def on_test_epoch_end(self, trainer, pl_module):
results = {
remove_postfix(k, '_epoch'): v
for k, v in trainer.logged_metrics.items()
if k.startswith('test_') and not k.endswith('_step')
}
tune.report(**results)
def align_text_fn(item, score_threshold, debug=False):
audio_path, recognition_text = item
#print("audio_path=",audio_path)
audio_dir = os.path.dirname(audio_path)
base_dir = os.path.dirname(audio_dir)
news_path = remove_postfix(audio_path.replace("audio", "assets"))
news_path = os.path.splitext(news_path)[0] + ".txt"
strip_fn = lambda line: line.strip().replace('"', '').replace("'", "")
#candidates = [strip_fn(line) for line in open(news_path, encoding='cp949').readlines()]
candidates = [
strip_fn(line)
for line in open(news_path, encoding='utf-8').readlines()
]
scores = { candidate: similarity(candidate, recognition_text) \
for candidate in candidates}
sorted_scores = sorted(scores.items(), key=operator.itemgetter(1))[::-1]
#jck add
if len(sorted_scores) < 2:
print("===========================================================")
print(">> error : candidates less than 2, news_path=[", news_path, "]")
print("===========================================================")
first, second = sorted_scores[0], sorted_scores[1]
if first[1] > second[1] and first[1] >= score_threshold:
found_text, score = first
aligned_text = search_optimal(found_text, recognition_text)
if debug:
print(" ", audio_path)
print(" ", recognition_text)
print("=> ", found_text)
print("==>", aligned_text)
print("=" * 30)
if aligned_text is not None:
result = {audio_path: add_punctuation(aligned_text)}
elif abs(
len(text_to_sequence(found_text)) -
len(text_to_sequence(recognition_text))) > 10:
result = {}
else:
result = {
audio_path: [add_punctuation(found_text), recognition_text]
}
else:
result = {}
if len(result) == 0:
result = {audio_path: [recognition_text]}
return result
def align_text_fn(item, score_threshold, debug=False):
audio_path, recognition_text = item
audio_name = os.path.splitext(os.path.basename(audio_path))[0]
audio_dir = os.path.dirname(audio_path)
base_dir = os.path.dirname(audio_dir)
news_path = remove_postfix(audio_path.replace("wavs", "assets"))
news_path = os.path.splitext(news_path)[0] + ".txt"
strip_fn = lambda line: line.strip().replace('"', '').replace("'", "")
candidates = [
strip_fn(line)
for line in open(news_path, 'r', encoding='utf-8').readlines()
]
scores = { candidate: similarity(candidate, recognition_text) \
for candidate in candidates}
sorted_scores = sorted(scores.items(), key=operator.itemgetter(1))[::-1]
first, second = sorted_scores[0], sorted_scores[1]
aligned_text = ""
if first[1] > second[1] and first[1] >= score_threshold:
found_text, score = first
aligned_text = search_optimal(found_text, recognition_text)
if debug:
print(" ", audio_path)
print(" ", recognition_text)
print("=> ", found_text)
print("==>", aligned_text)
print("=" * 30)
if aligned_text is not None:
aligned_text = add_punctuation(aligned_text)
elif abs(
len(text_to_sequence(found_text)) -
len(text_to_sequence(recognition_text))) > 10:
aligned_text = ""
else:
aligned_text = add_punctuation(found_text)
else:
aligned_text = ""
if aligned_text == "":
aligned_text = recognition_text
# Normalization
if debug:
print("aligned ==>", aligned_text)
aligned_text_noramlized = normalize(aligned_text)
result = {audio_name: [aligned_text, aligned_text_noramlized]}
return result
def on_epoch_end(self, trainer, pl_module):
results = {
remove_postfix(k, '_epoch'): v
for k, v in trainer.logged_metrics.items()
if (k.startswith('train_') or k.startswith('val_'))
and not k.endswith('_step')
}
results['mean_loss'] = results.get('val_loss', results['train_loss'])
if 'val_accuracy' in results:
results['mean_accuracy'] = results['val_accuracy']
# Checkpointing should be done *before* reporting
# https://docs.ray.io/en/master/tune/api_docs/trainable.html
with tune.checkpoint_dir(step=trainer.current_epoch) as checkpoint_dir:
trainer.save_checkpoint(
os.path.join(checkpoint_dir,
f"{type(pl_module).__name__}.ckpt"))
tune.report(**results)
def write_mkdocs_yaml_file(docs_directory) -> bool:
yml_directory = utils.remove_postfix(docs_directory, 'docs/')
base_file = yml_directory + 'base.yml'
mkdocs_file = yml_directory + 'mkdocs.yml'
if os.path.isfile(base_file):
shutil.copyfile(base_file, mkdocs_file)
else:
print(f'Could not find base.yml file in {yml_directory}.')
return False
try:
with open(mkdocs_file, 'a') as file:
ls = utils.yml_list_from_directory(docs_directory)
if len(ls) > 0:
file.writelines(ls)
return True
else:
print('Call to "utils.yml_list_from_directory" failed.')
return False
except OSError:
print(f'Could open file {mkdocs_file}.')
return False
def align_text_fn(item, score_threshold, debug=False):
audio_path, recognition_text = item
audio_dir = os.path.dirname(audio_path)
base_dir = os.path.dirname(audio_dir)
news_path = remove_postfix(audio_path.replace("audio", "assets"))
news_path = os.path.splitext(news_path)[0] + ".txt"
strip_fn = lambda line: line.strip().replace('"', '').replace("'", "")
candidates = [
strip_fn(line)
for line in open(news_path, encoding='cp949').readlines()
]
scores = { candidate: similarity(candidate, recognition_text) \
for candidate in candidates}
sorted_scores = sorted(scores.items(), key=operator.itemgetter(1))[::-1]
first, second = sorted_scores[0], sorted_scores[1]
'''
if first[1] > second[1] and first[1] >= score_threshold:
found_text, score = first
aligned_text, align_score = search_optimal(found_text, recognition_text, by_word=True)
'''
best_aligned_text = ''
found_text = ''
best_simscore = 0.0
for candidate, simscore in sorted_scores:
aligned_text, simscore = search_optimal(candidate, recognition_text)
if simscore > best_simscore:
best_aligned_text = aligned_text
best_simscore = simscore
found_text = candidate
if simscore == 1.0:
break
aligned_text = best_aligned_text
if len(best_aligned_text) != 0 and best_simscore >= score_threshold:
if debug:
print(" ", audio_path)
print(" ", recognition_text)
print("=> ", found_text)
print("==>", aligned_text)
print("=" * 30)
if aligned_text is not None:
result = {audio_path: add_punctuation(aligned_text)}
elif abs(
len(text_to_sequence(found_text)) -
len(text_to_sequence(recognition_text))) > 10:
result = {}
else:
result = {
audio_path: [add_punctuation(found_text), recognition_text]
}
else:
result = {}
if len(result) == 0:
result = {audio_path: [recognition_text]}
return result