compound.py

# -*- coding: utf-8 -*-
#!pip install wandb
#import wandb
#wandb.login()

from keras.models import Model
from keras.layers import Input, LSTM, Dense, Bidirectional, Concatenate
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.model_selection import KFold
from keras.layers import Layer
import keras.backend as K
from wandb.keras import WandbCallback
import pickle
from keras.models import load_model
from keras.callbacks import EarlyStopping

"""# Requirements"""

# returns lists ol(kridant), w1l(dhatu), w2l(pratya)
def get_data(path):
  with open(path) as f:
    lines=f.read().splitlines()
    ol=[]
    w1l=[]
    w2l=[]
    for line in lines:
        words=line.split(" ")
        kridant=words[0]
        dhatu=words[2]
        pratya=words[4]
        ol.append(kridant)
        w1l.append(dhatu)
        w2l.append(pratya)

  return w1l,w2l,ol

def get_texts(file_path,test_split=0.2,random_state=1,take_last_two=False):
  input_texts = []
  target_texts = []
  w1l,w2l,ol=get_data(file_path)

  print("Sandhi dataset created")
  ct=0
  for i in range(len(w1l)):
    #this condition avoids all the Satf~ and SAnac pratya
    if(take_last_two==False and w2l[i]=="Satf~" or w2l[i]=="SAnac"):
      ct+=1
      continue
    input_text = w1l[i] + '+' + w2l[i]
    target_text = ol[i]
    # We use "&" as the "start sequence" character for the targets, and "$" as "end sequence" character.
    target_text = '&' + target_text + '$'
    input_texts.append(input_text)
    target_texts.append(target_text)
  print("Total number of last two pratyas:",ct)
  if(test_split!=0):
    X_train, X_test, Y_train, Y_test = train_test_split(input_texts, target_texts, test_size=test_split, random_state=random_state)
    return X_train,X_test,Y_train,Y_test
  return input_texts,target_texts

def get_d(X,tokens):
  for sentence in X:
    for char in sentence:
      tokens.add(char)
  return tokens

def get_trained_model(architecture,X_train,Y_train,X_test=None,Y_test=None,latent_dim=32,batch_size=64,epochs=70,validation_split=0.2,verbose=1,use_wandb=False,model=None):
  re=True
  if(model==None):
    re=False
    model=Translator()
    model.latent_dim=latent_dim
    # getting model dictionary
    input_texts=[x for x in X_train]
    target_texts=[y for y in Y_train]
    if(X_test!=None and Y_test!=None):
      input_texts.extend(X_test)
      target_texts.extend(Y_test)
    characters=set()
    characters=get_d(input_texts,characters)
    characters=get_d(target_texts,characters) 

    model.max_encoder_seq_length = max([len(txt) for txt in input_texts])
    model.max_decoder_seq_length = max([len(txt) for txt in target_texts])

    # Using '*' for padding 
    characters.add('*')

    characters = sorted(list(characters))
    model.num_tokens = len(characters)

    print('Number of samples:', len(input_texts))
    print('Number of unique tokens:', model.num_tokens)
    print('Max sequence length for inputs:', model.max_encoder_seq_length)
    print('Max sequence length for outputs:', model.max_decoder_seq_length)

    model.token_index = dict([(char, i) for i, char in enumerate(characters)])
    model.reverse_target_char_index = dict((i, char) for char, i in model.token_index.items())


  encoder_input_data = np.zeros((len(X_train), model.max_encoder_seq_length,model.num_tokens), dtype='float32')
  decoder_input_data = np.zeros((len(X_train), model.max_decoder_seq_length, model.num_tokens), dtype='float32')
  decoder_target_data = np.zeros((len(X_train), model.max_decoder_seq_length, model.num_tokens), dtype='float32')

  for i, (input_text, target_text) in enumerate(zip(X_train, Y_train)):
    for t, char in enumerate(input_text):
      encoder_input_data[i, t, model.token_index[char]] = 1.
    encoder_input_data[i, t + 1:, model.token_index['*']] = 1.
    for t, char in enumerate(target_text):
      # decoder_target_data is ahead of decoder_input_data by one timestep
      decoder_input_data[i, t, model.token_index[char]] = 1.
      if t > 0:
        # decoder_target_data will be ahead by one timestep
        # and will not include the start character.
        decoder_target_data[i, t - 1, model.token_index[char]] = 1.
    decoder_input_data[i, t + 1:, model.token_index['*']] = 1.
    decoder_target_data[i, t:, model.token_index['*']] = 1.
  if(re==True):
    model.model=train(model.model,encoder_input_data,decoder_input_data,decoder_target_data,batch_size,epochs,validation_split,verbose,use_wandb,re=re)
    return model
  # train the model on text data
  model.model,model.encoder_model,model.decoder_model=architecture(model.latent_dim,model.num_tokens)
  model.model=train(model.model,encoder_input_data,decoder_input_data,decoder_target_data,batch_size,epochs,validation_split,verbose,use_wandb,re=re)
  
  #save the best model and last model before returning
  return model

# This function is just for logging results.
"""def use_wandb(project_name,run_name,batch_size,epochs,validation_split,latent_dim):
    wandb.init(project=project_name,name=run_name)
    config=wandb.config
    config.epochs=epochs
    config.batch_size=batch_size
    config.validation_split=validation_split
    config.latent_dim=latent_dim
"""  

def train(model,encoder_input_data,decoder_input_data,decoder_target_data,batch_size,epochs,validation_split,verbose,use_wandb=False,re=False):
  if(re==False):
    model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
  if(use_wandb==True):
    #print("Running:",run_name)
    es = EarlyStopping(monitor='val_accuracy', mode='max', verbose=1,patience=4)
    model.fit([encoder_input_data, decoder_input_data], decoder_target_data,batch_size=batch_size,epochs=epochs,validation_split=validation_split,verbose=verbose,callbacks=[WandbCallback(),es])
  else:
    es = EarlyStopping(monitor='val_accuracy', mode='max', verbose=1,patience=4)
    model.fit([encoder_input_data, decoder_input_data], decoder_target_data,batch_size=batch_size,epochs=epochs,validation_split=validation_split,verbose=verbose,callbacks=[es])
  return model


def normal_testing(X,Y,model,verbose=1,use_wandb=False,log_as="test_passed"):
  #file1 = open("Incorrect_predictions.txt", "w")    
  predictions=model.get_predictions(X)
  total=len(predictions)
  passed=0
  true=[a[1:-1] for a in Y]
  #passed=np.sum([1 for i in range(total) if(predictions[i]==true[i])])
  for i in range(total):
    if(predictions[i]==true[i]):
      passed+=1
    else:
      if(verbose==1):
        print(str(i)+'/'+str(total))
        print('-')
        print('Input sentence:   ', X[i])
        print('Decoded sentence: ', predictions[i])
        print('Expected sentence:', true[i])
  if(use_wandb==True):
    #wandb.log({log_as:str(passed)+"/"+str(total)})
    wandb.log({log_as:passed/total})
  print(log_as+":"+str(passed)+'/'+str(total))
  return passed,total


def K_Fold_testing(X,Y,model_architecture,k,latent_dim=32,batch_size=64,epochs=70,validation_split=0,verbose=0,use_wandb=False,run_name="None",log_as="test_passed"):
  acc=[]
  train_acc=[]
  if(use_wandb):
    wandb.config.num_k=k   
  kf=KFold(n_splits=k,random_state=1,shuffle=True)
  kf.get_n_splits(X,Y)
  i=1
  for train_index,test_index in kf.split(X):
    x_train=list(np.array(X)[train_index.astype(int)])
    x_test=list(np.array(X)[test_index.astype(int)])
    y_train=list(np.array(Y)[train_index.astype(int)])
    y_test=list(np.array(Y)[test_index.astype(int)])
    print("split_number:",i)
    model=get_trained_model(architecture=model_architecture,X_train=x_train,Y_train=y_train,X_test=x_test,Y_test=y_test,latent_dim=latent_dim,batch_size=batch_size,epochs=epochs,validation_split=validation_split,verbose=verbose,use_wandb=use_wandb)
    train_passed,train_out_of=normal_testing(x_train,y_train,model,verbose=0,use_wandb=use_wandb,log_as="train_passed")
    passed,outof=normal_testing(x_test,y_test,model,verbose=0,use_wandb=use_wandb,log_as="test_passed")
    train_acc.append(train_passed/train_out_of)
    acc.append(passed/outof)
    print("------------")
    i+=1
  kf_acc=np.sum(acc)/k
  kf_train_acc=np.sum(train_acc)/k
  if(use_wandb):
    wandb.log({"KFold_accuracy":kf_acc})
    wandb.log({"KFold_train_accuracy":kf_train_acc})
  print("KFold_train_acc=",kf_train_acc)
  print("KFold_test_acc=",kf_acc)

class Translator:
  def init(self):
    self.token_index=None
    self.num_tokens=0
    self.max_encoder_seq_length=0
    self.max_decoder_seq_length=0
    self.model=None
    self.encoder_model=None
    self.decoder_model=None
    self.reverse_target_char_index=None

  def dictionary_info(self):
    
    print("num_tokens",self.num_tokens)
    print("max_encoder_seq_length",self.max_encoder_seq_length)
    print("max_decoder_seq_length",self.max_decoder_seq_length)
    print("token_index")
    print(self.token_index)
    print("reverse_target_char_index")
    print(self.reverse_target_char_index)

  def vectorize(self,X):
    encoder_input_data = np.zeros((len(X), self.max_encoder_seq_length, self.num_tokens), dtype='float32')
    for i, input_text in enumerate(X):
      for t, char in enumerate(input_text):
        if char not in self.token_index:
          continue
        encoder_input_data[i, t, self.token_index[char]] = 1.
      encoder_input_data[i, t + 1:, self.token_index['*']] = 1.
    return encoder_input_data

  def get_predictions(self,X):
    ans=[]   
    encoder_input_data=self.vectorize(X)
    for seq_index in range(len(encoder_input_data)):
      input_seq = encoder_input_data[seq_index: seq_index + 1]
      decoded_sentence = self.decode_sequence(input_seq)
      #print(decoded_sentence,input_seq)
      decoded_sentence = decoded_sentence.strip()
      decoded_sentence = decoded_sentence.strip('$')
      ans.append(decoded_sentence)
    return ans

  def decode_sequence(self,input_seq):
    # Encode the input as state vectors.
    states_value = self.encoder_model.predict(input_seq)

    # Generate empty target sequence of length 1.
    target_seq = np.zeros((1, 1, self.num_tokens))
    # Populate the first character of target sequence with the start character.
    target_seq[0, 0, self.token_index['&']] = 1.

    # Sampling loop for a batch of sequences
    # (to simplify, here we assume a batch of size 1).
    stop_condition = False
    decoded_sentence = ''
    while not stop_condition:
      output_tokens, h, c = self.decoder_model.predict([target_seq] + states_value)
      # Sample a token
      sampled_token_index = np.argmax(output_tokens[0, -1, :])
      sampled_char = self.reverse_target_char_index[sampled_token_index]
      decoded_sentence += sampled_char

        # Exit condition: either hit max length
        # or find stop character.
      if (sampled_char == '$' or
        len(decoded_sentence) > self.max_decoder_seq_length):
          stop_condition = True

        # Update the target sequence (of length 1).
      target_seq = np.zeros((1, 1, self.num_tokens))
      target_seq[0, 0, sampled_token_index] = 1.

          # Update states
      states_value = [h, c]

    return decoded_sentence

# use maximum seq length for input before using i.e 18 for taddhita and 17 for kridant in our dataset. You can explicilty check the length for your dataset. It is however not needed for a model that does not use attention.
class attention(Layer):
    def __init__(self,**kwargs):
        super(attention,self).__init__(**kwargs)

    def build(self,input_shape):
        self.W=self.add_weight(name="att_weight",shape=(input_shape[-1],1),initializer="normal")
        self.b=self.add_weight(name="att_bias",shape=(18,1),initializer="zeros")                # Change the sequence length before using. It depends on the dataset and for us it was (16,1) for splitting and (17,1) for synthesis 
        super(attention, self).build(input_shape)

    def call(self,x):
        et=K.squeeze(K.tanh(K.dot(x,self.W)+self.b),axis=-1)
        #print(np.shape(et))
        #print("....shape et")
        at=K.softmax(et)
        at=K.expand_dims(at,axis=-1)
        output=x*at
        return K.sum(output,axis=1)

    def compute_output_shape(self,input_shape):
        return (input_shape[0],input_shape[-1])

    def get_config(self):
        return super(attention,self).get_config()


def get_model_attention(latent_dim,num_tokens):
  
  encoder_inputs = Input(shape=(None,num_tokens))
  encoder = Bidirectional(LSTM(latent_dim,return_sequences=True,return_state=True,recurrent_dropout=0.2))
  att_in,forward_h, forward_c, backward_h, backward_c=encoder(encoder_inputs)
  att_out=attention()(att_in)
    
  state_c = Concatenate()([forward_c, backward_c])
    
  encoder_states=[att_out,state_c]
    
  # Set up the decoder, using `encoder_states` as initial state.
  decoder_inputs = Input(shape=(None, num_tokens))
  # We set up our decoder to return full output sequences,
  # and to return internal states as well. We don't use the
  # return states in the training model, but we will use them in inference.
  decoder_lstm = LSTM(latent_dim*2, return_sequences=True, return_state=True,recurrent_dropout=0.2)
  decoder_outputs, _, _ = decoder_lstm(decoder_inputs, initial_state=encoder_states)
  decoder_dense = Dense(num_tokens, activation='softmax')
  decoder_outputs = decoder_dense(decoder_outputs)

  # Define the model that will turn
  # `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
  model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
      
  encoder_model = Model(encoder_inputs, encoder_states)
    
  decoder_state_input_h = Input(shape=(latent_dim*2,))
  decoder_state_input_c = Input(shape=(latent_dim*2,))
  decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
  decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs, initial_state=decoder_states_inputs)
  decoder_states = [state_h, state_c]
  decoder_outputs = decoder_dense(decoder_outputs)
  decoder_model = Model([decoder_inputs] + decoder_states_inputs, [decoder_outputs] + decoder_states)

  return model,encoder_model,decoder_model


def get_model(latent_dim,num_tokens):
  
  encoder_inputs = Input(shape=(None, num_tokens))
  encoder = Bidirectional(LSTM(latent_dim, return_state=True,recurrent_dropout=0.2))
  encoder_outputs, forward_h, forward_c, backward_h, backward_c = encoder(encoder_inputs)
  state_h = Concatenate()([forward_h, backward_h])
  state_c = Concatenate()([forward_c, backward_c])

  # We discard `encoder_outputs` and only keep the states.
  encoder_states = [state_h, state_c]

  # Set up the decoder, using `encoder_states` as initial state.
  decoder_inputs = Input(shape=(None, num_tokens))

  # We set up our decoder to return full output sequences,
  # and to return internal states as well. We don't use the
  # return states in the training model, but we will use them in inference.
  decoder_lstm = LSTM(latent_dim*2, return_sequences=True, return_state=True,recurrent_dropout=0.2)
  decoder_outputs, _, _ = decoder_lstm(decoder_inputs, initial_state=encoder_states)
  decoder_dense = Dense(num_tokens, activation='softmax')
  decoder_outputs = decoder_dense(decoder_outputs)

  # Define the model that will turn
  # `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
  model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
      
  encoder_model = Model(encoder_inputs, encoder_states)
    
  decoder_state_input_h = Input(shape=(latent_dim*2,))
  decoder_state_input_c = Input(shape=(latent_dim*2,))
  decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
  decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs, initial_state=decoder_states_inputs)
  decoder_states = [state_h, state_c]
  decoder_outputs = decoder_dense(decoder_outputs)
  decoder_model = Model([decoder_inputs] + decoder_states_inputs, [decoder_outputs] + decoder_states)

  return model,encoder_model,decoder_model

def save_model(filename,model_obj):
  mod=model_obj
  fileobj=open(filename,'wb')
  pickle.dump(mod,fileobj)
    
def restore_model(filename):                                                
  return pickle.load(open(filename,'rb'))

"""# Training"""

X_train, X_test, Y_train, Y_test=get_texts("Kosh_data/Taddhita.txt",random_state=1)

# can also use get_model instead of get_attention model to run the model without attention
translator=get_trained_model(get_model_attention,X_train,Y_train,X_test,Y_test,epochs=70,validation_split=0.1,batch_size=32,latent_dim=64,use_wandb=False)

normal_testing(X_test,Y_test,model=translator,verbose=0,use_wandb=False) #gives test accuracy

normal_testing(X_train,Y_train,model=translator,verbose=0,use_wandb=False,log_as="train_passed")  # gives train accuracy



"""## Narrowing epoch range"""

#translator=get_trained_model(get_model,X_train,Y_train,X_test,Y_test,epochs=10,verbose=2,batch_size=32,latent_dim=64)
#normal_testing(X_test,Y_test,model=translator,verbose=0)
#for i in range(9):
#  print("Increasing num of epochs by 10: total epochs=",10*(i+2))
#  translator=get_trained_model(get_model_attention,X_train,Y_train,X_test,Y_test,epochs=10,model=translator,verbose=2,batch_size=32,latent_dim=64)
#  a,b=normal_testing(X_test,Y_test,model=translator,verbose=0)
#  print("total_epochs=",10*(i+2),"test_acc=",(a*100)/b)

#normal_testing(X_test,Y_test,translator)

#translator=get_trained_model(get_model_attention,X_train,Y_train,X_test,Y_test,epochs=5,model=translator)

"""## K-Fold"""

#use_wandb("kridanta_synthesis","Swapped_parameters_Regularized_KFold_with_attention_1",batch_size=32,epochs=40,validation_split=0.1,latent_dim=64)
#X,Y=get_texts("/content/log.txt",test_split=0)
#K_Fold_testing(X,Y,get_model_attention,5,verbose=1,validation_split=0.1,use_wandb=True,latent_dim=64,batch_size=32,epochs=40)