def network():
# define encoder input data
enc_token_ids = fluid.layers.data(name="enc_token_ids", shape=[None, SEQ_MAX_LEN], dtype='int64')
enc_segment_ids = fluid.layers.data(name="enc_segment_ids", shape=[None, SEQ_MAX_LEN], dtype='int64')
enc_pos_ids = fluid.layers.data(name="enc_pos_ids", shape=[None, SEQ_MAX_LEN], dtype='int64')
enc_input_length = fluid.layers.data(name='enc_input_length', shape=[None, SEQ_MAX_LEN, 1], dtype='int64')
# define decoder input data
dec_token_ids = fluid.layers.data(name="token_ids", shape=[None, SEQ_MAX_LEN], dtype='int64')
dec_segment_ids = fluid.layers.data(name="segment_ids", shape=[None, SEQ_MAX_LEN], dtype='int64')
dec_pos_ids = fluid.layers.data(name="pos_ids", shape=[None, SEQ_MAX_LEN], dtype='int64')
dec_enc_slf_attn = fluid.layers.data(name='enc_slf_attn', shape=[None, SEQ_MAX_LEN, SEQ_MAX_LEN], dtype='int64')
# task label
dec_lm_label_mat = fluid.layers.data(name='lm_label_mat', shape=[None, SEQ_MAX_LEN], dtype='int64')
dec_lm_pos_mask = fluid.layers.data(name='lm_pos_mask', shape=[None, SEQ_MAX_LEN, SEQ_MAX_LEN], dtype='int64')
dec_lm_pos_len = fluid.layers.data(name='lm_pos_len', shape=[None, 1], dtype='int64')
goal_type_pos = fluid.layers.data(name="goal_type_pos", shape=[None, 2], dtype='int64')
goal_type_label = fluid.layers.data(name="goal_type_label", shape=[None], dtype='int64')
# enc_dec_mask
enc_dec_mask = fluid.layers.data(name='enc_dec_mask', shape=[None, SEQ_MAX_LEN, SEQ_MAX_LEN], dtype='int64')
# network
encode = Encoder(enc_token_ids, enc_pos_ids, enc_segment_ids,
enc_input_length, config)
enc_output = encode.get_sequence_output()
decode = Decoder(dec_token_ids, dec_pos_ids, dec_segment_ids,
dec_enc_slf_attn, config=config, enc_input=enc_output, enc_input_mask=enc_dec_mask)
loss, goal_type_acc = decode.pretrain(goal_type_pos, goal_type_label,
dec_lm_label_mat, dec_lm_pos_mask, dec_lm_pos_len)
input_name_list = [
enc_token_ids.name,
enc_segment_ids.name,
enc_pos_ids.name,
enc_input_length.name,
dec_token_ids.name,
dec_segment_ids.name,
dec_pos_ids.name,
dec_enc_slf_attn.name,
enc_dec_mask.name,
dec_lm_label_mat.name,
dec_lm_pos_mask.name,
dec_lm_pos_len.name,
goal_type_pos.name,
goal_type_label.name
]
return loss.name, goal_type_acc.name, input_name_list
# task label
dec_lm_label_mat = fluid.layers.data(name='lm_label_mat', shape=[None, SEQ_MAX_LEN], dtype='int64')
dec_lm_pos_mask = fluid.layers.data(name='lm_pos_mask', shape=[None, SEQ_MAX_LEN, SEQ_MAX_LEN], dtype='int64')
dec_lm_pos_len = fluid.layers.data(name='lm_pos_len', shape=[None, 1], dtype='int64')
goal_type_pos = fluid.layers.data(name="goal_type_pos", shape=[None, 2], dtype='int64')
goal_type_label = fluid.layers.data(name="goal_type_label", shape=[None], dtype='int64')
# enc_dec_mask
enc_dec_mask = fluid.layers.data(name='enc_dec_mask', shape=[None, SEQ_MAX_LEN, SEQ_MAX_LEN], dtype='int64')
# network
encode = Encoder(enc_token_ids, enc_pos_ids, enc_segment_ids,
enc_input_length, config)
enc_output = encode.get_sequence_output()
decode = Decoder(dec_token_ids, dec_pos_ids, dec_segment_ids,
dec_enc_slf_attn, config=config, enc_input=enc_output, enc_input_mask=enc_dec_mask)
dec_output = decode.get_sequence_output()
loss, goal_type_acc = decode.pretrain(goal_type_pos, goal_type_label,
dec_lm_label_mat, dec_lm_pos_mask, dec_lm_pos_len)
# loss
adam = fluid.optimizer.AdamOptimizer()
adam.minimize(loss)
# define executor
place = fluid.CUDAPlace(0) if USE_CUDA else fluid.CPUPlace()
