def build_beam_search_graph(self,
beam_size,
batch_size,
max_decode_length,
decode_alpha=0.6):
self.def_placeholder_and_components()
past_for_decoder = None
for i in range(0, self.input_num):
presents = self.encoder.encode(self.inputs[i], self.input_lens[i],
past_for_decoder)
if past_for_decoder is None:
past_for_decoder = presents
else:
past_for_decoder = tf.concat([past_for_decoder, presents],
axis=-2)
with tf.name_scope('beam_search'):
init_seq = tf.fill(dims=(batch_size, 1), value=self.sos_id)
seqs, scores = beamsearch.create_inference_graph(
init_seqs=init_seq,
state=past_for_decoder,
step_fn=self.decoder.decode_one_step,
hparams=self.hparams,
decode_length=max_decode_length,
batch_size=batch_size,
beam_size=beam_size,
decode_alpha=decode_alpha,
eos_id=self.eos_id,
ensemble=False,
concat_state_dim=None)
return seqs, scores
def ensemble_decoding_beam_search_graph(self,
context_list,
beam_size,
batch_size,
max_decode_length,
eos_id,
model_num,
decode_alpha=0.6):
def step(hparams, tokens, past=None, scope=None):
if scope is not None:
with tf.variable_scope(scope):
lm_output = model.model(hparams=hparams,
X=tokens,
past=past,
reuse=tf.AUTO_REUSE)
logits = lm_output['logits']
presents = lm_output['present']
presents.set_shape(
model.past_shape(hparams=hparams, batch_size=None))
return {
'logits': logits,
'presents': presents,
}
context_output_list = []
context_state_list = []
all_scopes = []
for i in range(0, model_num):
with tf.variable_scope('model_' + str(i)) as sc:
with tf.name_scope('sample_sequence'):
context_output_list.append(
step(self.hparams, context_list[i][:, :-1], scope=sc))
context_state_list.append(
context_output_list[-1]['presents'])
all_scopes.append('model_' + str(i))
with tf.name_scope('beam_search'):
init_seq = tf.expand_dims(context_list[0][:, -1], axis=1)
seqs, scores = beamsearch.create_inference_graph(
init_seqs=init_seq,
state=context_state_list,
step_fn=step,
hparams=self.hparams,
decode_length=max_decode_length,
batch_size=batch_size,
beam_size=beam_size,
decode_alpha=decode_alpha,
eos_id=eos_id,
scopes_for_ensemble=all_scopes,
ensemble=True,
concat_state_dim=None)
return seqs, scores
def build_beam_search_graph(self,
beam_size,
batch_size,
max_decode_length,
decode_alpha=0.6):
self.inputs = tf.placeholder(tf.int32, [1, None])
def step(hparams, tokens, past=None):
lm_output = model.model(hparams=hparams,
X=tokens,
past=past,
reuse=tf.AUTO_REUSE)
logits = lm_output['logits']
presents = lm_output['present']
presents.set_shape(
model.past_shape(hparams=hparams, batch_size=None))
return {
'logits': logits,
'presents': presents,
}
with tf.name_scope('sample_sequence'):
context_output = step(self.hparams, self.inputs[:, :-1])
context_state = context_output['presents']
with tf.name_scope('beam_search'):
init_seq = tf.expand_dims(self.inputs[:, -1], axis=1)
seqs, scores = beamsearch.create_inference_graph(
init_seqs=init_seq,
state=context_state,
step_fn=step,
hparams=self.hparams,
decode_length=max_decode_length,
batch_size=batch_size,
beam_size=beam_size,
decode_alpha=decode_alpha,
eos_id=self.eos_id,
ensemble=False,
concat_state_dim=-2)
return seqs, scores
def build_beam_search_graph(self,
beam_size,
batch_size,
max_decode_length,
decode_alpha=0.6):
self.def_placeholder_and_components()
emb_out = []
enc_h_out = []
past_for_decoder = []
for i in range(0, self.input_num):
past_length = 0
h = tf.gather(self.wte, self.inputs[i]) + tf.gather(
self.wpe, positions_for(self.inputs[i], past_length))
emb_out.append(h)
presents, h_enc = self.encoder.encode(h, self.input_lens[i])
enc_h_out.append(h_enc)
past_for_decoder.append(presents)
past_length = 0 if enc_h_out[0] is None else tf.shape(enc_h_out[0])[-2]
self.decoder.sef_var_for_beam_search(past_length,
enc_h_out,
beam_size=beam_size)
with tf.name_scope('beam_search'):
init_seq = tf.fill(dims=(batch_size, 1), value=self.sos_id)
seqs, scores = beamsearch.create_inference_graph(
init_seqs=init_seq,
state=past_for_decoder,
step_fn=self.decoder.decode_one_step,
hparams=self.hparams,
decode_length=max_decode_length,
batch_size=batch_size,
beam_size=beam_size,
decode_alpha=decode_alpha,
eos_id=self.eos_id,
ensemble=False,
concat_state_dim=None)
return seqs, scores