def create_rnn_op(self):
x = layers.data(shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False)
x.stop_gradient = False
rnn_0 = layers.StaticRNN()
with rnn_0.step():
x_t = rnn_0.step_input(x)
mem_pre = rnn_0.memory(shape=[-1, self.input_dim], batch_ref=x)
mem = layers.elementwise_add(x=mem_pre, y=x_t)
rnn_0.update_memory(mem_pre, mem)
rnn_0.output(mem)
rnn_1 = layers.StaticRNN()
with rnn_1.step():
mem_pre = rnn_1.memory(shape=[-1, self.input_dim], batch_ref=x)
x_t = rnn_1.step_input(x)
last_rnn_output = rnn_0()
last_rnn_sum = fluid.layers.reduce_sum(last_rnn_output)
mem = layers.elementwise_add(x=x_t, y=last_rnn_sum)
y = layers.elementwise_add(x=mem_pre, y=mem)
rnn_1.update_memory(mem_pre, mem)
rnn_1.output(y)
return rnn_1()
def setUp(self):
self.setup_program()
self.data_field = {"x", "h_boot"}
self.input_shape = (self.sent_len, self.batch_size, self.input_dim)
self.output_shape = (self.sent_len, self.batch_size, self.input_dim)
self.py_rnn = PySimpleRNN1(self.input_shape, self.output_shape)
with fluid.program_guard(self.main_program, self.startup_program):
x = layers.data(
shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False)
x.stop_gradient = False
h_boot = layers.data(
shape=[self.input_dim], dtype='float32', name='h_boot')
h_boot.stop_gradient = False
forward_only_rnn = layers.StaticRNN()
with forward_only_rnn.step():
h_pre = forward_only_rnn.memory(init=h_boot)
x_t = forward_only_rnn.step_input(x)
h = layers.scale(
x=layers.elementwise_add(
x=h_pre, y=x_t),
scale=self.py_rnn.scale)
forward_only_rnn.update_memory(h_pre, h)
forward_only_rnn.output(h)
forward_only_output = forward_only_rnn()
forward_only_output.stop_gradient = True
self.forward_only_output = layers.mean(forward_only_output)
rnn = layers.StaticRNN()
with rnn.step():
h_pre = rnn.memory(init=h_boot)
x_t = rnn.step_input(x)
h = layers.scale(
x=layers.elementwise_add(
x=h_pre, y=x_t),
scale=self.py_rnn.scale)
rnn.update_memory(h_pre, h)
rnn.output(h)
self.output = layers.mean(rnn())
def create_rnn_op(self):
x = layers.data(
shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False,
**self.p_info)
x.stop_gradient = False
h_boot = layers.data(
shape=[self.input_dim],
dtype='float32',
name='h_boot',
**self.p_info)
h_boot.stop_gradient = False
rnn = layers.StaticRNN(main_program=self.main_program)
with rnn.step():
h_pre = rnn.memory(init=h_boot)
x_t = rnn.step_input(x)
h = layers.scale(
x=layers.elementwise_add(
x=h_pre, y=x_t, **self.p_info),
scale=self.py_rnn.scale,
**self.p_info)
rnn.update_memory(h_pre, h)
rnn.output(h)
return rnn()
def create_rnn_op(self):
x = layers.data(shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False)
x.stop_gradient = False
h_boot1 = layers.data(shape=[self.batch_size, self.input_dim],
dtype='float32',
name='h_boot1',
append_batch_size=False)
h_boot1.stop_gradient = False
h_boot2 = layers.data(shape=[self.batch_size, self.input_dim],
dtype='float32',
name='h_boot2',
append_batch_size=False)
h_boot2.stop_gradient = False
rnn = layers.StaticRNN()
with rnn.step():
h_pre1 = rnn.memory(init=h_boot1)
h_pre2 = rnn.memory(init=h_boot2)
x_t = rnn.step_input(x)
mem1 = layers.scale(x=h_pre1, scale=1.0)
mem2 = layers.scale(x=h_pre2, scale=1.0)
out = layers.sums(input=[mem1, x_t, mem2])
rnn.update_memory(h_pre1, mem1)
rnn.update_memory(h_pre2, mem2)
rnn.output(out)
return rnn()
def create_rnn_op(self):
x = layers.data(shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False)
x.stop_gradient = False
h_boot = layers.data(shape=[self.input_dim],
dtype='float32',
name='h_boot')
h_boot.stop_gradient = False
rnn = layers.StaticRNN()
with rnn.step():
h_pre = rnn.memory(init=h_boot)
x_t = rnn.step_input(x)
temp_l = layers.fc(input=x_t,
size=self.input_dim,
param_attr='W',
bias_attr=False)
temp_r = layers.fc(input=h_pre,
size=self.input_dim,
param_attr='U',
bias_attr=False)
h = layers.sigmoid(x=layers.elementwise_add(x=temp_l, y=temp_r))
rnn.update_memory(h_pre, h)
rnn.output(h)
return rnn()
def create_rnn_op(self):
x = layers.data(shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False)
x.stop_gradient = False
emb = layers.data(
name='emb',
shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
append_batch_size=False)
emb.stop_gradient = False
w1 = layers.data(shape=[self.input_dim, self.input_dim],
dtype='float32',
name='w1',
append_batch_size=False)
w1.stop_gradient = False
w2 = layers.data(shape=[self.input_dim * 2, self.input_dim],
dtype='float32',
name='w2',
append_batch_size=False)
w2.stop_gradient = False
rnn = layers.StaticRNN()
def dot_attention(query, memory):
attn = layers.matmul(query, memory, transpose_y=True)
weight = layers.softmax(attn)
weight_memory = layers.matmul(weight, memory)
return weight_memory, weight
y = layers.matmul(emb, w1)
with rnn.step():
pre_h = rnn.memory(shape=(self.sent_len, self.input_dim),
batch_ref=x,
init_value=0.0)
step_in = rnn.step_input(x)
concat_in = layers.concat([step_in, pre_h], 1)
new_h = layers.matmul(concat_in, w2)
new_h = layers.unsqueeze(new_h, [1])
new_h, _ = dot_attention(new_h, y)
new_h = layers.squeeze(new_h, [1])
rnn.update_memory(pre_h, new_h)
rnn.step_output(new_h)
return rnn()
def static_rnn_net(self):
x = layers.data(shape=[SEQ_LEN, BATCH_SIZE, INPUT_DIM],
dtype="float32",
name="x",
append_batch_size=False)
x.stop_gradient = False
rnn = layers.StaticRNN()
with rnn.step():
x_t = rnn.step_input(x)
h_pre = rnn.memory(shape=[-1, INPUT_DIM], batch_ref=x_t)
h = layers.scale(x=layers.elementwise_add(x=h_pre, y=x_t),
scale=self.scale)
rnn.update_memory(h_pre, h)
rnn.output(h)
return layers.mean(rnn())
def create_rnn_op(self):
x = layers.data(shape=[self.sent_len, self.batch_size, self.input_dim],
dtype='float32',
name='x',
append_batch_size=False)
x.stop_gradient = False
rnn = layers.StaticRNN()
with rnn.step():
mem_pre = rnn.memory(shape=[-1, self.input_dim], batch_ref=x)
x_t = rnn.step_input(x)
mem = layers.elementwise_add(x=mem_pre, y=x_t)
rnn.update_memory(mem_pre, mem)
rnn.output(mem)
return rnn()
def create_rnn_op(self):
x = layers.data(shape=[self.sent_len, self.batch_size, self.input_dim],
dtype="float32",
name="x",
append_batch_size=False)
x.stop_gradient = False
h_boot = layers.data(shape=[self.input_dim],
dtype="float32",
name="h_boot")
h_boot.stop_gradient = True
rnn = layers.StaticRNN()
with rnn.step():
h_pre = rnn.memory(init=h_boot) # init doesn't have gradient
x_t = rnn.step_input(x)
temp_l = layers.fc(
input=x_t,
size=self.input_dim,
param_attr=ParamAttr(
name="W",
initializer=fluid.initializer.ConstantInitializer(1.0)),
bias_attr=False)
temp_r = layers.fc(
input=h_pre,
size=self.input_dim,
param_attr=ParamAttr(
name="U",
initializer=fluid.initializer.ConstantInitializer(0.0)),
bias_attr=False)
h = layers.sigmoid(x=layers.elementwise_add(temp_l, temp_r))
rnn.update_memory(h_pre, h)
rnn.output(h)
return rnn()
