s", node->name);
523 }
524 struct ggml_tensor * opt_step;
525 switch (optimizer) {
526 case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
527 opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, adamw_params);
528 break;
529 case GGML_OPT_OPTIMIZER_TYPE_SGD:
530 opt_step = ggml_opt_step_sgd(opt_ctx->ctx_compute, node, grad, adamw_params);
531 break;
532 default:
533 GGML_ABORT("fatal error");
534 }
535 ggml_format_name(opt_step, "%s step for %s", optimizer_name, node->name);
536 ggml_build_forward_expand(opt_ctx->gb_opt, opt_step);
537 }
538 }
539
540 if (!opt_ctx->buf_static) {
541 opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
542 opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
543 ggml_graph_reset(opt_ctx->gb_opt);
544 }
545
546 opt_ctx->buf_cpu = ggml_backend_alloc_ctx_tensors_from_buft(opt_ctx->ctx_cpu, ggml_backend_cpu_buffer_type());
547}
548
549ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
550 ggml_opt_context_t result = new struct ggml_opt_context;
551 result->backend_sched = params.backend_sched;
552 result->ctx_compute = params.ctx_compute;
553 result->loss_type = params.loss_type;
554 result->build_type = params.build_type;
555 result->build_type_alloc = params.build_type;
556 result->inputs = params.inputs;
557 result->outputs = params.outputs;
558 result->opt_period = params.opt_period;
559 result->get_opt_pars = params.get_opt_pars;
560 result->get_opt_pars_ud = params.get_opt_pars_ud;
561 result->optimizer = params.optimizer;
562
563 GGML_ASSERT(result->opt_period >= 1);
564
565 result->static_graphs = result->ctx_compute;
566
567 if (!result->static_graphs) {
568 GGML_ASSERT(!result->inputs);
569 GGML_ASSERT(!result->outputs);
570 return result;
571 }
572
573 GGML_ASSERT(result->inputs);
574 GGML_ASSERT(result->outputs);
575
576 result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
577 ggml_build_forward_expand(result->gf, result->outputs);
578
579 ggml_opt_build(result);
580
581 return result;
582}
583
584void ggml_opt_free(ggml_opt_context_t opt_ctx) {
585 if (opt_ctx == nullptr) {
586 return;
587 }
588 ggml_backend_buffer_free(opt_ctx->buf_static);
589 ggml_backend_buffer_free(opt_ctx->buf_cpu);
590 ggml_free(opt_ctx->ctx_static);
591 ggml_free(opt_ctx->ctx_cpu);
592 delete opt_ctx;
593}
594
595void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer) {
596 if (optimizer) {
597 ggml_graph_reset(opt_ctx->gb_opt);
598 opt_ctx->iter = 1;
599 } else {
600 ggml_graph_reset(opt_ctx->gb_grad);
601 }
602}
603
604bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx) {
605 return opt_ctx->static_graphs;
606}
607
608struct ggml_tensor * ggml_opt_inputs(ggml_opt_context_t opt_ctx) {
609 return opt_ctx->inputs;
610}
611
612struct ggml_tensor * ggml_opt_outputs(ggml_opt_context_t opt_ctx) {
613 return opt_ctx->outputs;
614}
615
616struct ggml_tensor * ggml_opt_labels(ggml_opt_context_t opt_ctx) {
617 return opt_ctx->labels;
618}
619
620struct ggml_tensor * ggml_opt_loss(ggml_opt_context_t opt_ctx) {
621 return opt_ctx->loss;
622}
623
624struct ggml_tensor * ggml_opt_pred(ggml_opt_context_t opt_ctx) {
625 return opt_ctx->pred;
626}
627
628struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx) {
629 return opt_ctx->ncorrect;
630}
631
632struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node) {
633 return ggml_graph_get_grad_acc(opt_ctx->gb_opt, node);
634}
635
636// ====== Optimization Result ======
637
638ggml_opt_result_t ggml_opt_result_init() {
639 return new ggml_opt_result;
640}
641
642void ggml_opt_result_free(ggml_opt_result_t result) {
643 delete result;
644}
645
646void ggml_opt_result_reset(ggml_opt_result_t result) {
647 result->ndata = 0;
648 result->loss.clear();
649 result->pred.clear();
650 result->ncorrect = 0;
651}
652
653void ggml_opt_result_ndata(ggml_opt_result_t result, int64_t * ndata) {
654 *ndata = result->ndata;
655}
656
657void ggml_opt_result_loss(ggml_opt_result_t result, double * loss, double * unc) {
658 const int64_t nbatches = result->loss.size(); // Number of physical batches.
659
660 if (nbatches == 0) {
661 *loss = 0.0;
662 *unc = NAN;
663 return;
664 }
665
666 double sum = 0.0;
667 double sum_squared = 0.0;
668
669 for (const float & loss : result->loss) {
670 // If the loss is per datapoint it was scaled by 1.0f/opt_period for each physical batch.
671 const float loss_scaled = result->loss_per_datapoint ? loss*result->opt_period : loss;
672 sum += loss_scaled;
673 sum_squared += loss_scaled*loss_scaled;
674 }
675
676 const double mean = sum/nbatches;
677 *loss = result->loss_per_datapoint ? mean : sum;
678
679 if (!unc) {
680 return;
681 }
682
683 if (nbatches < 2) {
684 *unc = NAN;
685 return;
686 }
687
688 const double var_sum = sum_squared/nbatches - mean*mean; // variance without Bessel's correction, i.e. nbatches/(nbatches-1)
689 *unc = result->loss_per_datapoint ? sqrt(var_sum / (nbatches - 1)) : sqrt(var_sum * nbatches/(nbatches - 1));
690}
691
692void ggml_opt_result_pred(ggml_opt_result_t result, int32_t * pred) {
693 for (size_t i = 0; i < result->pred.size(); ++i) {
694 pred[i] = result->pred[i];
695 }
696}
697
698void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, double * unc) {
699 *accuracy = result->ncorrect >= 0 ? double(result->ncorrect) / double(result->ndata) : NAN;
700
701 if (!unc) {
702 return;
703 }
704
705 *unc = result->ncorrect >= 0 && result->ndata >= 2 ?
706 sqrt((*accuracy) * (1.0 - (*accuracy)) / double(result->ndata - 1)) : NAN;
707}
708
709// ====== Computation ======
710
711void ggml_opt_prepare_alloc(
712 ggml_opt_context_t opt_ctx,
713 struct ggml_context * ctx_compute,
714 struct ggml_cgraph * gf,
715 struct ggml_tensor * inputs,
716 struct ggml_tensor * outputs) {
717 GGML_ASSERT(!opt_ctx->static_graphs);
718 opt_ctx->ctx_compute = ctx_compute;
719 opt_ctx->gf = gf;
720 opt_ctx->inputs = inputs;
721 opt_ctx->outputs = outputs;
722}
723
724void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) {
725 GGML_ASSERT(!opt_ctx->eval_ready);
726 if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period > 1 && opt_ctx->opt_i == 0) {
727 ggml_graph_reset(opt_ctx->gb_grad);
728 }
729 if (backward) {
730 const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
731 opt_ctx->build_type = opt_i_next == 0 ? GGML_OPT_BUILD_TYPE_OPT : GGML_OPT_BUILD_TYPE_GRAD;
732 } else {
733 opt_ctx->build_type = GGML_OPT_BUILD_TYPE_FORWARD;
734 }
735
736 if (!opt_ctx->static_graphs) {
737 ggml_opt_build(opt_ctx);
738 }
739
740 struct ggml_cgraph * graph = nullptr;
741 switch (opt_ctx->build_type) {
742 case GGML_OPT_BUILD_TYPE_FORWARD: {
743 graph = opt_ctx->gf;
744 } break;
745 case GGML_OPT_BUILD_TYPE_GRAD: {
746 graph = opt_ctx->gb_grad;
747 } break;
748 case GGML_OPT_BUILD_TYPE_OPT: {
749 graph = opt_ctx->gb_opt;
750 } break;
751 }
752 GGML_ASSERT(graph);
753
754 if (opt_ctx->allocated_graph == graph) {
755 opt_ctx->eval_ready = true;
756 return;
757 }
758
759 ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
760
761 if (opt_ctx->static_graphs) {
762 ggml_init_params params = {
763 /*.mem_size =*/ graph->size*ggml_tensor_overhead() + ggml_graph_overhead_custom(graph->size, graph->grads),
764 /*.mem_buffer =*/ nullptr,
765 /*.no_alloc =*/ true,
766 };
767 ggml_free(opt_ctx->ctx_copy);
768 opt_ctx->ctx_copy = ggml_init(params);
769
770 opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
771 } else {
772 opt_ctx->allocated_graph_copy = graph;
773 }
774
775 ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
776 opt_ctx->allocated_graph = graph;
777
778 opt_ctx->eval_ready = true;
779}
780
781void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) {
782 GGML_ASSERT(opt_ctx->eval_ready);
783 if (opt_ctx->allocated_graph == opt_ctx->gb_opt) {
784 const ggml_opt_optimizer_params & opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
785
786 switch (opt_ctx->optimizer) {
787 case GGML_OPT_OPTIMIZER_TYPE_ADAMW: {
788 GGML_ASSERT(opt_pars.adamw.alpha > 0.0f);
789 GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
790 GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
791 GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
792 GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
793 GGML_ASSERT(opt_pars.adamw.eps >= 0.0f);
794 GGML_ASSERT(opt_pars.adamw.wd >= 0.0f);
795 GGML_ASSERT(opt_pars.adamw.wd <= 1.0f);
796
797 // beta1, beta2 after applying warmup
798 const float beta1h = 1.0f / (1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
799 const float beta2h = 1.0f / (1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
800
801 float * adamw_par_data = ggml_get_data_f32(opt_ctx->opt_step_params);
802 adamw_par_data[0] = opt_pars.adamw.alpha;
803 adamw_par_data[1] = opt_pars.adamw.beta1;
804 adamw_par_data[2] = opt_pars.adamw.beta2;
805 adamw_par_data[3] = opt_pars.adamw.eps;
806 adamw_par_data[4] = opt_pars.adamw.wd;
807 adamw_par_data[5] = beta1h;
808 adamw_par_data[6] = beta2h;
809 } break;
810 case GGML_OPT_OPTIMIZER_TYPE_SGD: {
811 GGML_ASSERT(opt_pars.sgd.alpha > 0.0f);
812 GGML_ASSERT(opt_pars.sgd.wd >= 0.0f);
813 GGML_ASSERT(opt_pars.sgd.wd <= 1.0f);
814 float * sgd = ggml_get_data_f32(opt_ctx->opt_step_params);
815 sgd[0] = opt_pars.sgd.alpha;
816 sgd[1] = opt_pars.sgd.wd;
817 } break;
818 default:
819 GGML_ABORT("fatal error");
820 }
821 }
822
823 ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
824 opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
825 opt_ctx->opt_i = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
826
827 if (!opt_ctx->static_graphs) {
828 opt_ctx->gf = nullptr;
829 opt_ctx->gb_grad = nullptr;
830 opt_ctx->gb_opt = nullptr;
831 opt_ctx->allocated_graph = nullptr;
832 opt_ctx->allocated_graph_copy = nullptr;
833 }
834
835 opt_ctx->eval_ready = false;
836
837 if (!result) {
838 return;
839 }
840
841 if (result->ndata == 0) {
842 result->loss_per_datapoint = opt_ctx->loss_per_datapoint;
843 result->opt_period = opt_ctx->opt_period;
844 } else {
845 GGML_ASSERT(result->loss_per_datapoint == opt_ctx->loss_per_datapoint);
846 GGML_ASSERT(result->opt_period == opt_ctx->opt_period);
847 }
848
849 const int64_t ndata = opt_ctx->outputs->ne[1];
850 GGML_ASSERT(result->ndata == ndata*int64_t(result->loss.size()) && "varying batch size not supported");
851 result->ndata += ndata;
852
853 GGML_ASSERT(ggml_is_scalar(opt_ctx->loss));
854 GGML_ASSERT(opt_ctx->loss->type == GGML_TYPE_F32);
855 float loss;
856 ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
857 result->loss.push_back(loss);
858
859 if (opt_ctx->pred) {
860 GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
861 std::vector<int32_t> pred(ndata);
862 ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
863 result->pred.insert(result->pred.end(), pred.begin(), pred.end());
864 }
865
866 if (!opt_ctx->ncorrect || result->ncorrect < 0) {
867 result->ncorrect = -1;
868 return;
869 }
870
871 GGML_ASSERT(ggml_is_scalar(opt_ctx->ncorrect));
872 GGML_ASSERT(opt_ctx->ncorrect->type == GGML_TYPE_I64);
873 int64_t ncorrect;
874 ggml_backend_tensor_get(opt_ctx->ncorrect, &ncorrect, 0, ggml_nbytes(opt_ctx->ncorrect));
875 result->ncorrect += ncorrect;
876}
877
878// ====== High-Level Functions ======
879
880void ggml_opt_epoch(
881 ggml_opt_context_t opt_ctx,
882 ggml_opt_dataset_t dataset,
883 ggml_opt_result_t result_train,
884 ggml_opt_result_t result_eval,
885 int64_t idata_split,
886 ggml_opt_epoch_callback callback_train,
887 ggml_opt_epoch_callback callback_eval) {
888 GGML_ASSERT(ggml_opt_static_graphs(opt_ctx) && "ggml_opt_epoch requires static graphs");
889 struct ggml_tensor * inputs = ggml_opt_inputs(opt_ctx);
890 struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
891 struct ggml_tensor * data = ggml_opt_dataset_data(dataset);
892 GGML_ASSERT(data->ne[0] == inputs->ne[0]);
893
894 const int64_t ndata = data->ne[1];
895 const int64_t ndata_batch = inputs->ne[1];
896
897 GGML_ASSERT(data->ne[1] % inputs->ne[1] == 0);
898 const int64_t nbatches = ndata/ndata_batch;
899
900 idata_split = idata_split < 0 ? ndata : idata_split;
901 GGML_ASSERT(idata_split % ndata_batch == 0);
902 const int64_t ibatch_split = idata_split / ndata_batch;
903
904 int64_t ibatch = 0;
905 int64_t t_loop_start = ggml_time_us();
906 for (; ibatch < ibatch_split; ++ibatch) {
907 ggml_opt_alloc(opt_ctx, /*backward =*/ true);
908 ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
909 ggml_opt_eval(opt_ctx, result_train);
910 if (callback_train) {
911 callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
912 }
913 }
914 t_loop_start = ggml_time_us();
915 for (; ibatch < nbatches; ++ibatch) {
916 ggml_opt_alloc(opt_ctx, /*backward =*/ false);
917 ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
918 ggml_opt_eval(opt_ctx, result_eval);
919 if (callback_eval) {
920 callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
921 }
922 }
923}
924
925void ggml_opt_epoch_callback_progress_bar(
926 bool train,
927 ggml_opt_context_t opt_ctx,
928 ggml_opt_dataset_t dataset,
929 ggml_opt_result_t result,
930 int64_t ibatch,
931 int64_t ibatch_max,
932 int64_t t_start_us) {
933 fprintf(stderr, "%s[", train ? "train: " : "val: ");
934
935 // The progress bar consists of partially filled blocks, unicode has 8 separate fill levels.
936 constexpr int64_t bar_length = 8;
937 const int64_t ibatch8 = 8 * ibatch;
938 for (int64_t j = 0; j < bar_length; ++j) {
939 if (ibatch_max * (8*j + 8) / bar_length < ibatch8) {
940 fprintf(stderr, "\u2588"); // full block
941 } else if (ibatch_max * (8*j + 7) / bar_length < ibatch8) {
942 fprintf(stderr, "\u2589"); // 7/8 filled
943 } else if (ibatch_max * (8*j + 6) / bar_length < ibatch8) {
944 fprintf(stderr, "\u258A"); // 6/8 filled
945 } else if (ibatch_max * (8*j + 5) / bar_length < ibatch8) {
946 fprintf(stderr, "\u258B"); // 5/8 filled
947 } else if (ibatch_max * (8*j + 4) / bar_length < ibatch8) {
948 fprintf(stderr, "\u258C"); // 4/8 filled
949 } else if (ibatch_max * (8*j + 3) / bar_length < ibatch8) {
950 fprintf(stderr, "\u258D"); // 3/8 filled
951 } else if (ibatch_max * (8*j + 2) / bar_length < ibatch8) {
952 fprintf(stderr, "\u258E"); // 2/8 filled
953 } else if (ibatch_max * (8*j + 1) / bar_length < ibatch8) {
954 fprintf(stderr, "\u258F"); // 1/8 filled
955 } else {
956 fprintf(stderr, " ");
957 }
958 }
959
960 const int64_t batch_size = ggml_opt_inputs(opt_ctx)->ne[1];
961 const int64_t idata = ibatch*batch_size;
962 const int64_t idata_max = ibatch_max*batch_size;
963
964 double loss;
965 double loss_unc;
966 ggml_opt_result_loss(result, &loss, &loss_unc);
967
968 double accuracy;
969 double accuracy_unc;
970 ggml_opt_result_accuracy(result, &accuracy, &accuracy_unc);
971
972 const int64_t t_ibatch_us = ggml_time_us() - t_start_us;
973 int64_t t_ibatch_s = t_ibatch_us / 1000000;
974 const int64_t t_ibatch_h = t_ibatch_s / 3600;
975 t_ibatch_s -= t_ibatch_h * 3600;
976 const int64_t t_ibatch_m = t_ibatch_s / 60;
977 t_ibatch_s -= t_ibatch_m * 60;
978
979 const int64_t t_eta_us = t_ibatch_us * (ibatch_max - ibatch)/ibatch;
980 int64_t t_eta_s = t_eta_us / 1000000;
981 const int64_t t_eta_h = t_eta_s / 3600;
982 t_eta_s -= t_eta_h * 3600;
983 const int64_t t_eta_m = t_eta_s / 60;
984 t_eta_s -= t_eta_m * 60;
985
986 fprintf(stderr, "] data=%07" PRId64 "/%07" PRId64 " loss=%.5lf±%.5lf acc=%.2lf±%.2lf%% "
987 "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " \r",
988 idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
989 t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
990 if (ibatch == ibatch_max) {
991 fprintf(stderr, "\n");
992 }
993 fflush(stderr);
994
995 GGML_UNUSED(dataset);
996}
997
998void ggml_opt_fit(
999 ggml_backend_sched_t backend_sched,
1000 ggml_context * ctx_compute,
1001 ggml_tensor * inputs,
1002 ggml_tensor * outputs,
1003 ggml_opt_dataset_t dataset,
1004 enum ggml_opt_loss_type loss_type,
1005 enum ggml_opt_optimizer_type optimizer,
1006 ggml_opt_get_optimizer_params get_opt_pars,
1007 int64_t nepoch,
1008 int64_t nbatch_logical,
1009 float val_split,
1010 bool silent) {
1011 ggml_time_init();
1012 const int64_t t_start_us = ggml_time_us();
1013
1014 const int64_t ndata = ggml_opt_dataset_data(dataset)->ne[1];
1015 const int64_t nbatch_physical = inputs->ne[1];
1016 GGML_ASSERT(ndata % nbatch_logical == 0);
1017 GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
1018
1019 const int64_t opt_period = nbatch_logical / nbatch_physical;
1020 const int64_t nbatches_logical = ndata / nbatch_logical;
1021
1022 GGML_ASSERT(val_split >= 0.0f);
1023 GGML_ASSERT(val_split < 1.0f);
1024 const int64_t ibatch_split = int64_t(((1.0f - val_split) * nbatches_logical)) * opt_period; // train <-> val split index (physical)
1025 const int64_t idata_split = ibatch_split * nbatch_physical;
1026
1027 int64_t epoch = 1;
1028
1029 ggml_opt_params params = ggml_opt_default_params(backend_sched, loss_type);
1030 params.ctx_compute = ctx_compute;
1031 params.inputs = inputs;
1032 params.outputs = outputs;
1033 params.opt_period = opt_period;
1034 params.get_opt_pars = get_opt_pars;
1035 params.get_opt_pars_ud = &epoch;
1036 params.optimizer = optimizer;
1037 ggml_opt_context_t opt_ctx = ggml_opt_init(params);
1038
1039 // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
1040 if (nbatch_logical < ndata) {
1041 ggml_opt_dataset_shuffle(opt_ctx, dataset, -1); // Shuffle all data (train + validation).
1042 }
1043
1044 ggml_opt_result_t result_train = ggml_opt_result_init();
1045 ggml_opt_result_t result_val = ggml_opt_result_init();
1046
1047 ggml_opt_epoch_callback epoch_callback = silent ? nullptr : ggml_opt_epoch_callback_progress_bar;
1048
1049 for (; epoch <= nepoch; ++epoch) {
1050 if (nbatch_logical < idata_split) {
1051 ggml_opt_dataset_shuffle(opt_ctx, dataset, idata_split);
1052 }
1053
1054 ggml_opt_result_reset(result_train);
1055 ggml_opt_result_reset(result_val);
1056
1057 if (!silent) {
1058 fprintf(stderr, "%s: epoch %04" PRId64 "/%04" PRId64 ":\n", __func__, epoch, nepoch);
1059 }
1060 ggml_opt_epoch(opt_ctx, dataset, result_train, result_val, idata_split, epoch_callback, epoch_callback);
1061 if (!silent) {
1062 fprintf(stderr, "\n");
1063 }
1064 }
1065
1066 if (!silent) {
1067 int64_t t_total_s = (ggml_time_us() - t_start_us) / 1000000;
1068 const int64_t t_total_h = t_total_s / 3600;
1069 t_total_s -= t_total_h * 3600;
1070 const int64_t t_total_m = t_total_s / 60;
1071 t_total_s -= t_total_m * 60;
1072 fprintf(stderr, "%s: training took %02" PRId64 ":%02" PRId64 ":%02" PRId64 "\n", __func__, t_total_h, t_total_m, t_total_s);
1073 }
1074
1075 ggml_opt_free(opt_ctx);
1076 ggml_opt_result_free(result_train);
1077 ggml_opt_result_free(result_val);
1078}
1079
1080enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t c) {
1081 return c->optimizer;
1082}
1083
1084GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type o) {
1085 switch (o) {
1086 case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
1087 return "adamw";
1088 case GGML_OPT_OPTIMIZER_TYPE_SGD:
1089 return "sgd";
1090 default:
1091 return "undefined";
1092 };
1093}