validation_data=(x_val, y_val),
batch_size=512,
epochs=1000,
shuffle=True,
callbacks=[es, mc])
print('trained embedding shape:', model.layers[0].get_weights()[0].shape)
utils.save_embs_2_file(model, 0, tokenizer.word_index)
# test_loss, test_acc = model.evaluate(x_test, y_test)
# print('test loss:', test_loss, 'test acc:', test_acc)
gold_en = y_test
predicted_en = model.predict(x_test).argmax(axis=1)
gold_de = y_test_de
predicted_de = model.predict(x_test_de).argmax(axis=1)
utils.test_evaluation(gold_en, predicted_en, gold_de, predicted_de)
# de fine-tuning
FINETUNE = False
if FINETUNE:
print('performing classical fine-tuning...')
print('train:', de_train_dir)
print('dev:', de_dev_dir)
model2 = models.load_model('best_model.h5', compile=False)
# model2.layers[0].trainable = True
Adam = optimizers.Adam(learning_rate=0.0005)
model2.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['acc'])
print(model2.summary())
print(K.eval(model2.optimizer.lr))
save_best_only=True)
model2.fit(x_train,
y_train,
validation_data=(x_val, y_val),
batch_size=64,
epochs=100,
shuffle=True,
callbacks=[es, mc])
print('trained embedding shape:', model.layers[0].get_weights()[0].shape)
gold_en = y_test
predicted_en = model.predict(x_test).argmax(axis=1)
gold_de = y_test_de
predicted_de = model.predict(x_test_de).argmax(axis=1)
en_mic, de_mic, en_mac, de_mac = utils.test_evaluation(
gold_en, predicted_en, gold_de, predicted_de)
global_en_mic_train += en_mic
global_de_mic_train += de_mic
global_en_mac_train += en_mac
global_de_mac_train += de_mac
# de fine-tuning
FINETUNE = True
if FINETUNE:
print('performing classical fine-tuning...')
print('train:', de_train_dir)
print('dev:', de_dev_dir)
print('fine-tuning architecture...')
model3 = models.load_model('best_model.h5', compile=False)
model3.layers[0].trainable = False # freeze EmbLayer
for layer in model3.layers[1:]:
lr_d=1e-10,
units=128,
spatial_dr=0.5,
kernel_size1=4,
kernel_size2=4,
dense_units=64,
dr=0.2,
conv_size=32)
# test_loss, test_acc = model.evaluate(x_test, y_test)
# print('test loss:', test_loss, 'test acc:', test_acc)
gold = y_test
predicted = model.predict(x_test).argmax(axis=1)
gold2 = y_test2
predicted2 = model.predict(x_test2).argmax(axis=1)
utils.test_evaluation(gold, predicted)
utils.test_evaluation(gold2, predicted2)
# toy tests
toy_sents = tokenizer.texts_to_sequences([
'the cat sat on the mat', 'what a great movie', 'better not again',
'terrible, worst ever', 'best film ever', 'today is Tuesday'
])
toy_data = pad_sequences(toy_sents, maxlen=MAXLEN)
toy_gold = [1, 2, 0, 0, 2, 1]
prediction = model.predict(toy_data)
print(toy_gold)
print(prediction.argmax(axis=1))
# plot results
# utils.plot(history)