def train(train_model,
prediction_model,
model_name='baseline',
epochs=2,
batch_size=32,
validation_split=0.1):
# load training data
qa_data = QAData()
questions, good_answers, bad_answers = qa_data.get_training_data()
logger.info(
f'Training: epochs {epochs}, batch_size {batch_size}, validation_split {validation_split}'
)
# train the model
Y = np.zeros(shape=(questions.shape[0], ))
early_stopping = EarlyStopping(monitor='val_loss',
mode='min',
min_delta=0.01,
verbose=1,
patience=50,
restore_best_weights=True)
hist = train_model.fit([questions, good_answers, bad_answers],
Y,
epochs=epochs,
batch_size=batch_size,
validation_split=validation_split,
verbose=1,
callbacks=[early_stopping])
# save plot val_loss, loss
df = pd.DataFrame(hist.history)
df.insert(0, 'epochs', range(0, len(df)))
df = pd.melt(df, id_vars=['epochs'])
plot = ggplot(aes(x='epochs', y='value', color='variable'),
data=df) + geom_line()
filename = f'{model_name}_plot.png'
logger.info(f'saving loss, val_loss plot: {filename}')
plot.save(filename)
#save_model_architecture(prediction_model, model_name=model_name)
save_model_weights(train_model, model_name=model_name)
clear_session()
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 model
vocabulary = Vocabulary("./data/vocab_all.txt")
embedding_file = "./data/word2vec_100_dim.embeddings"
qa_model = QAModel()
train_model, predict_model = qa_model.get_bilstm_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/lstm_predict_weights_epoch_1.h5')
c = 0
c1 = 0
for i, d in enumerate(data):
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("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
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)