blob: f8c5094443b097bd3fe3465e05f13a2d09104ae1 (
plain) (
tree)
|
|
from blocks.algorithms import AdaDelta
from blocks.bricks import Tanh
from model.lstm import Model
dataset = 'data/logcompil.txt'
io_dim = 256
# An epoch will be composed of 'num_seqs' sequences of len 'seq_len'
# divided in chunks of lengh 'seq_div_size'
num_seqs = 50
seq_len = 5000
seq_div_size = 200
layers = [
{'dim': 1024,
'xreg': (768, 0.1, 10, 10, 10, 2)
},
{'dim': 1024,
'xreg': (768, 0.1, 10, 10, 10, 5)
},
{'dim': 1024,
},
]
activation_function = Tanh()
i2h_all = True # input to all hidden layers or only first layer
h2o_all = True # all hiden layers to output or only last layer
w_noise_std = 0.02
i_dropout = 0.5
l1_reg = 0
step_rule = AdaDelta()
# parameter saving freq (number of batches)
monitor_freq = 100
save_freq = 100
# used for sample generation and IRC mode
sample_temperature = 0.7 #0.5
# do we want to generate samples at times during training?
sample_len = 1000
sample_freq = 100
sample_init = '\nalex\ttu crois?\n'
|
