unit test tokenization + main scripts in example

.gitignore

@@ -4,6 +4,8 @@ build/
 
 bark
 encodec
+main
+tests/test-tokenizer
 
 *.o
 *.plist

CMakeLists.txt

@@ -75,7 +75,7 @@ option(BARK_CUDA_DMMV_F16 "bark: use 16 bit floats for dmmv CU
 option(BARK_CLBLAST "bark: use CLBlast" OFF)
 option(BARK_METAL "bark: use Metal" OFF)
 
-# option(BARK_BUILD_TESTS "bark: build tests" ${BARK_STANDALONE})
+option(BARK_BUILD_TESTS "bark: build tests" ${BARK_STANDALONE})
 
 #
 # Build info header
@@ -518,7 +518,7 @@ install(
 # programs, examples and tests
 #
 
-# if (BARK_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-#  include(CTest)
-#  add_subdirectory(tests)
-# endif ()
+if (BARK_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
+ include(CTest)
+ add_subdirectory(tests)
+endif ()

Makefile

@@ -2,8 +2,7 @@
 BUILD_TARGETS = bark
 
 # Binaries only useful for tests
-# TEST_TARGETS = tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0
-TEST_TARGETS =
+TEST_TARGETS = tests/test-tokenizer
 
 default: $(BUILD_TARGETS)
 
@@ -302,5 +301,17 @@ bark.o: bark.cpp bark.h
 clean:
  rm -vf *.o *.so *.dll encodec bark
 
-bark: bark.cpp  encodec.o ggml.o $(OBJS)
+bark: bark.cpp encodec.o ggml.o $(OBJS)
  $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+main: examples/main.cpp ggml.o bark.o encodec.o $(OBJS)
+ $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
+#
+# Test
+#
+
+tests: $(TEST_TARGETS)
+
+tests/test-tokenizer: tests/test-tokenizer.cpp ggml.o bark.o encodec.o $(OBJS)
+ $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)

bark.cpp

@@ -1460,56 +1460,3 @@ bool bark_generate_audio(
 
  return true;
 }
-
-int main() {
- const int64_t t_main_start_us = ggml_time_us();
-
- int64_t t_load_us = 0;
- int64_t t_eval_us = 0;
-
- bark_model model;
- std::string fname = "./ggml_weights";
-
- // load the model
- {
- const int64_t t_start_us = ggml_time_us();
-
- if(!bark_model_load(fname, model)) {
- fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, fname.c_str());
- return 1;
- }
-
- t_load_us = ggml_time_us() - t_start_us;
- }
-
- printf("\n");
-
- // forward pass
- const std::string prompt = "This is an audio";
- {
- const int64_t t_eval_us_start = ggml_time_us();
-
- // call to generate audio
- bark_generate_audio(model, model.vocab, prompt.data(), 4);
-
- t_eval_us = ggml_time_us() - t_eval_us_start;
- }
-
- // report timing
- {
- const int64_t t_main_end_us = ggml_time_us();
-
- printf("\n\n");
- printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
- printf("%s: eval time = %8.2f ms\n", __func__, t_eval_us/1000.0f);
- printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
- }
-
- // TODO: write wrapper
- ggml_free(model.coarse_model.ctx);
- ggml_free(model.fine_model.ctx);
- ggml_free(model.text_model.ctx);
- ggml_free(model.codec_model.ctx);
-
- return 0;
-}

bark.h

@@ -114,9 +114,17 @@ struct bark_model {
 
 bool gpt_model_load(const std::string& fname, gpt_model& model, bark_vocab& vocab, bool has_vocab);
 
+bool bark_model_load(const std::string & dirname, bark_model & model);
+
 void bert_tokenize(
  const bark_vocab& vocab, 
  const char * text, 
  int32_t * tokens, 
  int32_t * n_tokens, 
  int32_t n_max_tokens);
+
+bool bark_generate_audio(
+ bark_model model,
+ const bark_vocab& vocab,
+ const char * text,
+ const int n_threads);

examples/main.cpp

@@ -0,0 +1,55 @@
+#include "ggml.h"
+#include "bark.h"
+
+int main() {
+ const int64_t t_main_start_us = ggml_time_us();
+
+ int64_t t_load_us = 0;
+ int64_t t_eval_us = 0;
+
+ bark_model model;
+ std::string fname = "./ggml_weights";
+
+ // load the model
+ {
+ const int64_t t_start_us = ggml_time_us();
+
+ if(!bark_model_load(fname, model)) {
+ fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, fname.c_str());
+ return 1;
+ }
+
+ t_load_us = ggml_time_us() - t_start_us;
+ }
+
+ printf("\n");
+
+ // forward pass
+ const std::string prompt = "This is an audio";
+ {
+ const int64_t t_eval_us_start = ggml_time_us();
+
+ // call to generate audio
+ bark_generate_audio(model, model.vocab, prompt.data(), 4);
+
+ t_eval_us = ggml_time_us() - t_eval_us_start;
+ }
+
+ // report timing
+ {
+ const int64_t t_main_end_us = ggml_time_us();
+
+ printf("\n\n");
+ printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+ printf("%s: eval time = %8.2f ms\n", __func__, t_eval_us/1000.0f);
+ printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+ }
+
+ // TODO: write wrapper
+ ggml_free(model.coarse_model.ctx);
+ ggml_free(model.fine_model.ctx);
+ ggml_free(model.text_model.ctx);
+ ggml_free(model.codec_model.ctx);
+
+ return 0;
+}

tests/test-tokenizer.cpp

@@ -8,11 +8,11 @@
 static const std::map<std::string, std::vector<bark_vocab::id>> & k_tests()
 {
  static std::map<std::string, std::vector<bark_vocab::id>> _k_tests = {
- { "Hello World!", { 31178, 11356, 106, }, },
- { "Hello World", { 31178, 11356, }, },
- { " Hello World!", { 31178, 11356, 106, }, },
- { "this is an audio generated by bark", { 10531, 10124, 10151, 23685, 48918, 10155, 18121, 10174 }, },
- { "l'Amérique si c'est un rêve je le saurai ", { 180, 112, 28426, 10294, 171, 112, 10176, 10119, 89952, 10144, 10141, 11731, 33186 }, },
+ { "Hello World!", { 31178, 11356, 106,  }, },
+ { "Hello World", { 31178, 11356,  }, },
+ { " Hello World!", { 31178, 11356, 106,  }, },
+ { "this is an audio generated by bark", { 10531, 10124, 10151, 23685, 48918, 10155, 18121, 10174, }, },
+ { "l'Amérique si c'est un rêve je le saurai ", { 180, 112, 28426, 10294, 171, 112, 10176, 10119, 89952, 10144, 10141, 11731, 33186, }, },
  };
  return _k_tests;
 };
@@ -33,7 +33,7 @@ int main(int argc, char **argv) {
  {
  if(!gpt_model_load(fname, model.text_model, model.vocab, true)) {
  fprintf(stderr, "%s: invalid model file '%s' (bad text)\n", __func__, fname.c_str());
- return false;
+ return 1;
  }
  model.memsize += model.text_model.memsize;
  }