xinference 1.2.0__py3-none-any.whl → 1.2.2__py3-none-any.whl

xinference/thirdparty/melo/text/japanese_bert.py

@@ -0,0 +1,49 @@
+import torch
+from transformers import AutoTokenizer, AutoModelForMaskedLM
+import sys
+
+
+models = {}
+tokenizers = {}
+def get_bert_feature(text, word2ph, device=None, model_id='tohoku-nlp/bert-base-japanese-v3'):
+    global model
+    global tokenizer
+
+    if (
+        sys.platform == "darwin"
+        and torch.backends.mps.is_available()
+        and device == "cpu"
+    ):
+        device = "mps"
+    if not device:
+        device = "cuda"
+    if model_id not in models:
+        model = AutoModelForMaskedLM.from_pretrained(model_id).to(
+            device
+        )
+        models[model_id] = model
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+        tokenizers[model_id] = tokenizer
+    else:
+        model = models[model_id]
+        tokenizer = tokenizers[model_id]
+
+
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors="pt")
+        tokenized = tokenizer.tokenize(text)
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = model(**inputs, output_hidden_states=True)
+        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+
+    assert inputs["input_ids"].shape[-1] == len(word2ph), f"{inputs['input_ids'].shape[-1]}/{len(word2ph)}"
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T

xinference/thirdparty/melo/text/ko_dictionary.py

@@ -0,0 +1,44 @@
+# coding: utf-8
+# Add the word you want to the dictionary.
+etc_dictionary = {"1+1": "원플러스원", "2+1": "투플러스원"}
+
+
+english_dictionary = {
+    "KOREA": "코리아",
+    "IDOL": "아이돌",
+    "IT": "아이티",
+    "IQ": "아이큐",
+    "UP": "업",
+    "DOWN": "다운",
+    "PC": "피씨",
+    "CCTV": "씨씨티비",
+    "SNS": "에스엔에스",
+    "AI": "에이아이",
+    "CEO": "씨이오",
+    "A": "에이",
+    "B": "비",
+    "C": "씨",
+    "D": "디",
+    "E": "이",
+    "F": "에프",
+    "G": "지",
+    "H": "에이치",
+    "I": "아이",
+    "J": "제이",
+    "K": "케이",
+    "L": "엘",
+    "M": "엠",
+    "N": "엔",
+    "O": "오",
+    "P": "피",
+    "Q": "큐",
+    "R": "알",
+    "S": "에스",
+    "T": "티",
+    "U": "유",
+    "V": "브이",
+    "W": "더블유",
+    "X": "엑스",
+    "Y": "와이",
+    "Z": "제트",
+}

xinference/thirdparty/melo/text/korean.py

@@ -0,0 +1,192 @@
+# Convert Japanese text to phonemes which is
+# compatible with Julius https://github.com/julius-speech/segmentation-kit
+import re
+import unicodedata
+
+from transformers import AutoTokenizer
+
+from . import punctuation, symbols
+
+
+from num2words import num2words
+from melo.text.ko_dictionary import english_dictionary, etc_dictionary
+from anyascii import anyascii
+from jamo import hangul_to_jamo
+
+def normalize(text):
+    text = text.strip()
+    text = re.sub("[⺀-⺙⺛-⻳⼀-⿕々〇〡-〩〸-〺〻㐀-䶵一-鿃豈-鶴侮-頻並-龎]", "", text)
+    text = normalize_with_dictionary(text, etc_dictionary)
+    text = normalize_english(text)
+    text = text.lower()
+    return text
+
+
+def normalize_with_dictionary(text, dic):
+    if any(key in text for key in dic.keys()):
+        pattern = re.compile("|".join(re.escape(key) for key in dic.keys()))
+        return pattern.sub(lambda x: dic[x.group()], text)
+    return text
+
+
+def normalize_english(text):
+    def fn(m):
+        word = m.group()
+        if word in english_dictionary:
+            return english_dictionary.get(word)
+        return word
+
+    text = re.sub("([A-Za-z]+)", fn, text)
+    return text
+
+
+g2p_kr = None
+def korean_text_to_phonemes(text, character: str = "hangeul") -> str:
+    """
+
+    The input and output values look the same, but they are different in Unicode.
+
+    example :
+
+        input = '하늘' (Unicode : \ud558\ub298), (하 + 늘)
+        output = '하늘' (Unicode :\u1112\u1161\u1102\u1173\u11af), (ᄒ + ᅡ + ᄂ + ᅳ + ᆯ)
+
+    """
+    global g2p_kr  # pylint: disable=global-statement
+    if g2p_kr is None:
+        from g2pkk import G2p
+
+        g2p_kr = G2p()
+
+    if character == "english":
+        from anyascii import anyascii
+        text = normalize(text)
+        text = g2p_kr(text)
+        text = anyascii(text)
+        return text
+
+    text = normalize(text)
+    text = g2p_kr(text)
+    text = list(hangul_to_jamo(text))  # '하늘' --> ['ᄒ', 'ᅡ', 'ᄂ', 'ᅳ', 'ᆯ']
+    return "".join(text)
+
+def text_normalize(text):
+    # res = unicodedata.normalize("NFKC", text)
+    # res = japanese_convert_numbers_to_words(res)
+    # # res = "".join([i for i in res if is_japanese_character(i)])
+    # res = replace_punctuation(res)
+    text = normalize(text)
+    return text
+
+
+def distribute_phone(n_phone, n_word):
+    phones_per_word = [0] * n_word
+    for task in range(n_phone):
+        min_tasks = min(phones_per_word)
+        min_index = phones_per_word.index(min_tasks)
+        phones_per_word[min_index] += 1
+    return phones_per_word
+
+
+
+# tokenizer = AutoTokenizer.from_pretrained('cl-tohoku/bert-base-japanese-v3')
+
+model_id = 'kykim/bert-kor-base'
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+def g2p(norm_text):
+    tokenized = tokenizer.tokenize(norm_text)
+    phs = []
+    ph_groups = []
+    for t in tokenized:
+        if not t.startswith("#"):
+            ph_groups.append([t])
+        else:
+            ph_groups[-1].append(t.replace("#", ""))
+    word2ph = []
+    for group in ph_groups:
+        text = ""
+        for ch in group:
+            text += ch
+        if text == '[UNK]':
+            phs += ['_']
+            word2ph += [1]
+            continue
+        elif text in punctuation:
+            phs += [text]
+            word2ph += [1]
+            continue
+        # import pdb; pdb.set_trace()
+        # phonemes = japanese_text_to_phonemes(text)
+        # text = g2p_kr(text)
+        phonemes = korean_text_to_phonemes(text)
+        # import pdb; pdb.set_trace()
+        # # phonemes = [i for i in phonemes if i in symbols]
+        # for i in phonemes:
+        #     assert i in symbols, (group, norm_text, tokenized, i)
+        phone_len = len(phonemes)
+        word_len = len(group)
+
+        aaa = distribute_phone(phone_len, word_len)
+        assert len(aaa) == word_len
+        word2ph += aaa
+
+        phs += phonemes
+    phones = ["_"] + phs + ["_"]
+    tones = [0 for i in phones]
+    word2ph =  [1] + word2ph + [1]
+    assert len(word2ph) == len(tokenized) + 2
+    return phones, tones, word2ph
+
+def get_bert_feature(text, word2ph, device='cuda'):
+    from . import japanese_bert
+    return japanese_bert.get_bert_feature(text, word2ph, device=device, model_id=model_id)
+
+
+if __name__ == "__main__":
+    # tokenizer = AutoTokenizer.from_pretrained("./bert/bert-base-japanese-v3")
+    from text.symbols import symbols
+    text = "전 제 일의 가치와 폰타인 대중들이 한 일의 의미를 잘 압니다. 앞으로도 전 제 일에 자부심을 갖고 살아갈 겁니다"
+    import json
+
+    # genshin_data = json.load(open('/data/zwl/workspace/StarRail_Datasets/Index & Scripts/Index/1.3/Korean.json'))
+    genshin_data = json.load(open('/data/zwl/workspace/Genshin_Datasets/Index & Script/AI Hobbyist Version/Index/4.1/KR_output.json'))
+    from tqdm import tqdm
+    new_symbols = []
+    for key, item in tqdm(genshin_data.items()):
+        texts = item.get('voiceContent', '')
+        if isinstance(texts, list):
+            texts = ','.join(texts)
+        if texts is None:
+            continue
+        if len(texts) == 0:
+            continue
+
+        text = text_normalize(text)
+        phones, tones, word2ph = g2p(text)
+        bert = get_bert_feature(text, word2ph)
+        import  pdb; pdb.set_trace()
+        for ph in phones:
+            if ph not in symbols and ph not in new_symbols:
+                new_symbols.append(ph)
+                print('update!, now symbols:')
+                print(new_symbols)
+                with open('korean_symbol.txt', 'w') as f:
+                    f.write(f'{new_symbols}')
+
+        
+
+# if __name__ == '__main__':
+#     from pykakasi import kakasi
+#     # Initialize kakasi object
+#     kakasi = kakasi()
+
+#     # Set options for converting Chinese characters to Katakana
+#     kakasi.setMode("J", "H")  # Chinese to Katakana
+#     kakasi.setMode("K", "H")  # Hiragana to Katakana
+
+#     # Convert Chinese characters to Katakana
+#     conv = kakasi.getConverter()
+#     katakana_text = conv.do('ええ、僕はおきなと申します。こちらの小さいわらべは杏子。ご挨拶が遅れてしまいすみません。あなたの名は?')  # Replace with your Chinese text
+
+#     print(katakana_text)  # Output: ニーハオセカイ

xinference/thirdparty/melo/text/opencpop-strict.txt

@@ -0,0 +1,429 @@
+a	AA a
+ai	AA ai
+an	AA an
+ang	AA ang
+ao	AA ao
+ba	b a
+bai	b ai
+ban	b an
+bang	b ang
+bao	b ao
+bei	b ei
+ben	b en
+beng	b eng
+bi	b i
+bian	b ian
+biao	b iao
+bie	b ie
+bin	b in
+bing	b ing
+bo	b o
+bu	b u
+ca	c a
+cai	c ai
+can	c an
+cang	c ang
+cao	c ao
+ce	c e
+cei	c ei
+cen	c en
+ceng	c eng
+cha	ch a
+chai	ch ai
+chan	ch an
+chang	ch ang
+chao	ch ao
+che	ch e
+chen	ch en
+cheng	ch eng
+chi	ch ir
+chong	ch ong
+chou	ch ou
+chu	ch u
+chua	ch ua
+chuai	ch uai
+chuan	ch uan
+chuang	ch uang
+chui	ch ui
+chun	ch un
+chuo	ch uo
+ci	c i0
+cong	c ong
+cou	c ou
+cu	c u
+cuan	c uan
+cui	c ui
+cun	c un
+cuo	c uo
+da	d a
+dai	d ai
+dan	d an
+dang	d ang
+dao	d ao
+de	d e
+dei	d ei
+den	d en
+deng	d eng
+di	d i
+dia	d ia
+dian	d ian
+diao	d iao
+die	d ie
+ding	d ing
+diu	d iu
+dong	d ong
+dou	d ou
+du	d u
+duan	d uan
+dui	d ui
+dun	d un
+duo	d uo
+e	EE e
+ei	EE ei
+en	EE en
+eng	EE eng
+er	EE er
+fa	f a
+fan	f an
+fang	f ang
+fei	f ei
+fen	f en
+feng	f eng
+fo	f o
+fou	f ou
+fu	f u
+ga	g a
+gai	g ai
+gan	g an
+gang	g ang
+gao	g ao
+ge	g e
+gei	g ei
+gen	g en
+geng	g eng
+gong	g ong
+gou	g ou
+gu	g u
+gua	g ua
+guai	g uai
+guan	g uan
+guang	g uang
+gui	g ui
+gun	g un
+guo	g uo
+ha	h a
+hai	h ai
+han	h an
+hang	h ang
+hao	h ao
+he	h e
+hei	h ei
+hen	h en
+heng	h eng
+hong	h ong
+hou	h ou
+hu	h u
+hua	h ua
+huai	h uai
+huan	h uan
+huang	h uang
+hui	h ui
+hun	h un
+huo	h uo
+ji	j i
+jia	j ia
+jian	j ian
+jiang	j iang
+jiao	j iao
+jie	j ie
+jin	j in
+jing	j ing
+jiong	j iong
+jiu	j iu
+ju	j v
+jv	j v
+juan	j van
+jvan	j van
+jue	j ve
+jve	j ve
+jun	j vn
+jvn	j vn
+ka	k a
+kai	k ai
+kan	k an
+kang	k ang
+kao	k ao
+ke	k e
+kei	k ei
+ken	k en
+keng	k eng
+kong	k ong
+kou	k ou
+ku	k u
+kua	k ua
+kuai	k uai
+kuan	k uan
+kuang	k uang
+kui	k ui
+kun	k un
+kuo	k uo
+la	l a
+lai	l ai
+lan	l an
+lang	l ang
+lao	l ao
+le	l e
+lei	l ei
+leng	l eng
+li	l i
+lia	l ia
+lian	l ian
+liang	l iang
+liao	l iao
+lie	l ie
+lin	l in
+ling	l ing
+liu	l iu
+lo	l o
+long	l ong
+lou	l ou
+lu	l u
+luan	l uan
+lun	l un
+luo	l uo
+lv	l v
+lve	l ve
+ma	m a
+mai	m ai
+man	m an
+mang	m ang
+mao	m ao
+me	m e
+mei	m ei
+men	m en
+meng	m eng
+mi	m i
+mian	m ian
+miao	m iao
+mie	m ie
+min	m in
+ming	m ing
+miu	m iu
+mo	m o
+mou	m ou
+mu	m u
+na	n a
+nai	n ai
+nan	n an
+nang	n ang
+nao	n ao
+ne	n e
+nei	n ei
+nen	n en
+neng	n eng
+ni	n i
+nian	n ian
+niang	n iang
+niao	n iao
+nie	n ie
+nin	n in
+ning	n ing
+niu	n iu
+nong	n ong
+nou	n ou
+nu	n u
+nuan	n uan
+nun	n un
+nuo	n uo
+nv	n v
+nve	n ve
+o	OO o
+ou	OO ou
+pa	p a
+pai	p ai
+pan	p an
+pang	p ang
+pao	p ao
+pei	p ei
+pen	p en
+peng	p eng
+pi	p i
+pian	p ian
+piao	p iao
+pie	p ie
+pin	p in
+ping	p ing
+po	p o
+pou	p ou
+pu	p u
+qi	q i
+qia	q ia
+qian	q ian
+qiang	q iang
+qiao	q iao
+qie	q ie
+qin	q in
+qing	q ing
+qiong	q iong
+qiu	q iu
+qu	q v
+qv	q v
+quan	q van
+qvan	q van
+que	q ve
+qve	q ve
+qun	q vn
+qvn	q vn
+ran	r an
+rang	r ang
+rao	r ao
+re	r e
+ren	r en
+reng	r eng
+ri	r ir
+rong	r ong
+rou	r ou
+ru	r u
+rua	r ua
+ruan	r uan
+rui	r ui
+run	r un
+ruo	r uo
+sa	s a
+sai	s ai
+san	s an
+sang	s ang
+sao	s ao
+se	s e
+sen	s en
+seng	s eng
+sha	sh a
+shai	sh ai
+shan	sh an
+shang	sh ang
+shao	sh ao
+she	sh e
+shei	sh ei
+shen	sh en
+sheng	sh eng
+shi	sh ir
+shou	sh ou
+shu	sh u
+shua	sh ua
+shuai	sh uai
+shuan	sh uan
+shuang	sh uang
+shui	sh ui
+shun	sh un
+shuo	sh uo
+si	s i0
+song	s ong
+sou	s ou
+su	s u
+suan	s uan
+sui	s ui
+sun	s un
+suo	s uo
+ta	t a
+tai	t ai
+tan	t an
+tang	t ang
+tao	t ao
+te	t e
+tei	t ei
+teng	t eng
+ti	t i
+tian	t ian
+tiao	t iao
+tie	t ie
+ting	t ing
+tong	t ong
+tou	t ou
+tu	t u
+tuan	t uan
+tui	t ui
+tun	t un
+tuo	t uo
+wa	w a
+wai	w ai
+wan	w an
+wang	w ang
+wei	w ei
+wen	w en
+weng	w eng
+wo	w o
+wu	w u
+xi	x i
+xia	x ia
+xian	x ian
+xiang	x iang
+xiao	x iao
+xie	x ie
+xin	x in
+xing	x ing
+xiong	x iong
+xiu	x iu
+xu	x v
+xv	x v
+xuan	x van
+xvan	x van
+xue	x ve
+xve	x ve
+xun	x vn
+xvn	x vn
+ya	y a
+yan	y En
+yang	y ang
+yao	y ao
+ye	y E
+yi	y i
+yin	y in
+ying	y ing
+yo	y o
+yong	y ong
+you	y ou
+yu	y v
+yv	y v
+yuan	y van
+yvan	y van
+yue	y ve
+yve	y ve
+yun	y vn
+yvn	y vn
+za	z a
+zai	z ai
+zan	z an
+zang	z ang
+zao	z ao
+ze	z e
+zei	z ei
+zen	z en
+zeng	z eng
+zha	zh a
+zhai	zh ai
+zhan	zh an
+zhang	zh ang
+zhao	zh ao
+zhe	zh e
+zhei	zh ei
+zhen	zh en
+zheng	zh eng
+zhi	zh ir
+zhong	zh ong
+zhou	zh ou
+zhu	zh u
+zhua	zh ua
+zhuai	zh uai
+zhuan	zh uan
+zhuang	zh uang
+zhui	zh ui
+zhun	zh un
+zhuo	zh uo
+zi	z i0
+zong	z ong
+zou	z ou
+zu	z u
+zuan	z uan
+zui	z ui
+zun	z un
+zuo	z uo