from models import BiLSTM
from utils import aspects_sent_convert, read_file_lines, write_file_lines
from import_dy import dy
from random_inits import np
# Use a sentence level classifier to automatically label sentences.
if __name__ == "__main__":
label_files = sys.argv[1].split(',')
lines = list()
for label_file in label_files:
lines += read_file_lines(label_file)
model = dy.ParameterCollection()
W2I = pickle.load(open("models/SentenceLevelBiLSTM" + "/Model.meta", "rb"))
network = BiLSTM(1, 300, 150, W2I, model)
network.load_model("models/SentenceLevelBiLSTM/Model")
new_samples = list()
for i in range(0, len(lines), 1000):
print str(i) + " Sentences Labeled"
samples = aspects_sent_convert(lines[i:i + 1000])
for sample in samples:
dy.renew_cg()
sent = sample.sent
h = network(sent[0])
dist = network.classify(h)
prediction = np.argmax(dist.npvalue())
sample_str = sent[0] + u"<->" + str(prediction)
command_line_file.write(arg + "\n")
command_line_file.close()
else:
print "Model Folder Already exists."
sys.exit(1)
network.save_meta(model_path + "/Model.meta")
trainer = dy.AdamTrainer(model)
dists = np.load(args.dist_path)
xr_train(network,
trainer,
train,
dev,
args.epochs,
30,
model_path,
dists,
T_value=args.T,
batch_size=args.batch_size)
network.load_model(model_path + "/Model")
acc, loss, f1 = aspect_test(test, network, 30)
test_acc = "Test Accuarcy:" + str(acc)
test_loss = "Test Loss:" + str(loss)
test_f1 = "Test F1:" + str(f1)
print test_acc
print test_loss
print test_f1
test_results_file = open(model_path + '/Results/test_results.txt', 'w')
test_results_file.write(test_acc + "\n" + test_loss + "\n" + test_f1)
test_results_file.close()
dynet_config.set(random_seed=args.seed)
from train_funcs import *
dyparams = dy.DynetParams()
dyparams.set_autobatch(True)
from models import BiLSTM, LSTMAtt
model = dy.ParameterCollection()
model_params = model
W2I, vecs = pickle.load(
open("models/" + args.pretrained_model + "/Model.meta", "rb")), None
# Create the base bilstm network, which will be pretrained.
base_network = BiLSTM(1, 300, 150, W2I, model, vecs)
# Load the pretrained bilstm network
base_network.load_model("models/" + args.pretrained_model + "/Model")
# Add attention weights to the pretrained bilstm network.
network = LSTMAtt(base_network, model, 300, 150)
if not os.path.exists(model_path):
os.makedirs(model_path)
os.makedirs(model_path + "/Results")
command_line_file = open(model_path + "/Command_Lines.txt", 'w')
for arg in sys.argv:
command_line_file.write(arg + "\n")
command_line_file.close()
else:
print "Model Folder Already exists."
sys.exit(1)
network.save_meta(model_path + "/Model.meta")
trainer = dy.AdamTrainer(model)
# Finetune the model.
