def Train(model, batch_size, n_epochs, train_c_new, train_r_new, train_l_new,
dev_c_new, dev_r_new, dev_l_new):
print("Now training the model...")
# histories = Histories()
histories = my_callbacks.Histories()
start_time = time.time()
# start_time = time.time()
# compute_recall_ks(y_pred[:,0])
# print("---model evaluation time takes %s seconds ---" % (time.time() - start_time))
bestAcc = 0.0
patience = 0
print("\tbatch_size={}, nb_epoch={}".format(batch_size, n_epochs))
# for ep in range(1, args.n_epochs):
for ep in range(1, n_epochs):
model.fit([train_c_new, train_r_new],
train_l_new,
batch_size=batch_size,
epochs=n_epochs,
callbacks=[histories],
validation_data=([dev_c_new, dev_r_new], dev_l_new),
verbose=1)
curAcc = histories.accs[0]
if curAcc >= bestAcc:
bestAcc = curAcc
patience = 0
else:
patience = patience + 1
# classify the test set
y_pred = model.predict([test_c_new, test_r_new])
print("Perform on test set after Epoch: " + str(ep) + "...!")
recall_k = compute_recall_ks(y_pred[:, 0])
# stop training the model when patience = 10
if patience > 10:
print("Early stopping at epoch: " + str(ep))
break
print("---Training finished, model training time takes %s seconds ---" %
(time.time() - start_time))
return model
neg_branch = shared_cnn(neg_input)
#concatenated = merge([pos_branch, neg_branch], mode='concat',name="coherence_out")
concatenated = concatenate([pos_branch, neg_branch],
axis=-1,
name="coherence_out")
# output is two latent coherence score
final_model = Model([pos_input, neg_input], concatenated)
#final_model.compile(loss='ranking_loss', optimizer='adam')
final_model.compile(loss={'coherence_out': ranking_loss},
optimizer=opts.learn_alg)
# setting callback
histories = my_callbacks.Histories()
print(shared_cnn.summary())
#print(final_model.summary())
print("------------------------------------------------")
#writing model name
if opts.f_list != "":
ff = opts.f_list
m_type = "Ext.CNN."
else:
ff = "None"
m_type = "CNN."
model_name = opts.model_dir + m_type + str(opts.p_num) + "_" + str(opts.dropout_ratio) + "_"+ str(opts.emb_size) + "_"+ str(opts.maxlen) + "_" \
def main():
parser = argparse.ArgumentParser()
parser.register('type', 'bool', str2bool)
parser.add_argument('--emb_dim',
type=str,
default=300,
help='Embeddings dimension')
parser.add_argument('--hidden_size',
type=int,
default=300,
help='Hidden size')
parser.add_argument('--batch_size',
type=int,
default=256,
help='Batch size')
parser.add_argument('--n_epochs', type=int, default=50, help='Num epochs')
parser.add_argument('--lr',
type=float,
default=0.001,
help='Learning rate')
parser.add_argument('--optimizer',
type=str,
default='adam',
help='Optimizer')
parser.add_argument('--n_recurrent_layers',
type=int,
default=1,
help='Num recurrent layers')
parser.add_argument('--input_dir',
type=str,
default='./dataset/',
help='Input dir')
parser.add_argument('--save_model',
type='bool',
default=True,
help='Whether to save the model')
parser.add_argument('--model_fname',
type=str,
default='model/dual_encoder_lstm_classifier.h5',
help='Model filename')
parser.add_argument('--embedding_file',
type=str,
default='embeddings/glove.840B.300d.txt',
help='Embedding filename')
parser.add_argument('--seed', type=int, default=1337, help='Random seed')
args = parser.parse_args()
print('Model args: ', args)
np.random.seed(args.seed)
print("Starting...")
# first, build index mapping words in the embeddings set
# to their embedding vector
print('Now indexing word vectors...')
embeddings_index = {}
f = open(args.embedding_file, 'r')
for line in f:
values = line.split()
word = values[0]
try:
coefs = np.asarray(values[1:], dtype='float32')
except ValueError:
continue
embeddings_index[word] = coefs
f.close()
MAX_SEQUENCE_LENGTH, MAX_NB_WORDS, word_index = pickle.load(
open(args.input_dir + 'params.pkl', 'rb'))
print("MAX_SEQUENCE_LENGTH: {}".format(MAX_SEQUENCE_LENGTH))
print("MAX_NB_WORDS: {}".format(MAX_NB_WORDS))
print("Now loading embedding matrix...")
num_words = min(MAX_NB_WORDS, len(word_index)) + 1
embedding_matrix = np.zeros((num_words, args.emb_dim))
for word, i in word_index.items():
if i >= MAX_NB_WORDS:
continue
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None:
# words not found in embedding index will be all-zeros.
embedding_matrix[i] = embedding_vector
print("Now building dual encoder lstm model...")
# define lstm encoder
encoder = Sequential()
encoder.add(
Embedding(output_dim=args.emb_dim,
input_dim=MAX_NB_WORDS,
input_length=MAX_SEQUENCE_LENGTH,
weights=[embedding_matrix],
mask_zero=True,
trainable=True))
encoder.add(LSTM(units=args.hidden_size))
context_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
response_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
# encode the context and the response
context_branch = encoder(context_input)
response_branch = encoder(response_input)
concatenated = merge([context_branch, response_branch], mode='mul')
out = Dense((1), activation="sigmoid")(concatenated)
dual_encoder = Model([context_input, response_input], out)
dual_encoder.compile(loss='binary_crossentropy', optimizer=args.optimizer)
print(encoder.summary())
print(dual_encoder.summary())
print("Now loading UDC data...")
train_c, train_r, train_l = pickle.load(
open(args.input_dir + 'train.pkl', 'rb'))
test_c, test_r, test_l = pickle.load(
open(args.input_dir + 'test.pkl', 'rb'))
dev_c, dev_r, dev_l = pickle.load(open(args.input_dir + 'dev.pkl', 'rb'))
print('Found %s training samples.' % len(train_c))
print('Found %s dev samples.' % len(dev_c))
print('Found %s test samples.' % len(test_c))
print("Now training the model...")
histories = my_callbacks.Histories()
bestAcc = 0.0
patience = 0
print("\tbatch_size={}, nb_epoch={}".format(args.batch_size,
args.n_epochs))
for ep in range(1, args.n_epochs):
dual_encoder.fit([train_c, train_r],
train_l,
batch_size=args.batch_size,
epochs=1,
callbacks=[histories],
validation_data=([dev_c, dev_r], dev_l),
verbose=1)
curAcc = histories.accs[0]
if curAcc >= bestAcc:
bestAcc = curAcc
patience = 0
else:
patience = patience + 1
# classify the test set
y_pred = dual_encoder.predict([test_c, test_r])
print("Perform on test set after Epoch: " + str(ep) + "...!")
recall_k = compute_recall_ks(y_pred[:, 0])
# stop training the model when patience = 10
if patience > 10:
print("Early stopping at epoch: " + str(ep))
break
if args.save_model:
print("Now saving the model... at {}".format(args.model_fname))
dual_encoder.save(args.model_fname)
def main():
parser = argparse.ArgumentParser()
parser.register('type', 'bool', str2bool)
parser.add_argument('--emb_dim',
type=str,
default=300,
help='Embeddings dimension')
parser.add_argument('--hidden_size',
type=int,
default=300,
help='Hidden size')
parser.add_argument('--batch_size',
type=int,
default=256,
help='Batch size')
parser.add_argument('--n_epochs', type=int, default=50, help='Num epochs')
parser.add_argument('--lr',
type=float,
default=0.001,
help='Learning rate')
parser.add_argument('--optimizer',
type=str,
default='adam',
help='Optimizer')
parser.add_argument("--dropout_ratio",
type=float,
default=0.5,
help="ratio of cells to drop out")
parser.add_argument('--n_recurrent_layers',
type=int,
default=1,
help='Num recurrent layers')
parser.add_argument("--w_size",
type=int,
default=5,
help="window size length of neighborhood in words")
parser.add_argument("--pool_length",
type=int,
default=6,
help="length for max pooling")
parser.add_argument(
"--nb_filter",
type=int,
default=150,
help="nb of filter to be applied in convolution over words")
parser.add_argument('--input_dir',
type=str,
default='./dataset/',
help='Input dir')
parser.add_argument('--save_model',
type='bool',
default=True,
help='Whether to save the model')
parser.add_argument('--model_fname',
type=str,
default='model/dual_encoder_lstm_classifier.h5',
help='Model filename')
parser.add_argument('--embedding_file',
type=str,
default='embeddings/glove.840B.300d.txt',
help='Embedding filename')
parser.add_argument('--seed', type=int, default=1337, help='Random seed')
args = parser.parse_args()
print('Model args: ', args)
np.random.seed(args.seed)
if not os.path.exists(args.model_fname):
print("No pre-trained model...")
print("Start building model...")
# first, build index mapping words in the embeddings set
# to their embedding vector
print('Indexing word vectors.')
embeddings_index = {}
f = open(args.embedding_file, 'r')
for line in f:
values = line.split()
word = values[0]
#coefs = np.asarray(values[1:], dtype='float32')
try:
coefs = np.asarray(values[1:], dtype='float32')
except ValueError:
continue
embeddings_index[word] = coefs
f.close()
print("Now loading UDC data...")
train_c, train_r, train_l = pickle.load(
open(args.input_dir + 'train.pkl', 'rb'))
test_c, test_r, test_l = pickle.load(
open(args.input_dir + 'test.pkl', 'rb'))
dev_c, dev_r, dev_l = pickle.load(
open(args.input_dir + 'dev.pkl', 'rb'))
print('Found %s training samples.' % len(train_c))
print('Found %s dev samples.' % len(dev_c))
print('Found %s test samples.' % len(test_c))
MAX_SEQUENCE_LENGTH, MAX_NB_WORDS, word_index = pickle.load(
open(args.input_dir + 'params.pkl', 'rb'))
print("MAX_SEQUENCE_LENGTH: {}".format(MAX_SEQUENCE_LENGTH))
print("MAX_NB_WORDS: {}".format(MAX_NB_WORDS))
vocabs, E = init_vocab(args.emb_dim)
print("Now loading entity-grid data...")
train_egrid, train_label = load_and_numberize_egrids_with_labels(
filelist="./dataset/list.train",
maxlen=MAX_SEQUENCE_LENGTH,
w_size=args.w_size,
vocabs=vocabs)
dev_egrid, dev_label = load_and_numberize_egrids_with_labels(
filelist="./dataset/list.dev",
maxlen=MAX_SEQUENCE_LENGTH,
w_size=args.w_size,
vocabs=vocabs)
test_egrid, test_label = load_and_numberize_egrids_with_labels(
filelist="./dataset/list.test",
maxlen=MAX_SEQUENCE_LENGTH,
w_size=args.w_size,
vocabs=vocabs)
#print (train_label[:10])
#print (list(train_l[:10]))
#assert train_label == list(train_l)
#assert dev_label == list(dev_l)
#assert test_label == list(test_l)
#randomly shuffle the training data
#np.random.shuffle(train_egrid)
print("Now loading embedding matrix...")
num_words = min(MAX_NB_WORDS, len(word_index)) + 1
embedding_matrix = np.zeros((num_words, args.emb_dim))
for word, i in word_index.items():
if i >= MAX_NB_WORDS:
continue
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None:
# words not found in embedding index will be all-zeros.
embedding_matrix[i] = embedding_vector
print("Now building the dual encoder lstm model...")
encoder = Sequential()
encoder.add(
Embedding(output_dim=args.emb_dim,
input_dim=MAX_NB_WORDS,
input_length=MAX_SEQUENCE_LENGTH,
weights=[embedding_matrix],
mask_zero=True,
trainable=True))
encoder.add(LSTM(units=args.hidden_size))
print("Now building the CNN egrid model...")
sent_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
x = Embedding(output_dim=args.emb_dim,
weights=[E],
input_dim=len(vocabs),
input_length=MAX_SEQUENCE_LENGTH,
trainable=True)(sent_input)
x = Convolution1D(nb_filter=args.nb_filter,
filter_length=args.w_size,
border_mode='valid',
activation='relu',
subsample_length=1)(x)
x = MaxPooling1D(pool_length=args.pool_length)(x)
x = Dropout(args.dropout_ratio)(x)
x = Flatten()(x)
x = Dropout(args.dropout_ratio)(x)
x = Dense(300)(x)
cnn = Model(sent_input, x)
context_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
response_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
egrid_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
# these two models will share eveything from shared_cnn
context_branch = encoder(context_input)
response_branch = encoder(response_input)
context_branch_cnn = cnn(egrid_input)
concatenated = merge([context_branch, response_branch], mode='mul')
concatenated = merge([concatenated, context_branch_cnn], mode='concat')
out = Dense((1), activation="sigmoid")(concatenated)
model = Model([context_input, response_input, egrid_input], out)
model.compile(loss='binary_crossentropy', optimizer=args.optimizer)
print(model.summary())
print("Now training the model...")
histories = my_callbacks.Histories()
bestAcc = 0.0
patience = 0
print("\tbatch_size={}, nb_epoch={}".format(args.batch_size,
args.n_epochs))
for ep in range(1, args.n_epochs):
#model.fit([train_c, train_r], train_l,
#batch_size=args.batch_size, nb_epoch=1, callbacks=[histories],
#validation_data=([dev_c, dev_r], dev_l), verbose=1)
model.fit([train_c, train_r, train_egrid],
train_l,
batch_size=args.batch_size,
epochs=1,
callbacks=[histories],
validation_data=([dev_c, dev_r, dev_egrid], dev_l),
verbose=1)
#model.save(model_name + "_ep." + str(ep) + ".h5")
curAcc = histories.accs[0]
if curAcc >= bestAcc:
bestAcc = curAcc
patience = 0
else:
patience = patience + 1
#doing classify the test set
y_pred = model.predict([test_c, test_r, test_egrid])
print("Perform on test set after Epoch: " + str(ep) + "...!")
recall_k = compute_recall_ks(y_pred[:, 0])
#stop the model whch patience = 8
if patience > 10:
print("Early stopping at epoch: " + str(ep))
break
if args.save_model:
print("Now saving the model... at {}".format(args.model_fname))
model.save(args.model_fname)
else:
print("Found pre-trained model...")
model = K_load_model(args.model_fname)
return model
def main():
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
parser = argparse.ArgumentParser()
parser.register('type', 'bool', str2bool)
parser.add_argument('--emb_dim',
type=str,
default=300,
help='Embeddings dimension')
parser.add_argument('--emb_trainable',
type='bool',
default=True,
help='Whether fine tune embeddings')
parser.add_argument('--hidden_size',
type=int,
default=300,
help='Hidden size')
parser.add_argument('--hidden_size_lstm',
type=int,
default=200,
help='Hidden size')
parser.add_argument('--batch_size',
type=int,
default=256,
help='Batch size')
parser.add_argument('--n_epochs', type=int, default=50, help='Num epochs')
parser.add_argument('--lr',
type=float,
default=0.001,
help='Learning rate')
parser.add_argument('--optimizer',
type=str,
default='adam',
help='Optimizer')
parser.add_argument('--n_recurrent_layers',
type=int,
default=1,
help='Num recurrent layers')
parser.add_argument('--input_dir',
type=str,
default='./dataset/',
help='Input dir')
parser.add_argument('--save_model',
type='bool',
default=True,
help='Whether to save the model')
parser.add_argument('--model_fname',
type=str,
default='model/model.h5',
help='Model filename')
parser.add_argument('--embedding_file',
type=str,
default='embeddings/embeddings.vec',
help='Embedding filename')
parser.add_argument('--seed', type=int, default=1337, help='Random seed')
args = parser.parse_args()
print('Model args: ', args)
np.random.seed(args.seed)
print("Starting...")
# first, build index mapping words in the embeddings set
# to their embedding vector
print('Now indexing word vectors...')
embeddings_index = {}
f = open(args.embedding_file, 'r')
for line in f:
values = line.split()
word = values[0]
try:
coefs = np.asarray(values[1:], dtype='float32')
except ValueError:
continue
embeddings_index[word] = coefs
f.close()
MAX_SEQUENCE_LENGTH, MAX_NB_WORDS, word_index = pickle.load(
open(args.input_dir + 'params.pkl', 'rb'))
print("MAX_SEQUENCE_LENGTH: {}".format(MAX_SEQUENCE_LENGTH))
print("MAX_NB_WORDS: {}".format(MAX_NB_WORDS))
print("Now loading embedding matrix...")
num_words = min(MAX_NB_WORDS, len(word_index))
embedding_matrix = np.zeros((num_words, args.emb_dim))
for word, i in word_index.items():
if i >= MAX_NB_WORDS:
continue
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None:
embedding_matrix[i] = embedding_vector
print("Now building dual encoder lstm model...")
# define lstm encoder
encoder = Sequential()
encoder_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
embedding = Embedding(output_dim=args.emb_dim,
input_dim=MAX_NB_WORDS,
input_length=MAX_SEQUENCE_LENGTH,
weights=[embedding_matrix],
mask_zero=True,
trainable=args.emb_trainable)
embedded_input = embedding(encoder_input)
output = LSTM(units=args.hidden_size)(embedded_input)
encoder = Model(encoder_input, [output, embedded_input])
print(encoder.summary())
context_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
response_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32')
# encode the context and the response
context_branch, context_embed = encoder(context_input)
response_branch, response_embed = encoder(response_input)
# compute the sequence level similarity vector
S = keras.layers.multiply([context_branch, response_branch])
# compute the word level similarity matrix
embed_mul = keras.layers.dot([context_embed, response_embed], axes=2)
# transform the word level similarity matrix into a vector
W = LSTM(units=200)(embed_mul)
# concatenate the word and sequence level similarity vectors
concatenated = keras.layers.concatenate([S, W])
out = Dense((1), activation="sigmoid")(concatenated)
model = Model([context_input, response_input], out)
model.compile(loss='binary_crossentropy', optimizer=args.optimizer)
print(model.summary())
print("Now loading data...")
train_c, train_r, train_l = pickle.load(
open(args.input_dir + 'train.pkl', 'rb'))
test_c, test_r, test_l = pickle.load(
open(args.input_dir + 'test.pkl', 'rb'))
dev_c, dev_r, dev_l = pickle.load(open(args.input_dir + 'dev.pkl', 'rb'))
print('Found %s training samples.' % len(train_c))
print('Found %s dev samples.' % len(dev_c))
print('Found %s test samples.' % len(test_c))
print("Now training the model...")
histories = my_callbacks.Histories()
bestAcc = 0.0
patience = 0
print("\tbatch_size={}, nb_epoch={}".format(args.batch_size,
args.n_epochs))
for ep in range(1, args.n_epochs):
model.fit([train_c, train_r],
train_l,
batch_size=args.batch_size,
epochs=1,
callbacks=[histories],
validation_data=([dev_c, dev_r], dev_l),
verbose=1)
curAcc = histories.accs[0]
if curAcc >= bestAcc:
bestAcc = curAcc
patience = 0
if args.save_model:
print("Now saving the model... at {}".format(args.model_fname))
model.save(args.model_fname)
else:
patience = patience + 1
# stop training the model when patience = 5
if patience > 5:
print("Early stopping at epoch: " + str(ep))
break
