def __init__(self, state_num, action_num, experience_replay=True):
self.state_num = state_num
self.action_num = action_num
self.experience_replay = experience_replay
self.experience_pool = []
self.model = get_model(state_num, action_num)
train_conf = TrainerConfig()
train_conf.learning_rate = LEARNING_RATE
train_conf.weight_l2 = 0
self.trainer = SGDTrainer(self.model, train_conf)
self.trainer.training_names = []
self.trainer.training_variables = []
self.thread_lock = threading.Lock()
self.epsilon = EPSILON
self.tick = 0
def __init__(self, state_num, action_num, experience_replay=True):
self.state_num = state_num
self.action_num = action_num
self.experience_replay = experience_replay
self.experience_pool = []
self.model = get_model(state_num, action_num)
train_conf = TrainerConfig()
train_conf.learning_rate = LEARNING_RATE
train_conf.weight_l2 = 0
self.trainer = SGDTrainer(self.model, train_conf)
self.trainer.training_names = []
self.trainer.training_variables = []
self.thread_lock = threading.Lock()
self.epsilon = EPSILON
self.tick = 0
model.stack(Dropout(p=dropout_p_0), Dense(n, init=init, disable_bias=True), BatchNormalization(), Activation(activation))
#model.stack(Dropout(p=dropout_p_0), BatchNormalization())
for _ in range(T):
#model.stack(HighwayLayerLRDropoutBatchNorm(activation=activation, gate_bias=gate_bias, projection_dim=d, d_p_0 = dropout_p_h_0, d_p_1 = dropout_p_h_1, init=init))
model.stack(HighwayLayerLRDiagDropoutBatchNorm(activation=activation, gate_bias=gate_bias, projection_dim=d, d_p_0 = dropout_p_h_0, d_p_1 = dropout_p_h_1, init=init, quasi_ortho_init=True))
#model.stack(BatchNormalization(),Dropout(p=dropout_p_2), Dense(10, init=init))
model.stack(Dropout(p=dropout_p_2), Dense(10, init=init))
learning_rate_start = 3e-3
#learning_rate_target = 3e-7
#learning_rate_epochs = 100
#learning_rate_decay = (learning_rate_target / learning_rate_start) ** (1.0 / learning_rate_epochs)
conf = TrainerConfig()
conf.learning_rate = LearningRateAnnealer.learning_rate(learning_rate_start)
#conf.gradient_clipping = 1
conf.patience = 20
#conf.gradient_tolerance = 5
conf.avoid_nan = True
conf.min_improvement = 1e-10
#trainer = MomentumTrainer(model)
trainer = AdamTrainer(model, conf)
mnist = MiniBatches(MnistDataset(), batch_size=100)
#mnist = MiniBatches(MnistDatasetSmallValid(), batch_size=100)
#trainer.run(mnist, controllers=[IncrementalLearningRateAnnealer(trainer, 0, learning_rate_decay)])
trainer.run(mnist, controllers=[LearningRateAnnealer(trainer, 3, 14)])
logging.info('Setting best parameters for testing.')
import numpy as np
from deepy.layers.recurrent import RecurrentLayer, RecurrentNetwork
from deepy.conf import NetworkConfig, TrainerConfig
from deepy.utils.functions import FLOATX
from deepy import SGDTrainer
logging.basicConfig(level=logging.INFO)
if __name__ == '__main__':
net_conf = NetworkConfig(input_size=6)
net_conf.layers = [RecurrentLayer(size=10, activation='sigmoid', bptt=True)]
trainer_conf = TrainerConfig()
trainer_conf.learning_rate = 0.03
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)
data = np.array([[1,0,0,0,0,0],
[0,1,0,0,0,0],
[0,0,1,0,0,0],
[0,0,0,1,0,0],
[0,0,0,0,1,0],
[0,0,0,0,0,1],
[0,1,0,0,0,0],
[0,0,1,0,0,0],
HighwayLayerLRDiagDropoutBatchNorm(activation=activation,
gate_bias=gate_bias,
projection_dim=d,
d_p_0=dropout_p_h_0,
d_p_1=dropout_p_h_1,
init=init,
quasi_ortho_init=True))
#model.stack(BatchNormalization(),Dropout(p=dropout_p_2), Dense(10, init=init))
model.stack(Dropout(p=dropout_p_2), Dense(10, init=init))
learning_rate_start = 3e-3
#learning_rate_target = 3e-7
#learning_rate_epochs = 100
#learning_rate_decay = (learning_rate_target / learning_rate_start) ** (1.0 / learning_rate_epochs)
conf = TrainerConfig()
conf.learning_rate = LearningRateAnnealer.learning_rate(
learning_rate_start)
#conf.gradient_clipping = 1
conf.patience = 20
#conf.gradient_tolerance = 5
conf.avoid_nan = True
conf.min_improvement = 1e-10
#trainer = MomentumTrainer(model)
trainer = AdamTrainer(model, conf)
mnist = MiniBatches(MnistDataset(), batch_size=100)
#mnist = MiniBatches(MnistDatasetSmallValid(), batch_size=100)
#trainer.run(mnist, controllers=[IncrementalLearningRateAnnealer(trainer, 0, learning_rate_decay)])
trainer.run(mnist, controllers=[LearningRateAnnealer(trainer, 3, 14)])
logging.info('Setting best parameters for testing.')
if __name__ == '__main__':
ap = ArgumentParser()
ap.add_argument("--model", default=os.path.join(os.path.dirname(__file__), "models", "sequence_adding_100_2.gz"))
args = ap.parse_args()
model = NeuralRegressor(input_dim=2, input_tensor=3)
model.stack(IRNN(hidden_size=100, input_type="sequence",
output_type="one"),
Dense(1))
if os.path.exists(args.model):
model.load_params(args.model)
conf = TrainerConfig()
conf.learning_rate = LearningRateAnnealer.learning_rate(0.01)
conf.gradient_clipping = 3
conf.patience = 50
conf.gradient_tolerance = 5
conf.avoid_nan = False
trainer = SGDTrainer(model, conf)
annealer = LearningRateAnnealer(patience=20)
trainer.run(batch_set, controllers=[annealer])
model.save_params(args.model)
print "Identity matrix weight:"
print model.first_layer().W_h.get_value().diagonal()
