mlp_hidden_dims = [30]
lstm_hidden_dim = 20
noise_std = 0.1
dropout = 0.5
num_feats = 100
use_ensembling = False
randomize_feats = False
# step_rule = Momentum(learning_rate=0.01, momentum=0.9)
step_rule = AdaDelta()
# iter_scheme = LogregOrderTransposeIt(10, True, 'model_param/logreg_param.pkl', 500)
iter_scheme = RandomTransposeIt(None, True, num_feats, True)
valid_iter_scheme = RandomTransposeIt(None, True, None if use_ensembling else num_feats, True)
param_desc = '%s-%d-%s-%s-%s%s' % (repr(mlp_hidden_dims),
lstm_hidden_dim,
repr(noise_std),
repr(dropout),
'E' if use_ensembling else '',
'R' if randomize_feats else '')
pt_freq = 1
def construct_model(input_dim, out_dim):
# Construct the model
r = tensor.fmatrix('r')
x = tensor.fmatrix('x')
momentum = 0.9
if step_rule_name == 'adadelta':
step_rule = AdaDelta()
elif step_rule_name == 'rmsprop':
step_rule = RMSProp()
elif step_rule_name == 'momentum':
step_rule_name = "mom%s,%s" % (repr(learning_rate), repr(momentum))
step_rule = Momentum(learning_rate=learning_rate, momentum=momentum)
elif step_rule_name == 'adam':
step_rule = Adam()
else:
raise ValueError("No such step rule: " + step_rule_name)
ibatchsize = None
iter_scheme = RandomTransposeIt(ibatchsize, False, None, False)
valid_iter_scheme = RandomTransposeIt(ibatchsize, False, None, False)
w_noise_std = 0.05
r_dropout = 0.0
s_dropout = 0.0
i_dropout = 0.0
a_dropout = 0.0
center_feats = True
normalize_feats = True
randomize_feats = False
train_on_valid = False
reconstruction_penalty = 1
from blocks.bricks import Rectifier, MLP, Softmax, Tanh
from blocks.initialization import IsotropicGaussian, Constant
from blocks.filter import VariableFilter
from blocks.roles import WEIGHT
from blocks.graph import ComputationGraph, apply_noise
from datastream import RandomTransposeIt
learning_rate = 0.00001
momentum = 0.9
# step_rule = Momentum(learning_rate=learning_rate, momentum=momentum)
step_rule = AdaDelta()
ibatchsize = 100
iter_scheme = RandomTransposeIt(ibatchsize, True, None, True)
valid_iter_scheme = iter_scheme
noise_std = 0.1
randomize_feats = False
hidden_dims = [2, 2]
activation_functions = [Tanh() for _ in hidden_dims] + [Tanh()]
param_desc = '%s-%s-%s-mom%s,%s-i%d' % (
repr(hidden_dims), repr(noise_std), 'R' if randomize_feats else '',
repr(learning_rate), repr(momentum), ibatchsize)
pt_freq = 11