def output_h_vals(self, train=False):
if self.inputs_dict.has_key('input_single'):
input = self.get_input('input_single', train) #(nb_sample, input_dim)
X = TU.repeat(input, self.input_length) # (input_length, nb_sample, input_dim)
mask = None
else:
input = self.get_input('input_sequence', train) # (nb_sample, input_length, input_dim)
X = input.dimshuffle((1, 0, 2)) # (input_length, nb_sample, input_dim)
mask = self.get_input_mask('input_sequence',train) # (nb_sample, input_length)
if mask:
mask = T.cast(mask, dtype='int8').dimshuffle((1, 0, 'x')) # (input_length, nb_sample, 1)
#h_0 = T.zeros((X.shape[1], self.output_dim), X.dtype) # (nb_samples, output_dim)
h_0 = self._get_initial_state(X)
if mask:
h_vals, _ = theano.scan( self.step,
sequences=[mask, X],
outputs_info=h_0,
non_sequences=[self.W, self.U, self.b],
truncate_gradient=self.truncate_gradient,
go_backwards=self.go_backwards,
strict=True)
else:
h_vals, _ = theano.scan( self.step_no_mask,
sequences=[X],
outputs_info=h_0,
non_sequences=[self.W, self.U, self.b],
truncate_gradient=self.truncate_gradient,
go_backwards=self.go_backwards,
strict=True)
return h_vals #(input_length, nb_samples, output_dim)
def output_h_vals(self, train=False):
if self.inputs_dict.has_key('input_single'):
input = self.get_input('input_single', train) #(nb_sample, input_dim)
X = TU.repeat(input, self.input_length) # (input_length, nb_sample, input_dim)
mask = None
else:
input = self.get_input('input_sequence', train) # (nb_sample, input_length, input_dim)
X = input.dimshuffle((1, 0, 2)) # (input_length, nb_sample, input_dim)
mask = self.get_input_mask('input_sequence',train) # (nb_sample, input_length)
if mask:
mask = T.cast(mask, dtype='int8').dimshuffle((1, 0, 'x')) # (input_length, nb_sample, 1)
context = self.get_input('context', train) # (nb_samples, nb_context, context_dim)
att_ctx = T.dot(context, self.W_ctx_att) + self.b_att # (nb_samples, nb_context, context_dim) + (context_dim,)
h_0 = T.zeros((X.shape[1], self.output_dim), X.dtype) # (nb_samples, output_dim)
#h_0 = self._get_initial_state(X)
c_0 = T.zeros((X.shape[1], self.output_dim), X.dtype) # (nb_samples, output_dim)
if mask:
revals, _ = theano.scan( self.step,
sequences=[mask, X],
outputs_info=[h_0, c_0],
non_sequences=[context, att_ctx, self.W, self.V, self.U, self.b, self.W_h_att, self.w_att_prj],
truncate_gradient=self.truncate_gradient,
go_backwards=self.go_backwards,
strict=True)
else:
revals, _ = theano.scan( self.step_no_mask,
sequences=[X],
outputs_info=[h_0, c_0],
non_sequences=[context, att_ctx, self.W, self.V, self.U, self.b, self.W_h_att, self.w_att_prj],
truncate_gradient=self.truncate_gradient,
go_backwards=self.go_backwards,
strict=True)
return revals[0] #(input_length, nb_samples, output_dim)
def output(self, train=False):
X = self._default_input(train)
return TU.repeat(X, self.n).dimshuffle((1, 0, 2))