def get_baseline_model():
# get the train and predict model model
embedding_file, vocab_size = get_default_inputs_for_model()
qa_model = QAModel()
train_model, prediction_model = qa_model.get_lstm_cnn_model(
embedding_file, vocab_size)
logger.info('Default created: Baseline')
logger.info('enc_timesteps = 30,\
dec_timesteps = 30, hidden_dim = 50, filters = 500, kernel_sizes = [2, 3, 5, 7]'
)
return train_model, prediction_model
def get_small_model():
# small model
embedding_file, vocab_size = get_default_inputs_for_model()
enc_timesteps = 30
dec_timesteps = 30
hidden_dim = 10
filters = 20
qa_model = QAModel()
small_train_model, small_prediction_model = qa_model.get_lstm_cnn_model(
embedding_file,
vocab_size,
enc_timesteps=enc_timesteps,
dec_timesteps=dec_timesteps,
filters=filters,
hidden_dim=hidden_dim)
logger.info('Model created: Small')
logger.info(f'enc_timesteps = {enc_timesteps},\
dec_timesteps = {dec_timesteps},'
f' hidden_dim = {hidden_dim}, filters = {filters}, '
f'kernel_sizes = [2, 3, 5, 7]')
return small_train_model, small_prediction_model
def get_larger_model():
enc_timesteps = 30
dec_timesteps = 30
hidden_dim = 200
filters = 500
embedding_file, vocab_size = get_default_inputs_for_model()
qa_model = QAModel()
larger_train_model, larger_prediction_model = qa_model.get_lstm_cnn_model(
embedding_file,
vocab_size,
enc_timesteps=enc_timesteps,
dec_timesteps=dec_timesteps,
filters=filters,
hidden_dim=hidden_dim)
logger.info('Model created: Larger')
logger.info(f'enc_timesteps = {enc_timesteps},\
dec_timesteps = {dec_timesteps},'
f' hidden_dim = {hidden_dim}, filters = {filters}, '
f'kernel_sizes = [2, 3, 5, 7]')
return larger_train_model, larger_prediction_model
def main(mode='test', question=None, answers=None):
"""
This function is used to train, predict or test
Args:
mode (str): train/preddict/test
question (str): this contains the question
answers (list): this contains list of answers in string format
Returns:
index (integer): index of the most likely answer
"""
# get the train and predict model
vocabulary = Vocabulary("./data/vocab_all.txt")
embedding_file = "./data/word2vec_100_dim.embeddings"
qa_model = QAModel()
train_model, predict_model = qa_model.get_lstm_cnn_model(
embedding_file, len(vocabulary))
epoch = 1
if mode == 'train':
for i in range(epoch):
print('Training epoch', i)
# load training data
qa_data = QAData()
questions, good_answers, bad_answers = qa_data.get_training_data()
# train the model
Y = np.zeros(shape=(questions.shape[0], ))
train_model.fit([questions, good_answers, bad_answers],
Y,
epochs=1,
batch_size=64,
validation_split=0.1,
verbose=1)
# save the trained model
train_model.save_weights('model/train_weights_epoch_' +
str(epoch) + '.h5',
overwrite=True)
predict_model.save_weights('model/predict_weights_epoch_' +
str(epoch) + '.h5',
overwrite=True)
elif mode == 'predict':
# load the evaluation data
data = pickle.load(open("./data/dev.pkl", 'rb'))
random.shuffle(data)
# load weights from trained model
qa_data = QAData()
predict_model.load_weights('model/cnnlstm_predict_weights_epoch_1.h5')
c = 0
c1 = 0
total = 0
for i, d in enumerate(data):
print(i, len(d['question']))
#if(len(d['question']) > 7):
# pad the data and get it in desired format
indices, answers, question = qa_data.process_data(d)
total += 1
# get the similarity score
sims = predict_model.predict([question, answers])
n_good = len(d['good'])
max_r = np.argmax(sims)
max_n = np.argmax(sims[:n_good])
r = rankdata(sims, method='max')
c += 1 if max_r == max_n else 0
c1 += 1 / float(r[max_r] - r[max_n] + 1)
precision = c / float(total) #len(data))
mrr = c1 / float(total) #len(data))
print("Precision", precision)
print("MRR", mrr)
#break
precision = c / float(total) #len(data))
mrr = c1 / float(total) #len(data))
print("Precision", precision)
print("MRR", mrr)
elif mode == 'test':
# question and answers come from params
qa_data = QAData()
answers, question = qa_data.process_test_data(question, answers)
# load weights from the trained model
predict_model.load_weights('model/lstm_predict_weights_epoch_1.h5')
# get similarity score
sims = predict_model.predict([question, answers])
max_r = np.argmax(sims)
return max_r
@author: tarun
"""
from data import QAData, Vocabulary
from model import QAModel
import pickle
import numpy as np
import random
from keras.models import Model
import matplotlib.pyplot as plt
vocabulary = Vocabulary("./data/vocab_all.txt")
embedding_file = "./data/word2vec_100_dim.embeddings"
qa_model = QAModel()
train_model, predict_model = qa_model.get_lstm_cnn_model(embedding_file, len(vocabulary))
# layer_outputs = [predict_model.layers[0].output, predict_model.layers[1].output, predict_model.layers[2].layers[0].output, predict_model.layers[2].layers[0].output,
# predict_model.layers[2].layers[1].output, predict_model.layers[2].layers[2].get_output_at(0),
# predict_model.layers[2].layers[2].get_output_at(1), predict_model.layers[2].layers[3].get_output_at(0),
# predict_model.layers[2].layers[3].get_output_at(1), predict_model.layers[2].layers[4].get_output_at(0),
# predict_model.layers[2].layers[4].get_output_at(1), predict_model.layers[2].layers[5].output,
# predict_model.layers[2].layers[6].output, predict_model.layers[2].layers[7].get_output_at(0),
# predict_model.layers[2].layers[7].get_output_at(1), predict_model.layers[2].layers[8].get_output_at(0),
# predict_model.layers[2].layers[8].get_output_at(1), predict_model.layers[2].layers[9].get_output_at(0),
# predict_model.layers[2].layers[9].get_output_at(1), predict_model.layers[2].layers[10].get_output_at(0),
# predict_model.layers[2].layers[10].get_output_at(1), predict_model.layers[2].layers[11].output,
# predict_model.layers[2].layers[11].output, predict_model.layers[2].layers[13].get_output_at(0),
# predict_model.layers[2].layers[13].get_output_at(1), predict_model.layers[2].layers[14].output]
layer_outputs = [predict_model.layers[2].layers[13].get_output_at(0)]#, predict_model.layers[2].layers[13].get_output_at(1), predict_model.layers[2].layers[14].output]
def main(mode='test'):
# get the train and predict model model
vocabulary = Vocabulary("./data/vocab_all.txt")
embedding_file = "./data/word2vec_100_dim.embeddings"
qa_model = QAModel()
train_model, predict_model = qa_model.get_lstm_cnn_model(embedding_file, len(vocabulary))
epo = 100
if mode == 'train':
# load training data
qa_data = QAData()
questions, good_answers, bad_answers = qa_data.get_training_data()
callbacks = [EarlyStopping(monitor='val_loss', patience=20),
ModelCheckpoint(filepath='best_model.h5', monitor='val_loss', save_best_only=True)]
# train the model
Y = np.zeros(shape=(questions.shape[0],))
train_model.fit([questions, good_answers, bad_answers], Y, epochs=epo, batch_size=64, validation_split=0.1,
verbose=1, callbacks=callbacks)
# save the trained model
# train_model.save_weights('model/train_weights_epoch_' + str(epo) + '.h5', overwrite=True)
model = keras.models.load_model('best_model.h5')
model.save_weights('model/best_weights_epoch_' + str(epo) + '.h5', overwrite=True)
predict_model.save_weights('model/predict_weights_epoch_' + str(epo) + '.h5', overwrite=True)
elif mode == 'predict':
# load the evaluation data
data = pickle.load(open("./data/dev.pkl",'rb'))
random.shuffle(data)
# load weights from trained model
qa_data = QAData()
model_filenames = ['model/best_model.h5', 'model/predict_weights_epoch_' + str(epo) + '.h5']
for model_name in model_filenames:
predict_model.load_weights(model_name)
c = 0
c1 = 0
for i, d in enumerate(data):
if i%100 == 0:
print(i, len(data))
# pad the data and get it in desired format
indices, answers, question = qa_data.process_data(d)
# get the similarity score
sims = predict_model.predict([question, answers])
n_good = len(d['good'])
max_r = np.argmax(sims)
max_n = np.argmax(sims[:n_good])
r = rankdata(sims, method='max')
c += 1 if max_r == max_n else 0
c1 += 1 / float(r[max_r] - r[max_n] + 1)
precision = c / float(len(data))
mrr = c1 / float(len(data))
print(f'Results for: model: {model_name}')
print("Precision", precision)
print("MRR", mrr)
