97 lines
3.4 KiB
Python
97 lines
3.4 KiB
Python
|
import numpy as np
|
||
|
|
from keras_preprocessing.sequence import pad_sequences
|
||
|
|
from keras_preprocessing.text import Tokenizer
|
||
|
|
from keras.models import Sequential
|
||
|
|
from keras.layers import Dense, LSTM, Embedding, Dropout
|
||
|
|
from keras.callbacks import ModelCheckpoint
|
||
|
|
|
||
|
|
|
||
|
|
def recreate_model(predictors, labels, model, filepath, epoch_num):
|
||
|
|
model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
|
||
|
|
append_epochs(predictors, labels, model, epoch_num)
|
||
|
|
|
||
|
|
|
||
|
|
def append_epochs(predictors, labels, model, filepath, epoch_num):
|
||
|
|
checkpoint = ModelCheckpoint(filepath, monitor='loss', verbose=1, save_best_only=True, mode='min')
|
||
|
|
desired_callbacks = [checkpoint]
|
||
|
|
model.fit(predictors, labels, epochs=epoch_num, verbose=1, callbacks=desired_callbacks)
|
||
|
|
|
||
|
|
|
||
|
|
def generate_text(tokenizer, seed_text, next_words, model, max_seq_length):
|
||
|
|
for _ in range(next_words):
|
||
|
|
token_list = tokenizer.texts_to_sequences([seed_text])[0]
|
||
|
|
token_list = pad_sequences([token_list], maxlen=max_seq_length - 1, padding='pre')
|
||
|
|
predicted = np.argmax(model.predict(token_list), axis=-1)
|
||
|
|
output_word = ""
|
||
|
|
for word, index in tokenizer.word_index.items():
|
||
|
|
if index == predicted:
|
||
|
|
output_word = word
|
||
|
|
break
|
||
|
|
seed_text += " " + output_word
|
||
|
|
return seed_text
|
||
|
|
|
||
|
|
|
||
|
|
def start():
|
||
|
|
flag = -1
|
||
|
|
while flag < 1 or flag > 2:
|
||
|
|
flag = int(input("Select model and text (1 - eng, 2 - ru): "))
|
||
|
|
|
||
|
|
if flag == 1:
|
||
|
|
file = open("data.txt").read()
|
||
|
|
filepath = "model_eng.hdf5"
|
||
|
|
elif flag == 2:
|
||
|
|
file = open("rus_data.txt").read()
|
||
|
|
filepath = "model_rus.hdf5"
|
||
|
|
else:
|
||
|
|
exit(1)
|
||
|
|
|
||
|
|
tokenizer = Tokenizer()
|
||
|
|
tokenizer.fit_on_texts([file])
|
||
|
|
words_count = len(tokenizer.word_index) + 1
|
||
|
|
|
||
|
|
input_sequences = []
|
||
|
|
for line in file.split('\n'):
|
||
|
|
token_list = tokenizer.texts_to_sequences([line])[0]
|
||
|
|
for i in range(1, len(token_list)):
|
||
|
|
n_gram_sequence = token_list[:i + 1]
|
||
|
|
input_sequences.append(n_gram_sequence)
|
||
|
|
|
||
|
|
max_seq_length = max([len(x) for x in input_sequences])
|
||
|
|
input_sequences = pad_sequences(input_sequences, maxlen=max_seq_length, padding='pre')
|
||
|
|
|
||
|
|
predictors, labels = input_sequences[:, :-1], input_sequences[:, -1]
|
||
|
|
|
||
|
|
model = Sequential()
|
||
|
|
model.add(Embedding(words_count, 100, input_length=max_seq_length - 1))
|
||
|
|
model.add(LSTM(150))
|
||
|
|
model.add(Dropout(0.15))
|
||
|
|
model.add(Dense(words_count, activation='softmax'))
|
||
|
|
|
||
|
|
flag = input("Do you want to recreate the model ? (print yes): ")
|
||
|
|
if flag == 'yes':
|
||
|
|
flag = input("Are you sure? (print yes): ")
|
||
|
|
if flag == 'yes':
|
||
|
|
num = int(input("Select number of epoch: "))
|
||
|
|
if 0 < num < 100:
|
||
|
|
recreate_model(predictors, labels, model, filepath, num)
|
||
|
|
|
||
|
|
model.load_weights(filepath)
|
||
|
|
model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
|
||
|
|
|
||
|
|
flag = input("Do you want to train the model ? (print yes): ")
|
||
|
|
if flag == 'yes':
|
||
|
|
flag = input("Are you sure? (print yes): ")
|
||
|
|
if flag == 'yes':
|
||
|
|
num = int(input("Select number of epoch: "))
|
||
|
|
if 0 < num < 100:
|
||
|
|
append_epochs(predictors, labels, model, filepath, num)
|
||
|
|
|
||
|
|
flag = 'y'
|
||
|
|
while flag == 'y':
|
||
|
|
seed = input("Enter seed: ")
|
||
|
|
print(generate_text(tokenizer, seed, 25, model, max_seq_length))
|
||
|
|
flag = input("Continue? (print \'y\'): ")
|
||
|
|
|
||
|
|
|
||
|
|
start()
|