def test_train(self):
train_path = internal_resource("lm_test/train")
valid_path = internal_resource("lm_test/valid")
vocab = Vocab()
vocab.load(train_path)
train_data = RNNDataGenerator(vocab, train_path, target_vector=False,
history_len=-1, _just_test=False, fixed_length=False, progress=True)
valid_data = RNNDataGenerator(vocab, valid_path, target_vector=False,
history_len=-1, _just_test=False, fixed_length=False, progress=True)
net_conf = NetworkConfig(input_size=vocab.size)
net_conf.layers = [RecurrentLayer(size=50, activation='relu')]
trainer_conf = TrainerConfig()
trainer_conf.learning_rate = 0.1
trainer_conf.weight_l2 = 0.0001
trainer_conf.hidden_l2 = 0.0001
trainer_conf.monitor_frequency = trainer_conf.validation_frequency = trainer_conf.test_frequency = 1
network = RecurrentNetwork(net_conf)
trainer = SGDTrainer(network, config=trainer_conf)
start_time = time.time()
for k in list(trainer.train(train_data, valid_data)):
pass
print k
end_time = time.time()
network.save_params("/tmp/lmparam.gz")
print "elapsed time:", (end_time - start_time) / 60, "mins"
def get_wre_network(model_path=""):
net_conf = NetworkConfig(input_size=300)
net_conf.layers = [WRELayer(size=300)]
trainer_conf = TrainerConfig()
trainer_conf.learning_rate = 0.01
trainer_conf.weight_l2 = 0.0001
trainer_conf.hidden_l2 = 0.0001
trainer_conf.monitor_frequency = trainer_conf.validation_frequency = trainer_conf.test_frequency = 1
network = GeneralAutoEncoder(net_conf)
if os.path.exists(model_path):
network.load_params(model_path)
return network
def get_wre_network(model_path=""):
net_conf = NetworkConfig(input_size=300)
net_conf.layers = [WRELayer(size=300)]
trainer_conf = TrainerConfig()
trainer_conf.learning_rate = 0.01
trainer_conf.weight_l2 = 0.0001
trainer_conf.hidden_l2 = 0.0001
trainer_conf.monitor_frequency = trainer_conf.validation_frequency = trainer_conf.test_frequency = 1
network = GeneralAutoEncoder(net_conf)
if os.path.exists(model_path):
network.load_params(model_path)
return network
def test_train(self):
train_path = internal_resource("lm_test/train")
valid_path = internal_resource("lm_test/valid")
vocab = Vocab()
vocab.load(train_path)
train_data = RNNDataGenerator(vocab,
train_path,
target_vector=False,
history_len=-1,
_just_test=False,
fixed_length=False,
progress=True)
valid_data = RNNDataGenerator(vocab,
valid_path,
target_vector=False,
history_len=-1,
_just_test=False,
fixed_length=False,
progress=True)
net_conf = NetworkConfig(input_size=vocab.size)
net_conf.layers = [RecurrentLayer(size=50, activation='relu')]
trainer_conf = TrainerConfig()
trainer_conf.learning_rate = 0.1
trainer_conf.weight_l2 = 0.0001
trainer_conf.hidden_l2 = 0.0001
trainer_conf.monitor_frequency = trainer_conf.validation_frequency = trainer_conf.test_frequency = 1
network = RecurrentNetwork(net_conf)
trainer = SGDTrainer(network, config=trainer_conf)
start_time = time.time()
for k in list(trainer.train(train_data, valid_data)):
pass
print k
end_time = time.time()
network.save_params("/tmp/lmparam.gz")
print "elapsed time:", (end_time - start_time) / 60, "mins"
args = ap.parse_args()
print "[ARGS]", args
builder = WordDataBuilder(
"/home/hadoop/data/morpho/data/topwords.1500.txt")
"""
Setup network
"""
pretrain_model = args.premodel
model_path = args.model
net_conf = NetworkConfig(input_size=builder.input_size)
net_conf.layers = [WRELayer(size=100)]
trainer_conf = TrainerConfig()
trainer_conf.learning_rate = 0.01
trainer_conf.weight_l2 = 0.0001
trainer_conf.hidden_l2 = 0.0001
trainer_conf.monitor_frequency = trainer_conf.validation_frequency = trainer_conf.test_frequency = 1
network = GeneralAutoEncoder(net_conf)
trainer = AdaGradTrainer(network, config=trainer_conf)
"""
Run the network
"""
start_time = time.time()
if os.path.exists(pretrain_model):
network.load_params(pretrain_model)
return network
if __name__ == '__main__':
builder = ParaphraseClassifierDataBuilder("/home/hadoop/data/paraphrase/mspr/train3.pkl.gz",
"/home/hadoop/data/paraphrase/mspr/test3.pkl.gz")
# builder = ParaphraseClassifierDataBuilder("/home/hadoop/data/paraphrase/data/bank_classify_train2.pkl.gz",
# "/home/hadoop/data/paraphrase/data/bank_classify_test2.pkl.gz")
model_path = "/home/hadoop/play/model_zoo/parahrase_classifier4.gz"
net = get_classify_network()
if os.path.exists(model_path) and False:
net.load_params(model_path)
trainer_conf = TrainerConfig()
trainer_conf.learning_rate = 0.01
# trainer_conf.update_l1 = 0.0001
trainer_conf.weight_l2 = 0.0001
trainer_conf.hidden_l2 = 0.0001
trainer_conf.monitor_frequency = trainer_conf.validation_frequency = trainer_conf.test_frequency = 1
trainer = AdaDeltaTrainer(net, config=trainer_conf)
start_time = time.time()
c = 1
for _ in trainer.train(builder.train_data(), builder.valid_data(), builder.test_data()):
c += 1
if c >= 80:
trainer.test(0, builder.test_data())