app.py

from typing import List, Union
from pythainlp.tokenize import subword_tokenize,word_tokenize
from pythainlp.util import sound_syllable
from pythainlp.util import remove_tonemark
from pythainlp.khavee import KhaveeVerifier
import pythainlp as pythai
from pythainlp.tokenize import word_tokenize
from pythainlp.tokenize import subword_tokenize
from pythainlp.util import sound_syllable
from pythainlp.util import isthai
from pythainlp.transliterate import pronunciate
from pythainlp.spell import correct
from tqdm import tqdm
import numpy as np
import pandas as pd
kv = KhaveeVerifier()
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Thanravee/KarveeSaimai", local_files_only=False)
model = AutoModelForCausalLM.from_pretrained("Thanravee/KarveeSaimai", local_files_only=False)

# split text from \n to list and drop soi word ->  splitted wak list (no soi)
def split_klong(klong_text):
  splitted_klong = []
  klong_list = klong_text.split('-')
  klong_list = [klong for klong in klong_list if klong.strip()]
  for i in range(len(klong_list)):
    if i == 1 or i == 3 or i == 5:
      klong = klong_list[i]
      if klong[0] == ' ':
        klong = klong[1:]
      klong = klong.split(' ')
      splitted_klong.append(klong[0])
    else:
      splitted_klong.append(klong_list[i].replace(' ', ''))
  return splitted_klong

# subword tokenize wak with ssg and dict
def subword_token(wak, engine='ssg'):
  subword_tokenized = subword_tokenize(wak, engine='ssg')
  if len(subword_tokenized) != 5 and len(subword_tokenized) != 2:
      subword_tokenized = subword_tokenize(wak, engine='dict')
  return subword_tokenized


# check number of syllables -> [True, True, True, True, True, True, True, True] (len=8)
def subword_num(splitted_klong):
  checked = []
  two = [1,3,5]
  five = [0,2,4,6]
  for num in range(len(splitted_klong)):
    if num in two:
      checked.append(len(subword_token(splitted_klong[num])) == 2)
    elif num in five:
      checked.append(len(subword_token(splitted_klong[num])) == 5)
    elif num == 7:
      checked.append(len(subword_token(splitted_klong[num])) == 4)
  return checked

# check what word tone is
def find_tone(word):
  char_list = [*word]
  if "่" in char_list or sound_syllable(word) == 'dead':
    return "eak or dead"
  elif "้" in char_list:
    return "tou"
  else:
    return False

# check eaktou -> list[True, True, True, True, True, True, True, True] (len=8)
def check_eaktou(splitted_klong):
  checked = []
  for num in range(len(splitted_klong)):
    tokenzied_wak = subword_token(splitted_klong[num])
    if num == 0:
      checked.append(find_tone(tokenzied_wak[3]) == "eak or dead" and find_tone(tokenzied_wak[4]) == 'tou')
    elif num == 1:
      checked.append(True)
    elif num == 2:
      checked.append(find_tone(tokenzied_wak[1]) == "eak or dead")
    elif num == 3:
      checked.append(find_tone(tokenzied_wak[0]) == 'eak or dead' and find_tone(tokenzied_wak[1]) == 'tou')
    elif num == 4:
      checked.append(find_tone(tokenzied_wak[2]) == 'eak or dead')
    elif num == 5:
      checked.append(find_tone(tokenzied_wak[1]) == 'eak or dead')
    elif num == 6:
      checked.append(find_tone(tokenzied_wak[1]) == "eak or dead" and find_tone(tokenzied_wak[4]) == 'tou')
    elif num == 7:
      checked.append(find_tone(tokenzied_wak[0]) == "eak or dead" and find_tone(tokenzied_wak[1]) == 'tou')
  return checked

# last sound of wak from pronunciate tokenized last word of each wak
# ex [เสียงลือเสียงเล่าอ้าง] -> [อ้าง]
def sound_words(splitted_klong):
  sound_list = []
  for wak in splitted_klong:
    list_char = [*wak]
    if " " in list_char:
      wak = wak.split(" ")
      wak = wak[0]
    wak = word_tokenize(wak, engine="newmm")
    pronounce_word = pronunciate(wak[-1], engine="w2p")
    sound_list.append(pronounce_word.replace('ฺ', '').split('-')[-1])
  return sound_list

# check sampas -> [True, True, True]
# [0] = sampas wak 2-3, [1] = sampas wak 2-4, [2] sampas wak 4-7
def check_sampas(sound_list):
  checked = []
  if len(sound_list) > 2:
    checked.append(kv.is_sumpus(sound_list[1],sound_list[2]))
    if len(sound_list) > 4:
      checked.append(kv.is_sumpus(sound_list[1],sound_list[4]))
      if len(sound_list) > 6:
        checked.append(kv.is_sumpus(sound_list[3],sound_list[6]))
  else:
    checked.append(True)
  return checked

def main_check(klong_text):
  splitted_klong = split_klong(klong_text)
  checked_subword_num = subword_num(splitted_klong)
  if False in checked_subword_num:
    false_index = checked_subword_num.index(False)
    return 'syllable format error', false_index+1
  else:
    checked_eaktou = check_eaktou(splitted_klong)
    if False in checked_eaktou:
      false_index = checked_eaktou.index(False)
      return 'eaktou format error', false_index+1
    else:
      sound_list = sound_words(splitted_klong)
      checked_sampas = check_sampas(sound_list)
      if False in checked_sampas:
        wak_sampas = ['2 and 3', '2 and 5', '4 and 7']
        return 'sampas format error', wak_sampas[checked_sampas.index(False)]
      else:
        return True
    
    
def gen_prob_next_token(text:str, model, tokenizer):
  input_ids = tokenizer(text, return_tensors="pt")
  #look at tensor shape
  input_ids,input_ids['input_ids'].shape

  #get logit of the next token
  outputs = model(input_ids['input_ids'])
  logits = outputs.logits
  logits.shape #the size is equal to input token because it's predicting the next one

  #convert logit to prob; use the logits of the last input token
  import torch.nn.functional as F
  probs = F.softmax(logits[:, -1, :], dim=-1).squeeze()
  probs, probs.argmax()

  #match prob with vocab
  import pandas as pd
  df = pd.DataFrame(tokenizer.vocab.items(), columns=['token', 'token_id']).sort_values('token_id').reset_index(drop=True)

  df['prob'] = probs.detach().numpy()

  possible_token = df.sort_values('prob',ascending=False).reset_index()
  thai_only = [x if isthai(x) else None for x in possible_token['token']] # thai only
  possible_token['token'] = thai_only
  possible_token = possible_token.dropna()
  return possible_token


# filter broken word and get passed only 100 words
def gen_rules(probs, fast_gen=True):
  passed = []
  limiter = 5 if fast_gen else 100000000
  for prob in probs:
    if len(check_word(prob)) > 1 and len(subword_token(prob)) == 1 and '-' not in pronunciate(prob) and len(passed) <= limiter:
        passed.append(correct(prob))
  return passed


def check_word(word):
  alphabets = [alp for alp in [*word] if alp not in ['่','้','๊','๋','์']]
  if '์' in [*word]:
    alphabets = [*word][:-2]
  return alphabets

def generator(klong):
  prob = gen_prob_next_token(klong, model, tokenizer)
  new_prob = gen_rules(prob['token'].tolist())
  return new_prob

# get word with sampas
def get_sampassed(data:list, sampaswith):
  passed = []
  counter_exception = 0
  for possible_word in tqdm(data):
    possible_sampas = pronunciate(possible_word).split('-')[-1] # reduce word dimension
    sampaswith = pronunciate(sampaswith).split('-')[-1] # reduce word dimension
    try:
      if kv.is_sumpus(possible_sampas, sampaswith):
        passed.append(possible_word)
    except IndexError:
      counter_exception += 1
      continue
  assert len(passed) != counter_exception # if this failed mena that this function skipped all sampass which shouldn't be the case
  return passed

# get word with aek or too
def get_aek_too(data:list, ktype='aek'):
  passed = []
  for possible_word in tqdm(data):
      if kv.check_aek_too(possible_word) == ktype:
        passed.append(possible_word)
  return passed

def tone_gen(klong_text, gened_word, word_mark='no', sampas=False):
  splitted_klong = split_klong(klong_text)
  if word_mark == 'no' and sampas == False:
     probs = generator(klong_text)
     for prob in probs:
       if prob not in gened_word:
         gened_word.append(prob)
         return prob, gened_word
  elif word_mark == 'aek' and sampas == False:
    probs = generator(klong_text)
    aek = get_aek_too(probs)
    for prob in aek:
      if prob not in gened_word:
        gened_word.append(prob)
        return prob, gened_word
  elif word_mark == 'too' and sampas == False:
    probs = generator(klong_text)
    too = get_aek_too(probs, 'too')
    for prob in too:
      if prob not in gened_word:
        gened_word.append(prob)
        return prob, gened_word
  elif sampas == True and word_mark == 'no':
    probs = gen_prob_next_token(klong_text, model, tokenizer)
    probs = probs['token'][:500]
    passed = get_sampassed(probs, sound_words(splitted_klong)[1])
    for prob in passed:
       if prob not in gened_word:
         gened_word.append(prob)
         return prob, gened_word
  elif sampas == True and word_mark == 'too':
    probs = gen_prob_next_token(klong_text, model, tokenizer)
    probs = probs['token'][:500]
    passed = get_sampassed(probs, sound_words(splitted_klong)[3])
    for prob in passed:
       if prob not in gened_word and kv.check_aek_too(prob) == 'too':
         gened_word.append(prob)
         return prob, gened_word
     
def gen_klong(klong_text_input, gened_word):
  splitted_klong = split_klong(klong_text_input)
  klong_text = klong_text_input
  # วรรค 2, 4, 6
  if len(splitted_klong) in [1, 3, 5]:
    word_gen = 2
    if len(splitted_klong) == 1:
      # ฉันทลักษณ์ (none, none(sampas))
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      klong_text = klong_text + '-'
    elif len(splitted_klong) == 3:
      # ฉันทลักษณ์ (aek, too(sampas))
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='aek')
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word, 'too')
      klong_text = klong_text + prob
      klong_text = klong_text + '-'
    elif len(splitted_klong) == 5:
      # ฉันทลักษณ์ (none, aek)
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='aek')
      klong_text = klong_text + prob
      klong_text = klong_text + '-'

  # วรรค 3, 5, 7
  elif len(splitted_klong) in [2, 4, 6]:
    word_gen = 5
    if len(splitted_klong) == 2:
      # ฉันทลักษณ์ (none, aek, none, none, none(sampas))
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='aek')
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      sampas_word = sound_words(splitted_klong)[1]
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='no', sampas=True)
      klong_text = klong_text + prob
      klong_text = klong_text + '-'
    elif len(splitted_klong) == 4:
      # ฉันทลักษณ์ (none, none, aek, none, none(sampas))
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='aek')
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      sampas_word = sound_words(splitted_klong)[1]
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='no', sampas=True)
      klong_text = klong_text + prob
      klong_text = klong_text + '-'
    elif len(splitted_klong) == 6:
      # ฉันทลักษณ์ (none, aek, none, none, too(sampas))
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='aek')
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      prob, gened_word  = tone_gen(klong_text, gened_word)
      klong_text = klong_text + prob
      sampas_word = sound_words(splitted_klong)[1]
      prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='too', sampas=True)
      klong_text = klong_text + prob
      klong_text = klong_text + '-'
  # วรรค 8
  elif len(splitted_klong) == 7:
    # ฉันทลักษณ์ (eak, too, none, none)
    word_gen = 4
    prob, gened_word  = tone_gen(klong_text, gened_word, word_mark='aek')
    klong_text = klong_text + prob
    prob, gened_word  = tone_gen(klong_text, gened_word, 'too')
    klong_text = klong_text + prob
    prob, gened_word  = tone_gen(klong_text, gened_word)
    klong_text = klong_text + prob
    prob, gened_word  = tone_gen(klong_text, gened_word)
    klong_text = klong_text + prob
    klong_text = klong_text + '\n'
  return klong_text, gened_word

# main
def main(klong_text):
  gened_klong = []
  splitted = split_klong(klong_text)
  if main_check(klong_text) == True:
    wak_num = len(splitted)
    klong_text, gened_klong = gen_klong(klong_text, gened_klong)
    return klong_text
  else:
    return main_check(klong_text)

import gradio as gr

iface = gr.Interface(fn=main, inputs="text", outputs="text")
iface.launch(share=True, debug=True)