def initialize(self, weight_type="none"):
"""initialize weights
Parameters
----------
weight_type : string
type of weights: "none", "tanh", "sigmoid"
"""
if self.W_h is None:
self.W_h = util.init_weights("W_h",
self.out_dim,
self.in_dim,
weight_type=weight_type)
if self.W_t is None:
self.W_t = util.init_weights("W_t",
self.out_dim,
self.in_dim,
weight_type=weight_type)
if self.bias_h is None:
self.bias_h = util.init_weights("bias_h",
self.out_dim,
weight_type=weight_type)
if self.bias_t is None:
self.bias_t = util.shared_floatx_ones((self.out_dim, ),
value=self.gate_bias,
name="bias_t")
def initialize(self, weight_type="none"):
"""Initialize weights and bias
Parameters
----------
weight_type : string
type of weights: "none", "tanh", "sigmoid"
"""
if self.W==None:
self.W=util.init_weights("W", self.out_dim, self.in_dim, weight_type=weight_type);
if self.use_bias==True and self.bias==None:
self.bias=util.init_weights("bias", self.out_dim, weight_type=weight_type);
def initialize(self, weight_type="none"):
"""Initialize weights for RNN
"""
Layer.initialize(weight_type)
self.hidden = util.init_weights("Hidden",
self.out_dim,
weight_type=weight_type)
def initialize(self, weight_type="none"):
"""initialize weights
Parameters
----------
weight_type : string
type of weights: "none", "tanh", "sigmoid"
"""
if self.W_h is None:
self.W_h=util.init_weights("W_h", self.out_dim, self.in_dim, weight_type=weight_type);
if self.W_t is None:
self.W_t=util.init_weights("W_t", self.out_dim, self.in_dim, weight_type=weight_type);
if self.bias_h is None:
self.bias_h=util.init_weights("bias_h", self.out_dim, weight_type=weight_type);
if self.bias_t is None:
self.bias_t=util.shared_floatx_ones((self.out_dim,), value=self.gate_bias, name="bias_t");
def init_weights(self):
self.W_state=util.init_weights("W_state", self.out_dim*4, self.in_dim, weight_type="sigmoid");
self.W_cell_to_in=util.shared_floatx_nans((self.out_dim,), name="W cell to in");
self.W_cell_to_forget=util.shared_floatx_nans((self.out_dim,), name="W cell to forget");
self.W_cell_to_out=util.shared_floatx_nans((self.out_dim,), name="W cell to out");
self.init_state=util.shared_floatx_zeros((self.out_dim, ), name="initial states");
self.init_cell=util.shared_floatx_zeros((self.out_dim, ), name="initial cell");
def init_weights(self):
self.W_state = util.init_weights("W_state",
self.out_dim * 4,
self.in_dim,
weight_type="sigmoid")
self.W_cell_to_in = util.shared_floatx_nans((self.out_dim, ),
name="W cell to in")
self.W_cell_to_forget = util.shared_floatx_nans(
(self.out_dim, ), name="W cell to forget")
self.W_cell_to_out = util.shared_floatx_nans((self.out_dim, ),
name="W cell to out")
self.init_state = util.shared_floatx_zeros((self.out_dim, ),
name="initial states")
self.init_cell = util.shared_floatx_zeros((self.out_dim, ),
name="initial cell")
def initialize(self, weight_type="none"):
"""Initialize weights and bias
Parameters
----------
weight_type : string
type of weights: "none", "tanh", "sigmoid"
"""
# should have better implementation for convnet weights
fan_in = self.num_channels*np.prod(self.filter_size);
fan_out = self.num_filters*np.prod(self.filter_size);
filter_bound=np.sqrt(6./(fan_in + fan_out));
filter_shape=(self.num_filters, self.num_channels)+(self.filter_size);
self.filters = theano.shared(np.asarray(np.random.uniform(low=-filter_bound,
high=filter_bound,
size=filter_shape),
dtype='float32'),
borrow=True);
if self.use_bias==True:
self.bias=util.init_weights("bias", self.num_filters, weight_type=weight_type);
def initialize(self, weight_type="none"):
"""Initialize weights for RNN
"""
Layer.initialize(weight_type);
self.hidden=util.init_weights("Hidden", self.out_dim, weight_type=weight_type);
