def train_rnn(item_fc, h_0, output_type):
shifted_item_fc = fluid_sequence_advance(item_fc, OOV=0)
drnn = fluid.layers.DynamicRNN()
with drnn.block():
last_item_fc = drnn.step_input(shifted_item_fc)
cur_h_0 = drnn.memory(init=h_0, need_reorder=True)
# step_input will remove lod info
last_item_fc = layers.lod_reset(last_item_fc, cur_h_0)
next_h_0 = layers.dynamic_gru(
last_item_fc,
size=4,
h_0=cur_h_0,
param_attr=fluid.ParamAttr(name="item_gru.w_0"),
bias_attr=fluid.ParamAttr(name="item_gru.b_0"))
if output_type == 'c_Q':
Q = layers.fc(next_h_0,
1,
param_attr=fluid.ParamAttr(name="c_Q_fc.w_0"),
bias_attr=fluid.ParamAttr(name="c_Q_fc.b_0"))
drnn.output(Q)
elif output_type == 'max_Q':
Q = layers.fc(next_h_0,
1,
param_attr=fluid.ParamAttr(name="max_Q_fc.w_0"),
bias_attr=fluid.ParamAttr(name="max_Q_fc.b_0"))
drnn.output(Q)
# update
drnn.update_memory(cur_h_0, next_h_0)
drnn_output = drnn()
return drnn_output
def __call__(self, input, h_0=None):
return layers.dynamic_gru(
input=input,
size=size,
param_attr=self.attr_holder.param_attr,
bias_attr=self.attr_holder.bias_attr,
is_reverse=is_reverse,
gate_activation=gate_activation,
candidate_activation=candidate_activation,
h_0=h_0,
origin_mode=origin_mode)
def __call__(self, input, state=None):
gru_input = layers.fc(input=input,
size=self.hidden_dim * 3,
param_attr=self.input_param_attr,
bias_attr=self.input_bias_attr)
return layers.dynamic_gru(
input=gru_input,
size=self.hidden_dim,
param_attr=self.param_attr,
bias_attr=self.bias_attr,
is_reverse=self.is_reverse,
gate_activation=self.gate_activation,
candidate_activation=self.candidate_activation,
h_0=state,
origin_mode=self.origin_mode)
def gru_fun(gru_in, name=None, is_reverse=False):
""" gru fun """
fw_last_array = []
fw_in = gru_in
for i in range(num_layers):
fw_gru_in = layers.fc(
input=fw_in,
size=hidden_size * 3,
param_attr=fluid.ParamAttr(name=name + "_fc_w"),
bias_attr=fluid.ParamAttr(name=name + "_fc_b"))
fw_gru_out = layers.dynamic_gru(
input=fw_gru_in,
size=hidden_size,
param_attr=fluid.ParamAttr(name=name + "_w"),
bias_attr=fluid.ParamAttr(name=name + "_b"),
origin_mode=True,
is_reverse=is_reverse)
fw_in = fw_gru_out
if is_reverse:
fw_last_hidden = layers.sequence_first_step(fw_gru_out)
else:
fw_last_hidden = layers.sequence_last_step(fw_gru_out)
if last_mask:
fw_last_hidden = layers.elementwise_mul(fw_last_hidden,
last_mask,
axis=0)
fw_last_array.append(fw_last_hidden)
if num_layers == 1:
final_fw_last_hidden = layers.unsqueeze(fw_last_array[0], axes=[0])
else:
final_fw_last_hidden = layers.concat(fw_last_array, axis=0)
final_fw_last_hidden = layers.reshape(
final_fw_last_hidden, shape=[num_layers, -1, hidden_size])
final_fw_out = fw_in
return final_fw_out, final_fw_last_hidden