def build_model3(self):
assert not conf.hidden_size % 2 # 被2整除
self.uttn_enc_hidden_size = conf.hidden_size // 2
batch_size, num_turns, length = shape_list(self.multi_s1)
# embedding
# [batch,len,turn,embed]
multi_s1_embed, _ = embedding(tf.expand_dims(self.multi_s1, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=conf.pretrain_emb)
# [batch,len,embed]
s2_embed, _ = embedding(tf.expand_dims(self.s2, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=conf.pretrain_emb)
# uttn encoder
uttn_input = tf.reshape(multi_s1_embed, [-1, length, conf.embed_size]) # [batch*turn,len,embed]
uttn_mask = mask_nonpad_from_embedding(uttn_input) # [batch*turn,len] 1 for nonpad; 0 for pad
uttn_seqlen = tf.cast(tf.reduce_sum(uttn_mask, axis=-1), tf.int32) # [batch*turn]
# uttn-gru
self.encoder_uttn_rnn = Bi_RNN(cell_name='GRUCell', name='uttn_enc', hidden_size=self.uttn_enc_hidden_size, dropout_rate=self.dropout_rate)
_, uttn_embed = self.encoder_uttn_rnn(uttn_input, uttn_seqlen) # [batch*turn,len,2hid] [batch*turn,2hid]
uttn_embed = tf.reshape(uttn_embed, [batch_size, num_turns, self.uttn_enc_hidden_size * 2]) # [batch,turn,2hid] # 之后turn相当于len
# transformer ctx encoder
encoder_valid_mask = mask_nonpad_from_embedding(uttn_embed) # [batch,turn] 1 for nonpad; 0 for pad
encoder_input = add_timing_signal_1d(uttn_embed) # add position embedding
encoder_input = tf.layers.dropout(encoder_input, rate=self.dropout_rate) # dropout
encoder_output = transformer_encoder(encoder_input, encoder_valid_mask,
hidden_size=conf.hidden_size,
filter_size=conf.hidden_size * 4,
num_heads=conf.num_heads,
num_encoder_layers=conf.num_encoder_layers,
dropout=self.dropout_rate,
attention_dropout=self.dropout_rate,
relu_dropout=self.dropout_rate,
)
# transformer decoder
decoder_input = s2_embed
decoder_valid_mask = mask_nonpad_from_embedding(decoder_input) # [batch,len] 1 for nonpad; 0 for pad
decoder_input = shift_right(decoder_input) # 用pad当做eos
decoder_input = add_timing_signal_1d(decoder_input)
decoder_input = tf.layers.dropout(decoder_input, rate=self.dropout_rate) # dropout
decoder_output = transformer_decoder(decoder_input, encoder_output, decoder_valid_mask, encoder_valid_mask,
cache=None,
hidden_size=conf.hidden_size,
filter_size=conf.hidden_size * 4,
num_heads=conf.num_heads,
num_decoder_layers=conf.num_decoder_layers,
dropout=self.dropout_rate,
attention_dropout=self.dropout_rate,
relu_dropout=self.dropout_rate,
)
logits = proj_logits(decoder_output, conf.embed_size, conf.vocab_size, name='share_embedding')
onehot_s2 = tf.one_hot(self.s2, depth=conf.vocab_size) # [batch,len,vocab]
xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=onehot_s2) # [batch,len]
weights = tf.to_float(tf.not_equal(self.s2, 0)) # [batch,len] 1 for nonpad; 0 for pad
loss_num = xentropy * weights # [batch,len]
loss_den = weights # [batch,len]
loss = tf.reduce_sum(loss_num) / tf.reduce_sum(loss_den) # scalar
self.loss = loss
# transformer decoder infer
# 放在cache里面的在后面symbols_to_logits_fn函数中都会变成batch * beam
# 初始化缓存
cache = {
'layer_%d' % layer: {
# 用以缓存decoder过程前面已计算的k,v
'k': split_heads(tf.zeros([batch_size, 0, conf.embed_size]), conf.num_heads),
'v': split_heads(tf.zeros([batch_size, 0, conf.embed_size]), conf.num_heads)
} for layer in range(conf.num_decoder_layers)
}
for layer in range(conf.num_decoder_layers):
# 对于decoder每层均需与encoder顶层隐状态计算attention,相应均有特定的k,v可缓存
layer_name = 'layer_%d' % layer
with tf.variable_scope('decoder/%s/encdec_attention/multihead_attention' % layer_name):
k_encdec = tf.layers.dense(encoder_output, conf.embed_size, use_bias=False, name='k', reuse=tf.AUTO_REUSE)
k_encdec = split_heads(k_encdec, conf.num_heads)
v_encdec = tf.layers.dense(encoder_output, conf.embed_size, use_bias=False, name='v', reuse=tf.AUTO_REUSE)
v_encdec = split_heads(v_encdec, conf.num_heads)
cache[layer_name]['k_encdec'] = k_encdec
cache[layer_name]['v_encdec'] = v_encdec
cache['encoder_output'] = encoder_output
cache['encoder_mask'] = encoder_valid_mask
# position embedding
position_embedding = get_timing_signal_1d(conf.max_decode_len, conf.embed_size) # +eos [1,length+1,embed]
def symbols_to_logits_fn(ids, i, cache):
ids = ids[:, -1:] # [batch,1] 截取最后一个
target_embed, _ = embedding(tf.expand_dims(ids, axis=-1), conf.vocab_size, conf.embed_size, 'share_embedding', reuse=True) # [batch,1,hidden]
decoder_input = target_embed + position_embedding[:, i:i + 1, :] # [batch,1,hidden]
encoder_output = cache['encoder_output']
encoder_mask = cache['encoder_mask']
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
decoder_output = transformer_decoder(decoder_input, encoder_output, None, encoder_mask,
cache=cache, # 注意infer要cache
hidden_size=conf.embed_size,
filter_size=conf.embed_size * 4,
num_heads=6,
num_decoder_layers=6,
)
logits = proj_logits(decoder_output, conf.embed_size, conf.vocab_size, name='share_embedding') # [batch,1,vocab]
ret = tf.squeeze(logits, axis=1) # [batch,vocab]
return ret, cache
initial_ids = tf.zeros([batch_size], dtype=tf.int32) # <pad>为<sos>
def greedy_search_wrapper():
""" Greedy Search """
decoded_ids, scores = greedy_search(
symbols_to_logits_fn,
initial_ids,
conf.max_decode_len,
cache=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
def beam_search_wrapper():
""" Beam Search """
decoded_ids, scores = beam_search( # [batch,beam,len] [batch,beam]
symbols_to_logits_fn,
initial_ids,
conf.beam_size,
conf.max_decode_len,
conf.vocab_size,
alpha=0,
states=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
decoded_ids, scores = tf.cond(tf.equal(conf.beam_size, 1), greedy_search_wrapper, beam_search_wrapper)
self.decoded_ids = tf.identity(decoded_ids, name='decoded_ids') # [batch,beam/1,len]
self.scores = tf.identity(scores, name='scores') # [batch,beam/1]
self.global_step = tf.train.get_or_create_global_step()
self.optimizer = tf.train.AdamOptimizer(learning_rate=conf.lr)
self.train_op = self.optimizer.minimize(self.loss, global_step=self.global_step)
def build_model1(self):
self.uttn_enc_hidden_size = 256
self.ctx_enc_hidden_size = 256
batch_size, num_turns, length = shape_list(self.multi_s1)
# embedding
# [batch,len,turn,embed]
multi_s1_embed, _ = embedding(tf.expand_dims(self.multi_s1, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=conf.pretrain_emb)
# [batch,len,embed]
s2_embed, _ = embedding(tf.expand_dims(self.s2, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=conf.pretrain_emb)
# uttn encoder
uttn_input = tf.reshape(multi_s1_embed, [-1, length, conf.embed_size]) # [batch*turn,len,embed]
uttn_mask = mask_nonpad_from_embedding(uttn_input) # [batch*turn,len] 1 for nonpad; 0 for pad
uttn_seqlen = tf.cast(tf.reduce_sum(uttn_mask, axis=-1), tf.int32) # [batch*turn]
# uttn-gru
self.encoder_uttn_rnn = RNN(cell_name='GRUCell', name='uttn_enc', hidden_size=self.uttn_enc_hidden_size, dropout_rate=self.dropout_rate)
_, uttn_embed = self.encoder_uttn_rnn(uttn_input, uttn_seqlen) # [batch*turn,hid]
uttn_embed = tf.reshape(uttn_embed, [batch_size, num_turns, self.uttn_enc_hidden_size]) # [batch,turn,hid] # 之后turn相当于len
# ctx encoder
ctx_mask = mask_nonpad_from_embedding(uttn_embed) # [batch,turn] 1 for nonpad; 0 for pad
ctx_seqlen = tf.cast(tf.reduce_sum(ctx_mask, axis=-1), tf.int32) # [batch]
# ctx-gru
self.encoder_ctx_rnn = RNN(cell_name='GRUCell', name='ctx_enc', hidden_size=self.ctx_enc_hidden_size, dropout_rate=self.dropout_rate)
_, ctx_embed = self.encoder_ctx_rnn(uttn_embed, ctx_seqlen) # [batch,hid]
# rnn decoder train (no attention)
s2_mask = mask_nonpad_from_embedding(s2_embed) # [batch,len] 1 for nonpad; 0 for pad
s2_seqlen = tf.cast(tf.reduce_sum(s2_mask, axis=-1), tf.int32) # [batch]
decoder_input = shift_right(s2_embed) # 用pad当做eos
decoder_input = tf.layers.dropout(decoder_input, rate=self.dropout_rate) # dropout
# 输入拼上ctx
decoder_ctx = tf.tile(tf.expand_dims(ctx_embed, axis=1), [1, shape_list(decoder_input)[1], 1]) # [batch,len,hid]
decoder_input = tf.concat([decoder_input, decoder_ctx], axis=2)
self.decoder_rnn = RNN(cell_name='GRUCell', name='dec', hidden_size=conf.embed_size, dropout_rate=self.dropout_rate)
decoder_output, decoder_state = self.decoder_rnn(decoder_input, s2_seqlen)
logits = proj_logits(decoder_output, conf.embed_size, conf.vocab_size, name='share_embedding')
onehot_s2 = tf.one_hot(self.s2, depth=conf.vocab_size) # [batch,len,vocab]
xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=onehot_s2) # [batch,len]
weights = tf.to_float(tf.not_equal(self.s2, 0)) # [batch,len] 1 for nonpad; 0 for pad
loss_num = xentropy * weights # [batch,len]
loss_den = weights # [batch,len]
loss = tf.reduce_sum(loss_num) / tf.reduce_sum(loss_den) # scalar
self.loss = loss
# rnn decoder infer (no attention)
# 放在cache里面的在后面symbols_to_logits_fn函数中都会变成batch * beam
cache = {'state': self.decoder_rnn.cell.zero_state(batch_size, tf.float32), # [batch,hid]
'ctx': ctx_embed, # [batch,hid]
}
def symbols_to_logits_fn(ids, i, cache):
# ids [batch,length]
pred_target = ids[:, -1:] # [batch,1] 截取最后一个
target_embed, _ = embedding(tf.expand_dims(pred_target, axis=-1), conf.vocab_size, conf.embed_size, 'share_embedding') # [batch,1,embed]
decoder_input = tf.squeeze(target_embed, axis=1) # [batch,embed]
# 输入加上ctx
decoder_input = tf.concat([decoder_input, cache['ctx']], axis=-1)
# run rnn
decoder_output, cache['state'] = self.decoder_rnn.one_step(decoder_input, cache['state'])
logits = proj_logits(decoder_output, conf.embed_size, conf.vocab_size, name='share_embedding')
return logits, cache
initial_ids = tf.zeros([batch_size], dtype=tf.int32) # <pad>为<sos>
def greedy_search_wrapper():
""" Greedy Search """
decoded_ids, scores = greedy_search(
symbols_to_logits_fn,
initial_ids,
conf.max_decode_len,
cache=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
def beam_search_wrapper():
""" Beam Search """
decoded_ids, scores = beam_search( # [batch,beam,len] [batch,beam]
symbols_to_logits_fn,
initial_ids,
conf.beam_size,
conf.max_decode_len,
conf.vocab_size,
alpha=0,
states=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
decoded_ids, scores = tf.cond(tf.equal(conf.beam_size, 1), greedy_search_wrapper, beam_search_wrapper)
self.decoded_ids = tf.identity(decoded_ids, name='decoded_ids') # [batch,beam/1,len]
self.scores = tf.identity(scores, name='scores') # [batch,beam/1]
self.global_step = tf.train.get_or_create_global_step()
self.optimizer = tf.train.AdamOptimizer(learning_rate=conf.lr)
self.train_op = self.optimizer.minimize(self.loss, global_step=self.global_step)
def symbols_to_logits_fn(ids, i, cache):
# ids [batch,length]
pred_target = ids[:, -1:] # [batch,1] 截取最后一个
target_embed, _ = embedding(tf.expand_dims(pred_target, axis=-1), conf.vocab_size, conf.embed_size, 'share_embedding') # [batch,1,embed]
decoder_input = tf.squeeze(target_embed, axis=1) # [batch,embed]
dec_rnn_state = cache['dec_rnn_state']
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
# 内层循环是对于解码每一步 ctx_rnn上turn的每一步，最终生成ctx序列 turn_step
def inner_loop_condition(turn_step, *_):
return tf.less(turn_step, tf.reduce_max(ctx_seqlen))
def inner_loop_body(turn_step, ctx_rnn_state, ctx_rnn_output):
# 根据si, ctx_init_state 递归计算多个turn的句子ctx
q_antecedent = tf.concat([ctx_rnn_state, dec_rnn_state], axis=-1) # [batch, h]
q_antecedent = tf.tile(tf.expand_dims(q_antecedent, 1), [1, length, 1]) # [batch,len,h]
# 抽取每个batch的第i个turn
# sent_repre [batch,turn,len,h]
q_antecedent = tf.concat([cache['uttn_repre'][:, turn_step, :, :], q_antecedent], -1) # [batch,len,h]
uttn_mask_in_turn = tf.reshape(uttn_mask, [batch_size, num_turns, length])[:, turn_step, :] # [batch,len]
# word-level-attn
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='word_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='word_level_attn/layer2') # [batch,len,1]
energy = tf.squeeze(energy, -1) + (1. - uttn_mask_in_turn) * -1e9
alpha = tf.nn.softmax(energy) # [batch,len]
r_in_turn = tf.reduce_sum(tf.expand_dims(alpha, -1) * cache['uttn_repre'][:, turn_step, :, :], 1) # [batch,h]
ctx_rnn_output_, ctx_rnn_state = self.encoder_ctx_rnn.one_step(r_in_turn, ctx_rnn_state)
# attch
ctx_rnn_output = tf.concat([ctx_rnn_output, tf.expand_dims(ctx_rnn_output_, 1)], 1) # [batch,turn,h]
return turn_step + 1, ctx_rnn_state, ctx_rnn_output
# start inner loop
final_turn_step, final_state, ctx_rnn_output = tf.while_loop(
inner_loop_condition,
inner_loop_body,
loop_vars=[tf.constant(0, dtype=tf.int32),
cache['ctx_rnn_state'],
tf.zeros([shape_list(cache['ctx_rnn_state'])[0], 0, self.ctx_enc_hidden_size])],
shape_invariants=[
tf.TensorShape([]),
nest.map_structure(get_state_shape_invariants, init_ctx_encoder_state),
tf.TensorShape([None, None, self.ctx_enc_hidden_size]),
])
# ctx_rnn_output # [batch,turn,h]
# dec_rnn_state # [batch,h]
# ctx-level-attn
# q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, num_turns, 1]) # [batch,turn,h]
# 这样只拿当前batch中的尽可能小的turns数量而不是固定turn
q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, shape_list(ctx_rnn_output)[1], 1]) # [batch,turn,h]
q_antecedent = tf.concat([q_antecedent, ctx_rnn_output], 2) # [batch,turn,h]
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='ctx_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='ctx_level_attn/layer2') # [batch,turn,1]
energy = tf.squeeze(energy, -1) + (1. - ctx_mask) * -1e9 # [batch,turn]
alpha = tf.nn.softmax(energy) # [batch,turn]
ctx_input_in_dec = tf.reduce_sum(tf.expand_dims(alpha, -1) * ctx_rnn_output, 1) # [batch,h]
dec_rnn_input = tf.concat([ctx_input_in_dec, decoder_input], -1) # [batch,h]
dec_rnn_output_, dec_rnn_state = self.decoder_rnn.one_step(dec_rnn_input, dec_rnn_state)
cache['dec_rnn_state'] = dec_rnn_state
logits = proj_logits(dec_rnn_output_, conf.embed_size, conf.vocab_size, name='share_embedding')
return logits, cache
def loop_body(time_step, dec_rnn_state, dec_rnn_output):
# cal decoder ctx
# word level attention
# 内层循环是对于解码每一步 ctx_rnn上turn的每一步，最终生成ctx序列 turn_step
def inner_loop_condition(turn_step, *_):
return tf.less(turn_step, tf.reduce_max(ctx_seqlen))
def inner_loop_body(turn_step, ctx_rnn_state, ctx_rnn_output):
# 根据si, ctx_init_state 递归计算多个turn的句子ctx
q_antecedent = tf.concat([ctx_rnn_state, dec_rnn_state], axis=-1) # [batch, h]
q_antecedent = tf.tile(tf.expand_dims(q_antecedent, 1), [1, length, 1]) # [batch,len,h]
# 抽取每个batch的第i个turn
# sent_repre [batch,turn,len,h]
q_antecedent = tf.concat([uttn_repre[:, turn_step, :, :], q_antecedent], -1) # [batch,len,h]
uttn_mask_in_turn = tf.reshape(uttn_mask, [batch_size, num_turns, length])[:, turn_step, :] # [batch,len]
# word-level-attn
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='word_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='word_level_attn/layer2') # [batch,len,1]
energy = tf.squeeze(energy, -1) + (1. - uttn_mask_in_turn) * -1e9
alpha = tf.nn.softmax(energy) # [batch,len]
r_in_turn = tf.reduce_sum(tf.expand_dims(alpha, -1) * uttn_repre[:, turn_step, :, :], 1) # [batch,h]
ctx_rnn_output_, ctx_rnn_state = self.encoder_ctx_rnn.one_step(r_in_turn, ctx_rnn_state)
# attch
ctx_rnn_output = tf.concat([ctx_rnn_output, tf.expand_dims(ctx_rnn_output_, 1)], 1) # [batch,turn,h]
return turn_step + 1, ctx_rnn_state, ctx_rnn_output
# start inner loop
final_turn_step, final_state, ctx_rnn_output = tf.while_loop(
inner_loop_condition,
inner_loop_body,
loop_vars=[tf.constant(0, dtype=tf.int32),
init_ctx_encoder_state,
tf.zeros([batch_size, 0, self.ctx_enc_hidden_size])],
shape_invariants=[
tf.TensorShape([]),
nest.map_structure(get_state_shape_invariants, init_ctx_encoder_state),
tf.TensorShape([None, None, self.ctx_enc_hidden_size]),
])
# ctx_rnn_output # [batch,turn,h]
# dec_rnn_state # [batch,h]
# ctx-level-attn
# q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, num_turns, 1]) # [batch,turn,h]
# 这样只拿当前batch中的尽可能小的turns数量而不是固定turn
q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, shape_list(ctx_rnn_output)[1], 1]) # [batch,turn,h]
q_antecedent = tf.concat([q_antecedent, ctx_rnn_output], 2) # [batch,turn,h]
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='ctx_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='ctx_level_attn/layer2') # [batch,turn,1]
energy = tf.squeeze(energy, -1) + (1. - ctx_mask) * -1e9 # [batch,turn]
alpha = tf.nn.softmax(energy) # [batch,turn]
ctx_input_in_dec = tf.reduce_sum(tf.expand_dims(alpha, -1) * ctx_rnn_output, 1) # [batch,h]
dec_rnn_input = tf.concat([ctx_input_in_dec, decoder_input[:, time_step, :]], -1) # [batch,h]
dec_rnn_output_, dec_rnn_state = self.decoder_rnn.one_step(dec_rnn_input, dec_rnn_state)
dec_rnn_output = tf.concat([dec_rnn_output, tf.expand_dims(dec_rnn_output_, 1)], 1)
return time_step + 1, dec_rnn_state, dec_rnn_output
def build_model2(self):
self.uttn_enc_hidden_size = 256
self.ctx_enc_hidden_size = 256
batch_size, num_turns, length = shape_list(self.multi_s1)
# embedding
# [batch,len,turn,embed]
multi_s1_embed, _ = embedding(tf.expand_dims(self.multi_s1, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=conf.pretrain_emb)
# [batch,len,embed]
s2_embed, _ = embedding(tf.expand_dims(self.s2, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=conf.pretrain_emb)
# uttn encoder
uttn_input = tf.reshape(multi_s1_embed, [-1, length, conf.embed_size]) # [batch*turn,len,embed]
uttn_mask = mask_nonpad_from_embedding(uttn_input) # [batch*turn,len] 1 for nonpad; 0 for pad
uttn_seqlen = tf.cast(tf.reduce_sum(uttn_mask, axis=-1), tf.int32) # [batch*turn]
# uttn-gru
self.encoder_uttn_rnn = Bi_RNN(cell_name='GRUCell', name='uttn_enc', hidden_size=self.uttn_enc_hidden_size, dropout_rate=self.dropout_rate)
uttn_repre, uttn_embed = self.encoder_uttn_rnn(uttn_input, uttn_seqlen) # [batch*turn,len,2hid] [batch*turn,2hid]
uttn_embed = tf.reshape(uttn_embed, [batch_size, num_turns, self.uttn_enc_hidden_size * 2]) # [batch,turn,2hid]
uttn_repre = tf.reshape(uttn_repre, [batch_size, num_turns, length, self.uttn_enc_hidden_size * 2]) # [batch,turn,len,2hid]
# ctx encoder
ctx_mask = mask_nonpad_from_embedding(uttn_embed) # [batch,turn] 1 for nonpad; 0 for pad
ctx_seqlen = tf.cast(tf.reduce_sum(ctx_mask, axis=-1), tf.int32) # [batch]
# reverse turn
uttn_repre = tf.reverse_sequence(uttn_repre, seq_lengths=ctx_seqlen, seq_axis=1, batch_axis=0)
# ctx-gru
self.encoder_ctx_rnn = RNN(cell_name='GRUCell', name='ctx_enc', hidden_size=self.ctx_enc_hidden_size, dropout_rate=self.dropout_rate)
init_ctx_encoder_state = self.encoder_ctx_rnn.cell.zero_state(batch_size, tf.float32)
# rnn decoder train
s2_mask = mask_nonpad_from_embedding(s2_embed) # [batch,len] 1 for nonpad; 0 for pad
s2_seqlen = tf.cast(tf.reduce_sum(s2_mask, axis=-1), tf.int32) # [batch]
decoder_input = shift_right(s2_embed) # 用pad当做eos
decoder_input = tf.layers.dropout(decoder_input, rate=self.dropout_rate) # dropout
self.decoder_rnn = RNN(cell_name='GRUCell', name='dec', hidden_size=conf.embed_size, dropout_rate=self.dropout_rate)
init_decoder_state = self.decoder_rnn.cell.zero_state(batch_size, tf.float32)
# 两重循环 -_-
# 外层循环是decoder解码的每一步 time_step
def loop_condition(time_step, *_):
return tf.less(time_step, tf.reduce_max(s2_seqlen))
def loop_body(time_step, dec_rnn_state, dec_rnn_output):
# cal decoder ctx
# word level attention
# 内层循环是对于解码每一步 ctx_rnn上turn的每一步，最终生成ctx序列 turn_step
def inner_loop_condition(turn_step, *_):
return tf.less(turn_step, tf.reduce_max(ctx_seqlen))
def inner_loop_body(turn_step, ctx_rnn_state, ctx_rnn_output):
# 根据si, ctx_init_state 递归计算多个turn的句子ctx
q_antecedent = tf.concat([ctx_rnn_state, dec_rnn_state], axis=-1) # [batch, h]
q_antecedent = tf.tile(tf.expand_dims(q_antecedent, 1), [1, length, 1]) # [batch,len,h]
# 抽取每个batch的第i个turn
# sent_repre [batch,turn,len,h]
q_antecedent = tf.concat([uttn_repre[:, turn_step, :, :], q_antecedent], -1) # [batch,len,h]
uttn_mask_in_turn = tf.reshape(uttn_mask, [batch_size, num_turns, length])[:, turn_step, :] # [batch,len]
# word-level-attn
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='word_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='word_level_attn/layer2') # [batch,len,1]
energy = tf.squeeze(energy, -1) + (1. - uttn_mask_in_turn) * -1e9
alpha = tf.nn.softmax(energy) # [batch,len]
r_in_turn = tf.reduce_sum(tf.expand_dims(alpha, -1) * uttn_repre[:, turn_step, :, :], 1) # [batch,h]
ctx_rnn_output_, ctx_rnn_state = self.encoder_ctx_rnn.one_step(r_in_turn, ctx_rnn_state)
# attch
ctx_rnn_output = tf.concat([ctx_rnn_output, tf.expand_dims(ctx_rnn_output_, 1)], 1) # [batch,turn,h]
return turn_step + 1, ctx_rnn_state, ctx_rnn_output
# start inner loop
final_turn_step, final_state, ctx_rnn_output = tf.while_loop(
inner_loop_condition,
inner_loop_body,
loop_vars=[tf.constant(0, dtype=tf.int32),
init_ctx_encoder_state,
tf.zeros([batch_size, 0, self.ctx_enc_hidden_size])],
shape_invariants=[
tf.TensorShape([]),
nest.map_structure(get_state_shape_invariants, init_ctx_encoder_state),
tf.TensorShape([None, None, self.ctx_enc_hidden_size]),
])
# ctx_rnn_output # [batch,turn,h]
# dec_rnn_state # [batch,h]
# ctx-level-attn
# q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, num_turns, 1]) # [batch,turn,h]
# 这样只拿当前batch中的尽可能小的turns数量而不是固定turn
q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, shape_list(ctx_rnn_output)[1], 1]) # [batch,turn,h]
q_antecedent = tf.concat([q_antecedent, ctx_rnn_output], 2) # [batch,turn,h]
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='ctx_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='ctx_level_attn/layer2') # [batch,turn,1]
energy = tf.squeeze(energy, -1) + (1. - ctx_mask) * -1e9 # [batch,turn]
alpha = tf.nn.softmax(energy) # [batch,turn]
ctx_input_in_dec = tf.reduce_sum(tf.expand_dims(alpha, -1) * ctx_rnn_output, 1) # [batch,h]
dec_rnn_input = tf.concat([ctx_input_in_dec, decoder_input[:, time_step, :]], -1) # [batch,h]
dec_rnn_output_, dec_rnn_state = self.decoder_rnn.one_step(dec_rnn_input, dec_rnn_state)
dec_rnn_output = tf.concat([dec_rnn_output, tf.expand_dims(dec_rnn_output_, 1)], 1)
return time_step + 1, dec_rnn_state, dec_rnn_output
# start outer loop
final_time_step, final_state, dec_rnn_output = tf.while_loop(
loop_condition,
loop_body,
loop_vars=[tf.constant(0, dtype=tf.int32),
init_decoder_state,
tf.zeros([batch_size, 0, conf.embed_size])],
shape_invariants=[
tf.TensorShape([]),
nest.map_structure(get_state_shape_invariants, init_decoder_state),
tf.TensorShape([None, None, conf.embed_size]),
])
decoder_output = dec_rnn_output
logits = proj_logits(decoder_output, conf.embed_size, conf.vocab_size, name='share_embedding')
onehot_s2 = tf.one_hot(self.s2, depth=conf.vocab_size) # [batch,len,vocab]
xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=onehot_s2) # [batch,len]
weights = tf.to_float(tf.not_equal(self.s2, 0)) # [batch,len] 1 for nonpad; 0 for pad
loss_num = xentropy * weights # [batch,len]
loss_den = weights # [batch,len]
loss = tf.reduce_sum(loss_num) / tf.reduce_sum(loss_den) # scalar
self.loss = loss
# rnn decoder infer
# 放在cache里面的在后面symbols_to_logits_fn函数中都会变成batch * beam
cache = {'dec_rnn_state': self.decoder_rnn.cell.zero_state(batch_size, tf.float32), # [batch,hid]
'ctx_rnn_state': self.encoder_ctx_rnn.cell.zero_state(batch_size, tf.float32), # [batch,hid]
'uttn_repre': uttn_repre # [batch,turn,len,2hid]
}
def symbols_to_logits_fn(ids, i, cache):
# ids [batch,length]
pred_target = ids[:, -1:] # [batch,1] 截取最后一个
target_embed, _ = embedding(tf.expand_dims(pred_target, axis=-1), conf.vocab_size, conf.embed_size, 'share_embedding') # [batch,1,embed]
decoder_input = tf.squeeze(target_embed, axis=1) # [batch,embed]
dec_rnn_state = cache['dec_rnn_state']
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
# 内层循环是对于解码每一步 ctx_rnn上turn的每一步，最终生成ctx序列 turn_step
def inner_loop_condition(turn_step, *_):
return tf.less(turn_step, tf.reduce_max(ctx_seqlen))
def inner_loop_body(turn_step, ctx_rnn_state, ctx_rnn_output):
# 根据si, ctx_init_state 递归计算多个turn的句子ctx
q_antecedent = tf.concat([ctx_rnn_state, dec_rnn_state], axis=-1) # [batch, h]
q_antecedent = tf.tile(tf.expand_dims(q_antecedent, 1), [1, length, 1]) # [batch,len,h]
# 抽取每个batch的第i个turn
# sent_repre [batch,turn,len,h]
q_antecedent = tf.concat([cache['uttn_repre'][:, turn_step, :, :], q_antecedent], -1) # [batch,len,h]
uttn_mask_in_turn = tf.reshape(uttn_mask, [batch_size, num_turns, length])[:, turn_step, :] # [batch,len]
# word-level-attn
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='word_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='word_level_attn/layer2') # [batch,len,1]
energy = tf.squeeze(energy, -1) + (1. - uttn_mask_in_turn) * -1e9
alpha = tf.nn.softmax(energy) # [batch,len]
r_in_turn = tf.reduce_sum(tf.expand_dims(alpha, -1) * cache['uttn_repre'][:, turn_step, :, :], 1) # [batch,h]
ctx_rnn_output_, ctx_rnn_state = self.encoder_ctx_rnn.one_step(r_in_turn, ctx_rnn_state)
# attch
ctx_rnn_output = tf.concat([ctx_rnn_output, tf.expand_dims(ctx_rnn_output_, 1)], 1) # [batch,turn,h]
return turn_step + 1, ctx_rnn_state, ctx_rnn_output
# start inner loop
final_turn_step, final_state, ctx_rnn_output = tf.while_loop(
inner_loop_condition,
inner_loop_body,
loop_vars=[tf.constant(0, dtype=tf.int32),
cache['ctx_rnn_state'],
tf.zeros([shape_list(cache['ctx_rnn_state'])[0], 0, self.ctx_enc_hidden_size])],
shape_invariants=[
tf.TensorShape([]),
nest.map_structure(get_state_shape_invariants, init_ctx_encoder_state),
tf.TensorShape([None, None, self.ctx_enc_hidden_size]),
])
# ctx_rnn_output # [batch,turn,h]
# dec_rnn_state # [batch,h]
# ctx-level-attn
# q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, num_turns, 1]) # [batch,turn,h]
# 这样只拿当前batch中的尽可能小的turns数量而不是固定turn
q_antecedent = tf.tile(tf.expand_dims(dec_rnn_state, axis=1), [1, shape_list(ctx_rnn_output)[1], 1]) # [batch,turn,h]
q_antecedent = tf.concat([q_antecedent, ctx_rnn_output], 2) # [batch,turn,h]
h = tf.layers.dense(q_antecedent, 128, activation=tf.nn.tanh, use_bias=True, name='ctx_level_attn/layer1')
energy = tf.layers.dense(h, 1, use_bias=True, name='ctx_level_attn/layer2') # [batch,turn,1]
energy = tf.squeeze(energy, -1) + (1. - ctx_mask) * -1e9 # [batch,turn]
alpha = tf.nn.softmax(energy) # [batch,turn]
ctx_input_in_dec = tf.reduce_sum(tf.expand_dims(alpha, -1) * ctx_rnn_output, 1) # [batch,h]
dec_rnn_input = tf.concat([ctx_input_in_dec, decoder_input], -1) # [batch,h]
dec_rnn_output_, dec_rnn_state = self.decoder_rnn.one_step(dec_rnn_input, dec_rnn_state)
cache['dec_rnn_state'] = dec_rnn_state
logits = proj_logits(dec_rnn_output_, conf.embed_size, conf.vocab_size, name='share_embedding')
return logits, cache
initial_ids = tf.zeros([batch_size], dtype=tf.int32) # <pad>为<sos>
def greedy_search_wrapper():
""" Greedy Search """
decoded_ids, scores = greedy_search(
symbols_to_logits_fn,
initial_ids,
conf.max_decode_len,
cache=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
def beam_search_wrapper():
""" Beam Search """
decoded_ids, scores = beam_search( # [batch,beam,len] [batch,beam]
symbols_to_logits_fn,
initial_ids,
conf.beam_size,
conf.max_decode_len,
conf.vocab_size,
alpha=0,
states=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
decoded_ids, scores = tf.cond(tf.equal(conf.beam_size, 1), greedy_search_wrapper, beam_search_wrapper)
self.decoded_ids = tf.identity(decoded_ids, name='decoded_ids') # [batch,beam/1,len]
self.scores = tf.identity(scores, name='scores') # [batch,beam/1]
self.global_step = tf.train.get_or_create_global_step()
self.optimizer = tf.train.AdamOptimizer(learning_rate=conf.lr)
self.train_op = self.optimizer.minimize(self.loss, global_step=self.global_step)
def build_model3(self):
# biGRU encoder + bah_attn + GRU decoder
# embedding
# [batch,len,embed]
# pretrained_word_embeddings = np.load(f'{curr_dir}/pretrain_emb_300.npy')
pretrained_word_embeddings = None
s1_embed, _ = embedding(tf.expand_dims(self.s1, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=pretrained_word_embeddings)
s1_mask = mask_nonpad_from_embedding(s1_embed) # [batch,len] 1 for nonpad; 0 for pad
s1_seqlen = tf.cast(tf.reduce_sum(s1_mask, axis=-1), tf.int32) # [batch]
# encoder
encoder_input = s1_embed
encoder_input = tf.layers.dropout(encoder_input, rate=self.dropout_rate) # dropout
with tf.variable_scope('birnn_encoder'):
self.bilstm_encoder1 = Bi_RNN(cell_name='GRUCell', hidden_size=conf.hidden_size, dropout_rate=self.dropout_rate)
encoder_output, _ = self.bilstm_encoder1(encoder_input, s1_seqlen) # [batch,len,2hid]
batch_size = shape_list(encoder_input)[0]
# decoder
decoder_rnn = getattr(tf.nn.rnn_cell, 'GRUCell')(conf.hidden_size) # GRUCell/LSTMCell
encdec_atten = EncDecAttention(encoder_output, s1_seqlen, conf.hidden_size)
s2_embed, _ = embedding(tf.expand_dims(self.s2, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=pretrained_word_embeddings)
s2_mask = mask_nonpad_from_embedding(s2_embed) # [batch,len] 1 for nonpad; 0 for pad
s2_seqlen = tf.cast(tf.reduce_sum(s2_mask, -1), tf.int32) # [batch]
decoder_input = s2_embed
decoder_input = shift_right(decoder_input) # 用pad当做eos
decoder_input = tf.layers.dropout(decoder_input, rate=self.dropout_rate) # dropout
init_decoder_state = decoder_rnn.zero_state(batch_size, tf.float32)
# init_decoder_state = tf.nn.rnn_cell.LSTMStateTuple(c, h)
time_step = tf.constant(0, dtype=tf.int32)
rnn_output = tf.zeros([batch_size, 0, conf.hidden_size])
context_output = tf.zeros([batch_size, 0, conf.hidden_size * 2]) # 注意力
def loop_condition(time_step, *_):
return tf.less(time_step, tf.reduce_max(s2_seqlen))
def loop_body(time_step, prev_rnn_state, rnn_output, context_output):
# attention
s = prev_rnn_state if isinstance(decoder_rnn, tf.nn.rnn_cell.GRUCell) else prev_rnn_state.h
context = encdec_atten(s) # [batch,hidden]
context_output = tf.concat([context_output, tf.expand_dims(context, axis=1)], axis=1)
# construct rnn input
rnn_input = tf.concat([decoder_input[:, time_step, :], context], axis=-1) # [batch,hidden+] use attention
# rnn_input = decoder_input[:, time_step, :] # [batch,hidden] not use attention
# run rnn
current_output, rnn_state = decoder_rnn(rnn_input, prev_rnn_state)
# append to output bucket via length dim
rnn_output = tf.concat([rnn_output, tf.expand_dims(current_output, axis=1)], axis=1)
return time_step + 1, rnn_state, rnn_output, context_output
# start loop
final_time_step, final_state, rnn_output, context_output = tf.while_loop(
loop_condition,
loop_body,
loop_vars=[time_step, init_decoder_state, rnn_output, context_output],
shape_invariants=[
tf.TensorShape([]),
nest.map_structure(get_state_shape_invariants, init_decoder_state),
tf.TensorShape([None, None, conf.hidden_size]),
tf.TensorShape([None, None, conf.hidden_size * 2]),
])
# body_output = tf.concat([rnn_output, context_output], axis=-1)
# body_output = tf.layers.dense(body_output, self.hidden_size, activation=tf.nn.tanh, use_bias=True, name='body_output_layer')
decoder_output = rnn_output
logits = proj_logits(decoder_output, conf.embed_size, conf.vocab_size, name='share_embedding')
# logits = proj_logits(encoder_output[:,:,:300], conf.embed_size, conf.vocab_size, name='share_embedding')
onehot_s2 = tf.one_hot(self.s2, depth=conf.vocab_size) # [batch,len,vocab]
xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=onehot_s2) # [batch,len]
weights = tf.to_float(tf.not_equal(self.s2, 0)) # [batch,len] 1 for nonpad; 0 for pad
loss_num = xentropy * weights # [batch,len]
loss_den = weights # [batch,len]
loss = tf.reduce_sum(loss_num) / tf.reduce_sum(loss_den) # scalar
self.loss = loss
# self.sent_loss = tf.reduce_sum(loss_num, -1) / tf.reduce_sum(loss_den, -1) # [batch]
# decoder infer
cache = {'state': decoder_rnn.zero_state(batch_size, tf.float32)}
def symbols_to_logits_fn(ids, i, cache):
# ids [batch,length]
pred_target = ids[:, -1:] # 截取最后一个 [batch,1]
embed_target, _ = embedding(tf.expand_dims(pred_target, axis=-1), conf.vocab_size, conf.embed_size, 'share_embedding') # [batch,length,embed]
decoder_input = tf.squeeze(embed_target, axis=1) # [batch,embed]
# if use attention
s = cache['state'] if isinstance(decoder_rnn, tf.nn.rnn_cell.GRUCell) else cache['state'].h
context = encdec_atten(s, beam_size=conf.beam_size) # [batch,hidden]
decoder_input = tf.concat([decoder_input, context], axis=-1) # [batch,hidden+]
# run rnn
# with tf.variable_scope('rnn', reuse=tf.AUTO_REUSE):
decoder_output, cache['state'] = decoder_rnn(decoder_input, cache['state'])
logits = proj_logits(decoder_output, conf.hidden_size, conf.vocab_size, name='share_embedding')
return logits, cache
initial_ids = tf.zeros([batch_size], dtype=tf.int32) # <pad>为<sos>
def greedy_search_wrapper():
""" Greedy Search """
decoded_ids, scores = greedy_search(
symbols_to_logits_fn,
initial_ids,
max_decode_len=conf.max_decode_len,
cache=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
def beam_search_wrapper():
""" Beam Search """
decoded_ids, scores = beam_search( # [batch,beam,len] [batch,beam]
symbols_to_logits_fn,
initial_ids,
beam_size=conf.beam_size,
max_decode_len=conf.max_decode_len,
vocab_size=conf.vocab_size,
states=cache,
eos_id=conf.eos_id,
gamma=conf.gamma,
num_group=conf.num_group,
top_k=conf.top_k,
)
return decoded_ids, scores
decoded_ids, scores = tf.cond(tf.equal(conf.beam_size, 1), greedy_search_wrapper, beam_search_wrapper)
self.decoded_ids = tf.identity(decoded_ids, name='decoded_ids') # [batch,beam/1,len]
self.scores = tf.identity(scores, name='scores') # [batch,beam/1]
self.global_step = tf.train.get_or_create_global_step()
self.optimizer = tf.train.AdamOptimizer(learning_rate=conf.lr)
self.train_op = self.optimizer.minimize(self.loss, global_step=self.global_step)
def build_model2(self):
# biGRU encoder + bah_attn + GRU decoder
# embedding
# [batch,len,embed]
# pretrained_word_embeddings = np.load(f'{curr_dir}/pretrain_emb_300.npy')
pretrained_word_embeddings = None
s1_embed, _ = embedding(tf.expand_dims(self.s1, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=pretrained_word_embeddings)
s1_mask = mask_nonpad_from_embedding(s1_embed) # [batch,len] 1 for nonpad; 0 for pad
s1_seqlen = tf.cast(tf.reduce_sum(s1_mask, axis=-1), tf.int32) # [batch]
# encoder
encoder_input = s1_embed
encoder_input = tf.layers.dropout(encoder_input, rate=self.dropout_rate) # dropout
with tf.variable_scope('birnn_encoder'):
self.bilstm_encoder1 = Bi_RNN(cell_name='GRUCell', hidden_size=conf.hidden_size, dropout_rate=self.dropout_rate)
encoder_output, _ = self.bilstm_encoder1(encoder_input, s1_seqlen) # [batch,len,2hid]
# decoder
s2_embed, _ = embedding(tf.expand_dims(self.s2, -1), conf.vocab_size, conf.embed_size, name='share_embedding', pretrain_embedding=pretrained_word_embeddings)
s2_mask = mask_nonpad_from_embedding(s2_embed) # [batch,len] 1 for nonpad; 0 for pad
s2_seqlen = tf.cast(tf.reduce_sum(s2_mask, -1), tf.int32) # [batch]
decoder_input = s2_embed
decoder_input = shift_right(decoder_input) # 用pad当做eos
decoder_input = tf.layers.dropout(decoder_input, rate=self.dropout_rate) # dropout
decoder_rnn = tf.nn.rnn_cell.DropoutWrapper(getattr(tf.nn.rnn_cell, 'GRUCell')(conf.hidden_size), # GRUCell/LSTMCell
input_keep_prob=1.0 - self.dropout_rate)
attention_mechanism = getattr(tf.contrib.seq2seq, 'BahdanauAttention')(
conf.hidden_size,
encoder_output,
memory_sequence_length=s1_seqlen,
name='BahdanauAttention',
)
cell = tf.contrib.seq2seq.AttentionWrapper(decoder_rnn,
attention_mechanism,
output_attention=False,
name='attention_wrapper',
)
with tf.variable_scope('decoder'):
decoder_output, _ = tf.nn.dynamic_rnn(
cell,
decoder_input,
s2_seqlen,
initial_state=None, # 默认用0向量初始化
dtype=tf.float32,
time_major=False
) # 默认scope是rnn e.g.decoder/rnn/kernal
logits = proj_logits(decoder_output, conf.embed_size, conf.vocab_size, name='share_embedding')
# logits = proj_logits(encoder_output[:,:,:300], conf.embed_size, conf.vocab_size, name='share_embedding')
onehot_s2 = tf.one_hot(self.s2, depth=conf.vocab_size) # [batch,len,vocab]
xentropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=onehot_s2) # [batch,len]
weights = tf.to_float(tf.not_equal(self.s2, 0)) # [batch,len] 1 for nonpad; 0 for pad
loss_num = xentropy * weights # [batch,len]
loss_den = weights # [batch,len]
loss = tf.reduce_sum(loss_num) / tf.reduce_sum(loss_den) # scalar
self.loss = loss
# self.sent_loss = tf.reduce_sum(loss_num, -1) / tf.reduce_sum(loss_den, -1) # [batch]
"""
decoder infer
"""
# batch_size = shape_list(encoder_output)[0]
batch_size = shape_list(encoder_output)[0]
last_dim = shape_list(encoder_output)[-1]
tile_encoder_output = tf.tile(tf.expand_dims(encoder_output, 1), [1, conf.beam_size, 1, 1])
tile_encoder_output = tf.reshape(tile_encoder_output, [batch_size * conf.beam_size, -1, last_dim])
tile_s1_seqlen = tf.tile(tf.expand_dims(s1_seqlen, 1), [1, conf.beam_size])
tile_s1_seqlent = tf.reshape(tile_s1_seqlen, [-1])
# 因为tf.BahdanauAttention在初始时就要指定beam_size来tile memory, 所以这里初始化另外一个专用于推断的beam_size tiled的并共享参数
# 不过验证还是有问题，不能用reuse=True 报 Variable memory_layer_1/kernel does not exist, or was not created with tf.get_variable()
# 故不建议使用
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
attention_mechanism_decoder = getattr(tf.contrib.seq2seq, 'BahdanauAttention')(
conf.hidden_size,
tile_encoder_output,
memory_sequence_length=tile_s1_seqlent,
name='BahdanauAttention',
)
cell_decoder = tf.contrib.seq2seq.AttentionWrapper(decoder_rnn,
attention_mechanism_decoder,
output_attention=False,
name='attention_wrapper',
)
initial_state = cell_decoder.zero_state(batch_size * conf.beam_size, tf.float32) # 内部会检查batch_size与encoder_output是否一致,需乘beam_size
# 初始化缓存
# 区分能否设在cache: cache的值在beam_search过程中会expand和merge,需要tensor rank大于1
cache = {
'cell_state': initial_state.cell_state,
'attention': initial_state.attention,
'alignments': initial_state.alignments,
'attention_state': initial_state.attention_state,
}
unable_cache = {
'alignment_history': initial_state.alignment_history,
# 'time': initial_state.time
}
# 将cache先变回batch,beam_search过程会expand/merge/gather,使得state是符合batch*beam的
cache = nest.map_structure(lambda s: s[:batch_size], cache)
def symbols_to_logits_fn(ids, i, cache):
nonlocal unable_cache
ids = ids[:, -1:]
target = tf.expand_dims(ids, axis=-1) # [batch,1,1]
embedding_target, _ = embedding(target, conf.vocab_size, conf.hidden_size, 'share_embedding', reuse=True)
input = tf.squeeze(embedding_target, axis=1) # [batch,hid]
# 合并 cache和unable_cache为state
state = cell_decoder.zero_state(batch_size * conf.beam_size, tf.float32).clone(
cell_state=cache['cell_state'],
attention=cache['attention'],
alignments=cache['alignments'],
attention_state=cache['attention_state'],
alignment_history=unable_cache['alignment_history'],
# time=unable_cache['time'],
time=tf.convert_to_tensor(i, dtype=tf.int32),
)
with tf.variable_scope('decoder/rnn', reuse=tf.AUTO_REUSE):
output, state = cell_decoder(input, state)
# 分开cache和unable_cache
cache['cell_state'] = state.cell_state
cache['attention'] = state.attention
cache['alignments'] = state.alignments
cache['attention_state'] = state.attention_state
unable_cache['alignment_history'] = state.alignment_history
# unable_cache['time'] = state.time
body_output = output # [batch,hidden]
logits = proj_logits(body_output, conf.embed_size, conf.vocab_size, name='share_embedding')
return logits, cache
initial_ids = tf.zeros([batch_size], dtype=tf.int32) # <pad>为<sos>
def greedy_search_wrapper():
""" Greedy Search """
decoded_ids, scores = greedy_search(
symbols_to_logits_fn,
initial_ids,
max_decode_len=conf.max_decode_len,
cache=cache,
eos_id=conf.eos_id,
)
return decoded_ids, scores
def beam_search_wrapper():
""" Beam Search """
decoded_ids, scores = beam_search( # [batch,beam,len] [batch,beam]
symbols_to_logits_fn,
initial_ids,
conf.beam_size,
conf.max_decode_len,
conf.vocab_size,
states=cache,
eos_id=conf.eos_id,
gamma=conf.gamma,
num_group=conf.num_group,
top_k=conf.top_k,
)
return decoded_ids, scores
decoded_ids, scores = tf.cond(tf.equal(conf.beam_size, 1), greedy_search_wrapper, beam_search_wrapper)
self.decoded_ids = tf.identity(decoded_ids, name='decoded_ids') # [batch,beam/1,len]
self.scores = tf.identity(scores, name='scores') # [batch,beam/1]
self.global_step = tf.train.get_or_create_global_step()
self.optimizer = tf.train.AdamOptimizer(learning_rate=conf.lr)
self.train_op = self.optimizer.minimize(self.loss, global_step=self.global_step)