Esempio n. 1
0
 def __init__(self, input_length, input_dim, output_dim, context_dim, attention_hidden_dim, 
              name='AttentionLSTM', truncate_gradient=-1, go_backwards=False,
              weight_init = 'glorot_uniform', inner_init = 'orthogonal', bias_init = 'zero', forget_bias_init = 'one',
              activation='tanh', attention_activation='tanh', inner_activation='hard_sigmoid'):
     super(AttentionLSTM_X, self).__init__()
     self.input_length = input_length
     self.input_dim = input_dim
     self.output_dim = output_dim
     self.context_dim = context_dim
     self.attention_hidden_dim = attention_hidden_dim
     self.name=name
     self.truncate_gradient = truncate_gradient
     self.go_backwards = go_backwards
     
     self.required_input_sets = [['input_single', 'context'], ['input_sequence', 'context'], ['input_sequence', 'input_mask', 'context']]
     self.output_names = ['output_last', 'output_sequence', 'output_sequence_with_alpha', 'output_last_with_alpha']
     self.required_function_sets = [['weight_init', 'inner_init', 'bias_init', 'forget_bias_init', 'activation', 'attention_activation']]
     self.set_output('output_last', self.output_last)
     self.set_output('output_sequence', self.output_sequence)
     self.set_output('output_last_with_alpha', self.output_last_with_alpha)
     self.set_output('output_sequence_with_alpha', self.output_sequence_with_alpha)
     self.set_function('activation', activations.get(activation))
     self.set_function('attention_activation', activations.get(attention_activation))
     self.set_function('inner_activation', activations.get(inner_activation))
     self.set_function('weight_init', initializations.get(weight_init))
     self.set_function('inner_init', initializations.get(weight_init))
     self.set_function('bias_init', initializations.get(bias_init))
     self.set_function('forget_bias_init', initializations.get(forget_bias_init))
Esempio n. 2
0
File: core.py Progetto: lxastro/dlx
 def __init__(self, input_dim, output_dim, name='Dense', weight_init='glorot_uniform', bias_init='zero', activation='linear'):
     super(Dense, self).__init__()
     self.input_dim = input_dim
     self.output_dim = output_dim
     self.name=name
     
     self.required_input_sets = [['input']]
     self.output_names = ['output']
     self.required_function_sets = [['weight_init', 'bias_init', 'activation']]
     self.set_output('output', self.output)
     self.set_function('activation', activations.get(activation))
     self.set_function('weight_init', initializations.get(weight_init))
     self.set_function('bias_init', initializations.get(bias_init))
Esempio n. 3
0
 def __init__(self, input_length, input_dim, output_dim, name='RNN', truncate_gradient=-1, go_backwards=False,
              weight_init = 'glorot_uniform', inner_init = 'orthogonal', bias_init = 'zero', activation='sigmoid'):
     super(RNN, self).__init__()
     self.input_length = input_length
     self.input_dim = input_dim
     self.output_dim = output_dim
     self.name=name
     self.truncate_gradient = truncate_gradient
     self.go_backwards = go_backwards
     
     self.required_input_sets = [['input_single'], ['input_sequence'], ['input_sequence', 'input_mask']]
     self.output_names = ['output_last', 'output_sequence']
     self.required_function_sets = [['weight_init', 'inner_init', 'bias_init', 'activation']]
     self.set_output('output_last', self.output_last)
     self.set_output('output_sequence', self.output_sequence)
     self.set_function('activation', activations.get(activation))
     self.set_function('weight_init', initializations.get(weight_init))
     self.set_function('inner_init', initializations.get(weight_init))
     self.set_function('bias_init', initializations.get(bias_init))
Esempio n. 4
0
from dlx import initializations
import numpy as np

print '\n------------------------------------------------------------'
print 'Test: dlx.initializations'

def two(shape):
    return 2. * np.ones(shape)

init_dict = {'uniform': (2,3,4),
             'normal': (2,3,4),
             'lecun_uniform':[2,3,4],
             'glorot_normal':(2,3,4),
             'glorot_uniform':(2,3,4),
             'he_normal':[2,3,4],
             'he_uniform':(2,3,4),
             'orthogonal':(4,4),
             'identity':(4,4),
             'zero':(2,3,4),
             'one':(2,3,4),
             two: (2,3,4)
            }

for fun, shape in init_dict.items():
    val = initializations.get(fun)(shape)
    print fun, shape, val.dtype, ':'
    print val
    print