def get_attention(att_class, **kwargs):
import OpLSTMCustom
recurrent_transform = RecurrentTransform.get_dummy_recurrent_transform(
att_class.name, **kwargs)
assert isinstance(recurrent_transform, att_class)
f = OpLSTMCustom.register_func(recurrent_transform)
return f
def scan(self,
x,
z,
non_sequences,
i,
outputs_info,
W_re,
W_in,
b,
go_backwards=False,
truncate_gradient=-1):
assert self.parent.recurrent_transform
import OpLSTMCustom
op = OpLSTMCustom.register_func(self.parent.recurrent_transform)
custom_vars = self.parent.recurrent_transform.get_sorted_custom_vars()
initial_state_vars = self.parent.recurrent_transform.get_sorted_state_vars_initial(
)
# See OpLSTMCustom.LSTMCustomOp.
# Inputs args are: Z, c, y0, i, W_re, custom input vars, initial state vars
# Results: (output) Y, (gates and cell state) H, (final cell state) d, state vars sequences
op_res = op(z[::-(2 * go_backwards - 1)], outputs_info[1],
outputs_info[0], i[::-(2 * go_backwards - 1)],
T.ones((i.shape[1], ), 'float32'), W_re,
*(custom_vars + initial_state_vars))
result = [op_res[0], op_res[2].dimshuffle('x', 0, 1)] + op_res[3:]
assert len(result) == len(outputs_info)
return result
def scan(self, x, z, non_sequences, i, outputs_info, W_re, W_in, b, go_backwards = False, truncate_gradient = -1):
assert self.parent.recurrent_transform
import OpLSTMCustom
op = OpLSTMCustom.register_func(self.parent.recurrent_transform)
custom_vars = self.parent.recurrent_transform.get_sorted_custom_vars()
initial_state_vars = self.parent.recurrent_transform.get_sorted_state_vars_initial()
# See OpLSTMCustom.LSTMCustomOp.
# Inputs args are: Z, c, y0, i, W_re, custom input vars, initial state vars
# Results: (output) Y, (gates and cell state) H, (final cell state) d, state vars sequences
op_res = op(z[::-(2 * go_backwards - 1)],
outputs_info[1], outputs_info[0], i[::-(2 * go_backwards - 1)], T.ones((i.shape[1],),'float32'), W_re, *(custom_vars + initial_state_vars))
result = [ op_res[0], op_res[2].dimshuffle('x',0,1) ] + op_res[3:]
assert len(result) == len(outputs_info)
return result
def get_attention(att_class, **kwargs): import OpLSTMCustom recurrent_transform = RecurrentTransform.get_dummy_recurrent_transform(att_class.name, **kwargs) assert isinstance(recurrent_transform, att_class) f = OpLSTMCustom.register_func(recurrent_transform) return f