def lstmp_encoder(input_seq, gate_size, h_0, c_0, para_name, proj_size, args):
# A lstm encoder implementation with projection.
# Linear transformation part for input gate, output gate, forget gate
# and cell activation vectors need be done outside of dynamic_lstm.
# So the output size is 4 times of gate_size.
init = None
init_b = None
input_proj = layers.fc(input=input_seq,
param_attr=fluid.ParamAttr(name=para_name +
'_gate_w',
initializer=init),
size=gate_size * 4,
act=None,
bias_attr=False)
hidden, cell = layers.dynamic_lstmp(
input=input_proj,
size=gate_size * 4,
proj_size=proj_size,
h_0=h_0,
c_0=c_0,
use_peepholes=False,
proj_clip=proj_clip,
cell_clip=cell_clip,
proj_activation="identity",
param_attr=fluid.ParamAttr(initializer=init),
bias_attr=fluid.ParamAttr(initializer=init_b))
return hidden, cell, input_proj
def test_dynamic_lstmp(self):
program = Program()
with program_guard(program):
hidden_dim, proj_dim = 16, 8
seq_data = layers.data(
name='seq_data', shape=[10, 10], dtype='float32', lod_level=1)
fc_out = layers.fc(input=seq_data, size=4 * hidden_dim)
self.assertIsNotNone(
layers.dynamic_lstmp(
input=fc_out, size=4 * hidden_dim, proj_size=proj_dim))
print(str(program))
def test_dynamic_lstmp(self):
program = Program()
with program_guard(program):
hidden_dim, proj_dim = 16, 8
seq_data = layers.data(
name='seq_data', shape=[10, 10], dtype='float32', lod_level=1)
fc_out = layers.fc(input=seq_data, size=4 * hidden_dim)
self.assertIsNotNone(
layers.dynamic_lstmp(
input=fc_out, size=4 * hidden_dim, proj_size=proj_dim))
print(str(program))
def __call__(self, input):
return layers.dynamic_lstmp(
input=input,
size=size,
proj_size=proj_size,
param_attr=self.attr_holder.param_attr,
bias_attr=self.attr_holder.bias_attr,
use_peepholes=use_peepholes,
is_reverse=is_reverse,
gate_activation=gate_activation,
cell_activation=cell_activation,
candidate_activation=candidate_activation,
proj_activation=proj_activation,
dtype=dtype,
cell_clip=cell_clip,
proj_clip=proj_clip)
def lstmp_encoder(input_seq, gate_size, h_0, c_0, para_name, proj_size, test_mode, args):
# A lstm encoder implementation with projection.
# Linear transformation part for input gate, output gate, forget gate
# and cell activation vectors need be done outside of dynamic_lstm.
# So the output size is 4 times of gate_size.
if args.para_init:
init = fluid.initializer.Constant(args.init1)
init_b = fluid.initializer.Constant(0.0)
else:
init = None
init_b = None
input_seq = dropout(input_seq, test_mode, args)
input_proj = layers.fc(input=input_seq,
param_attr=fluid.ParamAttr(
name=para_name + '_gate_w', initializer=init),
size=gate_size * 4,
act=None,
bias_attr=False)
if args.debug:
layers.Print(input_seq, message='input_seq', summarize=10)
layers.Print(input_proj, message='input_proj', summarize=10)
hidden, cell = layers.dynamic_lstmp(
input=input_proj,
size=gate_size * 4,
proj_size=proj_size,
h_0=h_0,
c_0=c_0,
use_peepholes=False,
proj_clip=args.proj_clip,
cell_clip=args.cell_clip,
proj_activation="identity",
param_attr=fluid.ParamAttr(initializer=init),
bias_attr=fluid.ParamAttr(initializer=init_b))
return hidden, cell, input_proj
