def bilstm(X_train, X_test, Y_train, Y_test, wordembeddings):
np.random.seed(1234)
tf.random.set_seed(1234)
random.seed(1234)
max_length_sentence = X_train.str.split().str.len().max()
tokenizer = Tokenizer(filters='!"#$%&()*+,-./:;<=>?@[\\]^_`{|}~\t\n\'',
lower=True)
tokenizer.fit_on_texts(X_train)
word_index = tokenizer.word_index
EMBEDDING_DIM = 300
vocabulary_size = len(word_index) + 1
print('Found %s unique tokens.' % len(word_index))
sequences_train = tokenizer.texts_to_sequences(X_train)
sequences_valid = tokenizer.texts_to_sequences(X_test)
X_train = pad_sequences(sequences_train, maxlen=max_length_sentence)
X_val = pad_sequences(sequences_valid, maxlen=X_train.shape[1])
y_train = np.asarray(Y_train)
y_val = np.asarray(Y_test)
#print(word_index)
'''
print('Shape of data tensor:', X_train.shape)
print('Shape of data tensor:', X_val.shape)
print('Shape of data tensor:', y_train.shape)
print('Shape of data tensor:', y_val.shape)
print(X_train)
print("*"*100)
print(X_val)
print("*"*100)
print(y_train)
print("*"*100)
print(y_val)
'''
embedding_matrix = np.zeros((vocabulary_size, EMBEDDING_DIM))
for word, i in word_index.items():
if (word in wordembeddings.keys()):
embedding_vector = wordembeddings[word]
if len(embedding_vector) == 0: #if array is empty
embedding_vector = wordembeddings[word.title()]
if len(embedding_vector) == 0:
embedding_vector = wordembeddings[word.upper()]
if len(embedding_vector) == 0:
embedding_vector = np.array([
round(np.random.rand(), 8) for i in range(0, 300)
])
else:
#print("WORD NOT IN DICT",word)
embedding_vector = np.array(
[round(np.random.rand(), 8) for i in range(0, 300)])
if len(embedding_vector) != 0:
embedding_matrix[i] = embedding_vector
embedding_layer = Embedding(vocabulary_size,
EMBEDDING_DIM,
weights=[embedding_matrix],
trainable=False) #Try with True
inputs = Input(shape=(X_train.shape[1], ))
model = (Embedding(vocabulary_size,
EMBEDDING_DIM,
input_length=max_length_sentence,
weights=[embedding_matrix]))(inputs)
model = Bidirectional(GRU(64))(
model) # !!!!!!! CHANGE THIS FOR OTHER MODELS
model = (Dense(900, activation='relu'))(model)
model = (Dense(400, activation='relu'))(model)
model = (Dense(250, activation='relu'))(model)
model = (Dense(204, activation='softmax'))(model)
model = Model(inputs=inputs, outputs=model)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
model.summary()
callbacks = [EarlyStopping(monitor='val_loss')]
hist_adam = model.fit(
X_train,
y_train,
batch_size=1000,
epochs=200,
verbose=1,
validation_data=(X_val, y_val),
callbacks=callbacks
) #!!!!!!!!!!!!!!!!!!!!!!!CHANGE BATCH SIZE TO 1000 #change epochs to 200
model.save(config.bigru_prepocessed_dataset1_chai
) # !!!!!!! CHANGE THIS FOR OTHER MODELS
y_pred = model.predict(X_val)
print(y_pred)
y_val_class = pd.DataFrame(y_val).idxmax(axis=1)
print(y_val_class)
y_val_class_argmax = np.argmax(y_val, axis=1)
y_pred_class_argmax = np.argmax(y_pred, axis=1)
y_pred_class = pd.DataFrame(y_pred).idxmax(axis=1)
print(y_pred_class)
print(classification_report(y_val_class, y_pred_class))
plt.suptitle('Optimizer : Adam', fontsize=10)
plt.ylabel('Loss', fontsize=16)
plt.xlabel('Epoch', fontsize=14)
plt.plot(hist_adam.history['loss'], color='b', label='Training Loss')
plt.plot(hist_adam.history['val_loss'], color='r', label='Validation Loss')
plt.legend(loc='upper right')
plt.savefig(
'/home/ubuntu/asset_classification/results/bigru_model_dataset1_preprocessed_chai.png'
) # !!!!!!! CHANGE THIS FOR OTHER MODELS
tf.keras.utils.plot_model(
model, to_file=config.bigru_architecture,
show_shapes=True) # !!!!!!! CHANGE THIS FOR OTHER MODELS
return (y_pred, y_val_class, y_pred_class, y_val_class_argmax,
y_pred_class_argmax)
encoder = Bidirectional(LSTM(HIDDEN_SIZE),
merge_mode='sum',
name="encoder_lstm")(inputs)
decoder = RepeatVector(max(sent_lens))(encoder)
decoder = Bidirectional(LSTM(EMBED_SIZE, return_sequences=True),
merge_mode='sum')(decoder)
autoencoder = Model(inputs, decoder)
autoencoder.compile(optimizer='sgd', loss='mse')
num_train_step = len(X_train) // BATCH_SIZE
num_test_step = len(X_test) // BATCH_SIZE
hist = autoencoder.fit_generator(train_gen, steps_per_epoch = num_train_step, epochs = 20, validation_data = test_gen, \
validation_steps=num_test_step)
encoder = Model(autoencoder.input,
autoencoder.get_layer("encoder_lstm").output)
def compare_cosine_similarity(x, y):
return np.dot(x, y) / (np.linalg.norm(x, 2) * np.linalg.norm(y, 2))
for i in range(10):
Xtest, Ytest = next(test_gen)
Ytest_ = autoencoder.predict(Xtest)
Xvec = encoder.predict(Xtest)
Yvec = encoder.predict(Ytest_)
for j in range(Xvec.shape[0]):
print(compare_cosine_similarity(Xvec[j], Yvec[j]))
def train(is_debug=False):
train_data = open("dataset/atis-2.train.w-intent.iob", "r").readlines()
test_data = open("dataset/atis-2.dev.w-intent.iob", "r").readlines()
train_data_ed = data_pipeline(train_data)
test_data_ed = data_pipeline(test_data)
word2index, index2word, slot2index, index2slot, intent2index, index2intent = \
get_info_from_training_data(train_data_ed)
# print("slot2index: ", slot2index)
# print("index2slot: ", index2slot)
index_train = to_index(train_data_ed, word2index, slot2index, intent2index)
index_test = to_index(test_data_ed, word2index, slot2index, intent2index)
intents = [item[3] for item in index_train]
intent_labels = np.eye(intent_size)[np.array(intents)]
intent_train = [item[0] for item in index_train]
intent_train = np.array(intent_train)
slot_train = [item[2] for item in index_train]
slot_train = np.array(slot_train)
slot_train_target = np.insert(slot_train,
slot_train.shape[1] - 1,
values=0,
axis=1)
slot_train_target = np.delete(slot_train_target, 0, axis=1)
slot_train_target = np.eye(slot_size)[slot_train_target]
seq = np.array(index_test[0][0])
import tensorflow as tf
from keras.layers import Lambda
print_func = Lambda(lambda x: tf.Print(x, [tf.shape(x)]))
squeeze = Lambda(lambda x: tf.squeeze(x, axis=2))
# encoder define
input_voc = Input(shape=(None, 1))
embedding_voc = Embedding(input_dim=vocab_size,
output_dim=embedding_size,
mask_zero=True)
embedding_voc_out = embedding_voc(input_voc)
encoder_lstm1 = Bidirectional(LSTM(units=hidden_size,
dropout=0.7,
return_sequences=True),
merge_mode="concat")
embedding_voc_out = squeeze(embedding_voc_out)
# embedding_voc_out = K.backend.squeeze(embedding_voc_out, axis=0)
encoder_lstm1_out = encoder_lstm1(embedding_voc_out)
# encoder
encoder = Bidirectional(LSTM(units=hidden_size,
dropout=0.7,
return_sequences=False,
return_state=True),
merge_mode="concat")
# encoder output, states
encoder_out, forward_h, forward_c, backward_h, backward_c = encoder(
encoder_lstm1_out)
encoder_state = [forward_h, forward_c, backward_h, backward_c]
# intent
intent = Dense(intent_size, activation="linear")(encoder_out)
intent = Dense(intent_size, activation="softmax")(intent)
# encode state
forward_h = K.layers.concatenate([forward_h, backward_h], 1)
forward_c = K.layers.concatenate([forward_c, backward_c], 1)
encoder_state = [forward_h, forward_c]
# decoder define
input_slot = Input(shape=(None, 1))
embedding_slot = Embedding(input_dim=slot_size,
output_dim=embedding_size,
mask_zero=True)
embedding_slot_out = embedding_slot(input_slot)
embedding_slot_out = squeeze(embedding_slot_out)
decoder_lstm1 = LSTM(units=hidden_size * 2,
dropout=0.7,
return_state=True,
return_sequences=True)
decoder_lstm1_out, forward_h, forward_c = decoder_lstm1(
embedding_slot_out, initial_state=[forward_h, forward_c])
decoder = LSTM(units=hidden_size * 2,
dropout=0.7,
return_sequences=True,
return_state=False)
decoder_output = decoder(decoder_lstm1_out)
dense1 = Dense(slot_size, activation="linear")
dense1_out = dense1(decoder_output)
dense2 = Dense(slot_size, activation="softmax")
dense2_out = dense2(dense1_out)
model = Model(inputs=[input_voc, input_slot], outputs=[intent, dense2_out])
# print(model.summary())
def intent_slot_loss(y_true, y_pred):
y_slot_true = y_true[0]
y_intent_true = y_true[1]
y_slot_pred = y_pred[0]
y_intent_pred = y_pred[1]
return K.losses.categorical_crossentropy(
y_slot_true, y_slot_pred) + K.losses.categorical_crossentropy(
y_intent_true, y_intent_pred)
model.compile(optimizer="adam",
loss=categorical_crossentropy,
metrics=["acc"])
# acc = model.fit([np.expand_dims(intent_train, 2), np.expand_dims(slot_train, 2)], [intent_labels, slot_train_target], batch_size=batch_size, epochs=1)
# print(acc)
# model.save_weights("nlu.hdf5")
model.load_weights("nlu.hdf5")
## inference
encoder = Model(input_voc, encoder_state)
intenter = Model(input_voc, intent)
# print(encoder.summary())
# decoder
decoder_in = Input(shape=(1, ))
decoder_state_in_h = Input(shape=(hidden_size * 2, ))
decoder_state_in_c = Input(shape=(hidden_size * 2, ))
decoder_state_in = [decoder_state_in_h, decoder_state_in_c]
decoder_out, decoder_h, decoder_c = decoder_lstm1(
embedding_slot(decoder_in), initial_state=decoder_state_in)
decoder_state_out = [decoder_h, decoder_c] #output
decoder_output = decoder(decoder_out)
dense1_out = dense1(decoder_output)
dense2_out = dense2(dense1_out) #output
decoder_model = Model([decoder_in] + decoder_state_in,
[dense2_out] + decoder_state_out)
print(decoder_model.summary())
i = 0
print(seq)
seq = np.expand_dims(np.array(seq), 0)
seq = np.expand_dims(np.array(seq), 2)
encoder_state = encoder.predict(seq)
intent = intenter.predict(seq)
intent = np.argmax(intent)
print(index2intent[intent])
index = slot2index["<STR>"]
while i < input_steps:
i = i + 1
decoder_out, state_h, state_c = decoder_model.predict(
[np.array([index])] + encoder_state)
index = np.argmax(decoder_out)
encoder_state = [state_h, state_c]
print(index2slot[index])
history = model.fit(X_train,
np.array(y_train),
batch_size=BATCH_SIZE,
epochs=EPOCHS,
validation_split=0.1,
verbose=2)
def pred2label(pred):
out = []
for pred_i in pred:
out_i = []
for p in pred_i:
p_i = np.argmax(p)
out_i.append(idx2tag[p_i].replace("PADword", "O"))
out.append(out_i)
return out
pred_y = model.predict(X_test)
pred_y = np.argmax(pred_y, axis=-1)
y_test_true = np.argmax(y_test, -1)
pred_y = [[index_to_tag[i].replace("PADword", "O")
for i in pred_y[index]][0:len(y_test[index])]
for index in range(len(pred_y))]
y_test_true = [[index_to_tag[i] for i in row] for row in y_test]
print('LSTM Classification Report\n',
metrics.flat_classification_report(pred_y, y_test_true))
# show figure
plt.show()
score = model.evaluate([X_w_te,np.array(X_c_te).reshape((len(X_c_te), max_len, max_len_char))], np.array(y_te), batch_size=batch_size,verbose=1)
print(model.metrics_names)
print("Score:")
print(score)
# ## Prediction on test set
from seqeval.metrics import precision_score, recall_score, f1_score, classification_report
# print("Input:")
# print(X_te[0])
# print("Supposed output:")
# print(y_te)
# print(np.array(y_te))
test_pred = model.predict([X_w_te,np.array(X_c_te).reshape((len(X_c_te), max_len, max_len_char))], verbose=1)
# print("Prediction result:")
# print(test_pred[0])
idx2tag = {i: w for w,i in tags2idx.items()}
tags_size = len(idx2tag)
idx2tag[tags_size] = 'O'
def pred2label(pred):
out = []
for pred_i in pred:
out_i = []
for p in pred_i:
p_i = np.argmax(p)
out_i.append(idx2tag[p_i].replace("PAD", "O"))
out.append(out_i)
return out
all_train = all[:int(0.8 * all.shape[0]), ...]
Y_train = Y[:int(0.8 * all.shape[0]), ...]
Y_train_dense = np.reshape(Y_train, (Y_train.shape[0], Y_train.shape[1]))
Y_train_dense = np.argmax(Y_train_dense, axis=-1)
all_test = all[int(0.8 * all.shape[0]):, ...]
Y_test = Y[int(0.8 * all.shape[0]):, ...]
# pu.db
Y_test_dense = np.reshape(Y_test, (Y_test.shape[0], Y_test.shape[1]))
Y_test_dense = np.argmax(Y_test_dense, axis=-1)
for i in xrange(100):
print i
model.fit(all_train, Y_train, batch_size=1000, epochs=5, verbose=1)
Y_pred_train = model.predict(all_train, batch_size=1000)
Y_pred_test = model.predict(all_test, batch_size=1000)
Y_pred_train_dense = np.reshape(
Y_pred_train, (Y_pred_train.shape[0], Y_pred_train.shape[1]))
Y_pred_train_dense = np.argmax(Y_pred_train_dense, axis=-1)
Y_pred_test_dense = np.reshape(
Y_pred_test, (Y_pred_test.shape[0], Y_pred_test.shape[1]))
Y_pred_test_dense = np.argmax(Y_pred_test_dense, axis=-1)
train_acc = np.sum(
Y_pred_train_dense == Y_train_dense) * 100.0 / len(Y_pred_train_dense)
val_acc = np.sum(
Y_pred_test_dense == Y_test_dense) * 100.0 / len(Y_pred_test_dense)
