def __init__(self, options):
"""Initializes the parameters for a LSTM layer of a recurrent neural network.
:type options: dict
:param options: a dictionary of training options
"""
self.options = options
# Initialize the parameters.
self.init_params = OrderedDict()
nin = self.options['dim_word']
dim = self.options['dim']
W = numpy.concatenate([normalized_weight(nin, dim),
normalized_weight(nin, dim),
normalized_weight(nin, dim)],
axis=1)
self.init_params['encoder_W'] = W
n_gates = 3
self.init_params['encoder_b'] = numpy.zeros((n_gates * dim,)).astype('float32')
U = numpy.concatenate([orthogonal_weight(dim),
orthogonal_weight(dim),
orthogonal_weight(dim)],
axis=1)
self.init_params['encoder_U'] = U
Wx = normalized_weight(nin, dim)
self.init_params['encoder_Wx'] = Wx
Ux = orthogonal_weight(dim)
self.init_params['encoder_Ux'] = Ux
self.init_params['encoder_bx'] = numpy.zeros((dim,)).astype('float32')
def __init__(self, options):
"""Initializes the parameters for a feed-forward layer of a neural
network.
:type options: dict
:param options: a dictionary of training options
"""
# Create the parameters.
self.init_params = OrderedDict()
nin = options['dim']
nout = options['dim_word']
self.init_params['ff_logit_lstm_W'] = normalized_weight(nin, nout, scale=0.01, ortho=False)
self.init_params['ff_logit_lstm_b'] = numpy.zeros((nout,)).astype('float32')
nin = options['dim_word']
nout = options['dim_word']
self.init_params['ff_logit_prev_W'] = normalized_weight(nin, nout, scale=0.01, ortho=False)
self.init_params['ff_logit_prev_b'] = numpy.zeros((nout,)).astype('float32')
nin = options['dim_word']
nout = options['n_words']
self.init_params['ff_logit_W'] = normalized_weight(nin, nout, scale=0.01, ortho=True)
self.init_params['ff_logit_b'] = numpy.zeros((nout,)).astype('float32')
def __init__(self, options):
"""Initializes the parameters for a LSTM layer of a recurrent neural network.
:type options: dict
:param options: a dictionary of training options
"""
self.options = options
# Initialize the parameters.
self.init_params = OrderedDict()
nin = self.options['dim_word']
dim = self.options['dim']
W = numpy.concatenate([
normalized_weight(nin, dim),
normalized_weight(nin, dim),
normalized_weight(nin, dim)
],
axis=1)
self.init_params['encoder_W'] = W
n_gates = 3
self.init_params['encoder_b'] = numpy.zeros(
(n_gates * dim, )).astype('float32')
U = numpy.concatenate([
orthogonal_weight(dim),
orthogonal_weight(dim),
orthogonal_weight(dim)
],
axis=1)
self.init_params['encoder_U'] = U
Wx = normalized_weight(nin, dim)
self.init_params['encoder_Wx'] = Wx
Ux = orthogonal_weight(dim)
self.init_params['encoder_Ux'] = Ux
self.init_params['encoder_bx'] = numpy.zeros((dim, )).astype('float32')
def __init__(self, options):
"""Initializes the parameters for the first layer of a neural network
language model, which creates the word embeddings.
:type options: dict
:param options: a dictionary of training options
"""
# Initialize the parameters.
self.init_params = OrderedDict()
nin = options['n_words']
nout = options['dim_word']
self.init_params['Wemb'] = normalized_weight(nin, nout)
def __init__(self, options):
"""Initializes the parameters for the first layer of a neural network
language model, which creates the word embeddings.
:type options: dict
:param options: a dictionary of training options
"""
# Initialize the parameters.
self.init_params = OrderedDict()
nin = options['n_words']
nout = options['dim_word']
self.init_params['Wemb'] = normalized_weight(nin, nout)
def __init__(self, options):
"""Initializes the parameters for a feed-forward layer of a neural
network.
:type options: dict
:param options: a dictionary of training options
"""
# Create the parameters.
self.init_params = OrderedDict()
nin = options['dim']
nout = options['dim_word']
self.init_params['ff_logit_lstm_W'] = normalized_weight(nin,
nout,
scale=0.01,
ortho=False)
self.init_params['ff_logit_lstm_b'] = numpy.zeros(
(nout, )).astype('float32')
nin = options['dim_word']
nout = options['dim_word']
self.init_params['ff_logit_prev_W'] = normalized_weight(nin,
nout,
scale=0.01,
ortho=False)
self.init_params['ff_logit_prev_b'] = numpy.zeros(
(nout, )).astype('float32')
nin = options['dim_word']
nout = options['n_words']
self.init_params['ff_logit_W'] = normalized_weight(nin,
nout,
scale=0.01,
ortho=True)
self.init_params['ff_logit_b'] = numpy.zeros(
(nout, )).astype('float32')