def sampling(self, inputs, sampling_type):
decode_len = inputs['decode_len']
user_feature = self.user_encode(inputs)
item_embedding = self._build_embeddings(inputs, self.item_slot_names)
item_fc = self.item_fc_op(item_embedding)
pos = fluid_sequence_get_pos(item_fc)
pos_embed = self.dict_data_embed_op['pos'](pos)
if self._candidate_encode:
cand_encoding = self.candidate_encode(item_fc)
init_hidden = self.candidate_encode_fc_op(
layers.concat([user_feature, cand_encoding], 1))
else:
init_hidden = user_feature
eps = inputs['eps'] if sampling_type == 'eps_greedy' else None
eta = inputs['eta'] if sampling_type == 'softmax' else None
sampled_id = self.sampling_rnn(item_fc,
h_0=init_hidden,
pos_embed=pos_embed,
forward_func=self.sampling_rnn_forward,
sampling_type=sampling_type,
eps=eps,
eta=eta)
sampled_id = self._cut_by_decode_len(
layers.lod_reset(sampled_id, item_fc), decode_len)
return sampled_id
def forward(self, inputs, mode):
"""
don't use sequence_expand for backward
"""
is_test = True if (mode in ['test', 'inference']) else False
# encode
user_embedding = self._build_embeddings(inputs, self.user_slot_names)
user_feature = self.user_feature_fc_op(user_embedding)
# item embed and pos embed
item_embedding = self._build_embeddings(inputs, self.item_slot_names)
item_fc = self.item_fc_op(item_embedding)
pos = fluid_sequence_get_pos(item_fc)
pos_embed = self.dict_data_embed_op['pos'](pos)
input_embed = layers.concat([
item_fc, pos_embed,
layers.sequence_expand_as(user_feature, item_fc)
], 1)
# transformer
trans_in = self.input_embed_fc_op(input_embed)
decoding = self.transformer_decode(is_test, trans_in)
click_prob = self.output_fc2_op(self.output_fc1_op(decoding))
return click_prob
def dynamic_rnn(self, item_fc, h_0, output_type=None, double_type=None, double_id=None):
drnn = fluid.layers.DynamicRNN()
pos = fluid_sequence_get_pos(item_fc)
with drnn.block():
cur_item_fc = drnn.step_input(item_fc)
cur_h_0 = drnn.memory(init=h_0, need_reorder=True)
cur_item_fc = layers.lod_reset(cur_item_fc, cur_h_0)
next_h_0 = self.simple_step_rnn(cur_item_fc, h_0=cur_h_0)
if output_type == 'c_Q':
Q = self.out_Q_fc2_op(self.out_Q_fc1_op(next_h_0))
drnn.output(Q)
elif output_type in ['max_Q', 'double_Q']:
# batch_size = 2
# item_fc: lod = [0,4,7]
# cur_h_0: lod = [0,1,2]
item_fc = drnn.static_input(item_fc)
pos = drnn.static_input(pos)
cur_step = drnn.memory(shape=[1], dtype='int64', value=0)
expand_h_0 = layers.sequence_expand(cur_h_0, item_fc) # lod = [0,1,2,3,4,5,6,7]
new_item_fc = layers.lod_reset(item_fc, expand_h_0) # lod = [0,1,2,3,4,5,6,7]
next_expand_h_0 = self.simple_step_rnn(new_item_fc, expand_h_0) # lod = [0,1,2,3,4,5,6,7]
next_expand_h_0 = layers.lod_reset(next_expand_h_0, item_fc) # lod = [0,4,7]
expand_Q = self.out_Q_fc2_op(self.out_Q_fc1_op(next_expand_h_0))
cur_step_id = layers.slice(cur_step, axes=[0, 1], starts=[0, 0], ends=[1, 1])
mask = layers.cast(pos >= cur_step_id, 'float32')
expand_Q = expand_Q * mask
if output_type == 'max_Q':
max_Q = layers.sequence_pool(expand_Q, 'max') # lod = [0,1,2]
drnn.output(max_Q)
elif output_type == 'double_Q':
if double_type == 'max_id':
max_id = self.eps_greedy_sampling(expand_Q, mask, eps=0)
drnn.output(max_id)
elif double_type == 'double_Q':
cur_double_id = drnn.step_input(double_id)
double_Q = fluid_sequence_index(expand_Q, cur_double_id)
drnn.output(double_Q)
# update
next_step = cur_step + 1
drnn.update_memory(cur_step, next_step)
elif output_type == 'hidden':
drnn.output(next_h_0)
else:
raise NotImplementedError(output_type)
# update
drnn.update_memory(cur_h_0, next_h_0)
drnn_output = drnn()
return drnn_output
def forward(self, inputs):
"""forward"""
user_feature = self.user_encode(inputs)
item_embedding = self._build_embeddings(inputs, self.item_slot_names)
item_fc = self.item_fc_op(item_embedding)
pos = fluid_sequence_get_pos(item_fc)
pos_embed = self.dict_data_embed_op['pos'](pos)
gru_input = self.item_gru_fc_op(layers.concat([item_fc, pos_embed], 1))
item_gru = self.item_gru_op(gru_input, h_0=user_feature)
click_prob = self.out_fc2_op(self.out_fc1_op(item_gru))
return click_prob
def forward(self, inputs):
"""forward"""
user_feature = self.user_encode(inputs)
item_embedding = self._build_embeddings(inputs, self.item_slot_names)
item_fc = self.item_fc_op(item_embedding)
pos = fluid_sequence_get_pos(item_fc)
pos_embed = self.dict_data_embed_op['pos'](pos)
item_concat = layers.concat([
item_fc, pos_embed,
layers.sequence_expand_as(user_feature, item_fc)
], 1)
item_concat_fc = self.item_concat_fc_op(item_concat)
click_prob = self.out_fc2_op(self.out_fc1_op(item_concat_fc))
return click_prob
def forward(self, inputs, mode):
"""forward"""
# encode
user_embedding = self._build_embeddings(inputs, self.user_slot_names)
user_feature = self.user_feature_fc_op(user_embedding)
# item embed + pos embed
item_embedding = self._build_embeddings(inputs, self.item_slot_names)
item_fc = self.item_fc_op(item_embedding)
pos = fluid_sequence_get_pos(item_fc)
pos_embed = self.dict_data_embed_op['pos'](pos)
# item gru
gru_input = self.item_gru_fc_op(layers.concat([item_fc, pos_embed], 1))
item_gru_forward = self.item_gru_forward_op(gru_input,
h_0=user_feature)
item_gru_backward = self.item_gru_backward_op(gru_input,
h_0=user_feature)
item_gru = layers.concat([item_gru_forward, item_gru_backward], axis=1)
click_prob = self.out_click_fc2_op(self.out_click_fc1_op(item_gru))
return click_prob
def forward(self, inputs, output_type):
"""forward"""
decode_len = inputs['decode_len']
user_feature = self.user_encode(inputs)
item_embedding = self._build_embeddings(inputs, self.item_slot_names)
item_fc = self.item_fc_op(item_embedding)
atten_item_fc = self.atten_item_fc_op(item_embedding)
cand_encoding = self.candidate_encode(item_fc)
pos = fluid_sequence_get_pos(item_fc)
pos_embed = self.dict_data_embed_op['pos'](pos)
init_hidden = self.candidate_encode_fc_op(
layers.concat([user_feature, cand_encoding], 1))
item_Q = self.train_rnn(item_fc,
atten_item_fc,
init_hidden,
pos,
pos_embed,
output_type=output_type)
item_Q = self._cut_by_decode_len(layers.lod_reset(item_Q, item_fc),
decode_len)
return item_Q