def get_multi_hop_model(rnn_dim, c2c: bool, q2c: bool, res_rnn: bool,
res_self_att: bool, post_merge: bool, encoder: str,
merge_type: str, num_c2c_hops: int):
recurrent_layer = CudnnGru(rnn_dim, w_init=TruncatedNormal(stddev=0.05))
answer_encoder = BinaryAnswerEncoder()
res_model = get_res_fc_seq_fc(model_rnn_dim=rnn_dim,
rnn=res_rnn,
self_att=res_self_att)
context_to_context = \
AttentionWithPostMapper(BiAttention(TriLinear(bias=True), True), post_mapper=res_model) if c2c else None
question_to_context = \
AttentionWithPostMapper(BiAttention(TriLinear(bias=True), True), post_mapper=res_model) if q2c else None
if encoder == 'max':
sequence_encoder = MaxPool(map_layer=None,
min_val=0,
regular_reshape=True)
elif encoder == 'rnn':
sequence_encoder = CudnnGruEncoder(rnn_dim,
w_init=TruncatedNormal(stddev=0.05))
else:
raise NotImplementedError()
if merge_type == 'max':
attention_merger = MaxMerge(
pre_map_layer=None,
post_map_layer=(res_model if post_merge else None))
else:
attention_merger = WeightedMerge(
pre_map_layer=None,
post_map_layer=(res_model if post_merge else None),
weight_type=merge_type)
return MultiHopContextsToQuestionModel(
encoder=QuestionsAndParagraphsEncoder(answer_encoder),
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d",
word_vec_init_scale=0,
learn_unk=False,
cpu=True),
char_embed=CharWordEmbedder(LearnedCharEmbedder(word_size_th=14,
char_th=50,
char_dim=20,
init_scale=0.05,
force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True),
embed_mapper=SequenceMapperSeq(
VariationalDropoutLayer(0.8),
recurrent_layer,
VariationalDropoutLayer(0.8),
),
question_to_context_attention=question_to_context,
context_to_context_attention=context_to_context,
c2c_hops=num_c2c_hops,
context_to_question_attention=BiAttention(TriLinear(bias=True), True),
attention_merger=attention_merger,
sequence_encoder=sequence_encoder,
predictor=BinaryFixedPredictor())
def get_model(rnn_dim):
recurrent_layer = CudnnGru(rnn_dim, w_init=TruncatedNormal(stddev=0.05))
answer_encoder = BinaryAnswerEncoder()
return ContextPairRelevanceModel(
encoder=QuestionsAndParagraphsEncoder(answer_encoder),
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d",
word_vec_init_scale=0,
learn_unk=False,
cpu=True),
char_embed=CharWordEmbedder(LearnedCharEmbedder(word_size_th=14,
char_th=50,
char_dim=20,
init_scale=0.05,
force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True),
embed_mapper=SequenceMapperSeq(
VariationalDropoutLayer(0.8),
recurrent_layer,
# VariationalDropoutLayer(0.8), # fixme probably doesn't belong here
),
question_to_context_attention=None,
context_to_context_attention=None,
context_to_question_attention=None,
sequence_encoder=MaxPool(map_layer=None,
min_val=0,
regular_reshape=True),
merger=MergeTwoContextsConcatQuestion(),
predictor=BinaryFixedPredictor())
def get_reread_model(rnn_dim, use_elmo, encoder_keep_rate=0.8, reread_keep_rate=0.8,
two_phase_att=False, res_rnn=True, res_self_att=False,
multiply_iteration_probs=False, reformulate_by_context=False,
rank_first=False, rank_second=False, reread_rnn_dim=None,
first_rank_lambda=1.0, second_rank_lambda=1.0,
ranking_gamma=1.0):
recurrent_layer = CudnnGru(rnn_dim, w_init=TruncatedNormal(stddev=0.05))
answer_encoder = IterativeAnswerEncoder(group=rank_first or rank_second)
embed_mapper = SequenceMapperSeq(VariationalDropoutLayer(encoder_keep_rate), recurrent_layer)
if res_rnn or res_self_att:
res_model = get_res_fc_seq_fc(model_rnn_dim=rnn_dim, rnn=res_rnn, self_att=res_self_att,
keep_rate=reread_keep_rate)
else:
res_model = FullyConnected(rnn_dim * 2, activation="relu")
attention = AttentionWithPostMapper(BiAttention(TriLinear(bias=True), True),
post_mapper=res_model)
use_c2q = two_phase_att or not reformulate_by_context
use_q2c = two_phase_att or reformulate_by_context
elmo_model = None
if use_elmo:
print("Using Elmo!")
elmo_model = get_hotpot_elmo()
lm_reduce = MapperSeq(
ElmoLayer(0, layer_norm=False, top_layer_only=False),
DropoutLayer(0.5),
)
embed_mapper = ElmoWrapper(input_append=True, output_append=False, rnn_layer=embed_mapper, lm_reduce=lm_reduce)
return IterativeContextReReadModel(
encoder=QuestionsAndParagraphsEncoder(answer_encoder, use_sentence_segments=True, paragraph_as_sentence=True),
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d", word_vec_init_scale=0, learn_unk=False, cpu=True),
char_embed=CharWordEmbedder(
LearnedCharEmbedder(word_size_th=14, char_th=50, char_dim=20, init_scale=0.05, force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True
),
elmo_model=elmo_model,
embed_mapper=embed_mapper,
sequence_encoder=MaxPool(map_layer=None, min_val=VERY_NEGATIVE_NUMBER, regular_reshape=True),
sentences_encoder=SentenceMaxEncoder(),
sentence_mapper=None,
post_merger=None,
merger=WithConcatOptions(sub=False, hadamard=True, dot=True, raw=True),
reread_mapper=None if reread_rnn_dim is None else CudnnGru(reread_rnn_dim, w_init=TruncatedNormal(stddev=0.05)),
pre_attention_mapper=None, # VariationalDropoutLayer(reread_keep_rate),
context_to_question_attention=attention if use_c2q else None,
question_to_context_attention=attention if use_q2c else None,
reformulate_by_context=reformulate_by_context,
multiply_iteration_probs=multiply_iteration_probs,
first_predictor=BinaryNullPredictor(rank_first, ranking_lambda=first_rank_lambda, gamma=ranking_gamma),
second_predictor=BinaryNullPredictor(rank_second, ranking_lambda=second_rank_lambda, gamma=ranking_gamma),
max_batch_size=512
)
def get_model_with_yes_no(rnn_dim: int, use_elmo, keep_rate=0.8):
recurrent_layer = CudnnGru(rnn_dim, w_init=TruncatedNormal(stddev=0.05))
embed_mapper = SequenceMapperSeq(
VariationalDropoutLayer(keep_rate),
recurrent_layer,
VariationalDropoutLayer(keep_rate),
)
elmo_model = None
if use_elmo:
print("Using Elmo!")
elmo_model = get_hotpot_elmo()
lm_reduce = MapperSeq(
ElmoLayer(0, layer_norm=False, top_layer_only=False),
DropoutLayer(0.5),
)
embed_mapper = ElmoWrapper(input_append=True, output_append=False, rnn_layer=embed_mapper, lm_reduce=lm_reduce)
answer_encoder = GroupedSpanAnswerEncoderWithYesNo(group=True)
predictor = BoundsPredictor(
ChainBiMapper(
first_layer=recurrent_layer,
second_layer=recurrent_layer
),
span_predictor=IndependentBoundsGroupedWithYesNo()
)
return AttentionQAWithYesNo(
encoder=QuestionsAndParagraphsEncoder(answer_encoder, use_sentence_segments=False),
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d", word_vec_init_scale=0, learn_unk=False, cpu=True),
char_embed=CharWordEmbedder(
LearnedCharEmbedder(word_size_th=14, char_th=50, char_dim=20, init_scale=0.05, force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True
),
elmo_model=elmo_model,
embed_mapper=embed_mapper,
question_mapper=None,
context_mapper=None,
memory_builder=NullBiMapper(),
attention=BiAttention(TriLinear(bias=True), True),
match_encoder=SequenceMapperSeq(FullyConnected(rnn_dim * 2, activation="relu"),
ResidualLayer(SequenceMapperSeq(
VariationalDropoutLayer(keep_rate),
recurrent_layer,
VariationalDropoutLayer(keep_rate),
StaticAttentionSelf(TriLinear(bias=True), ConcatWithProduct()),
FullyConnected(rnn_dim * 2, activation="relu"),
)),
VariationalDropoutLayer(keep_rate)),
predictor=predictor,
yes_no_question_encoder=MaxPool(map_layer=None, min_val=0, regular_reshape=True),
yes_no_context_encoder=MaxPool(map_layer=None, min_val=0, regular_reshape=True)
)
def get_res_fc_seq_fc(model_rnn_dim, rnn: bool, self_att: bool, keep_rate=0.8):
seq_mapper = []
if not rnn and not self_att:
raise NotImplementedError()
if rnn:
seq_mapper.extend([VariationalDropoutLayer(keep_rate),
CudnnGru(model_rnn_dim, w_init=TruncatedNormal(stddev=0.05))])
if self_att:
seq_mapper.extend([VariationalDropoutLayer(keep_rate),
StaticAttentionSelf(TriLinear(bias=True), ConcatWithProduct())])
seq_mapper.append(FullyConnected(model_rnn_dim * 2, activation="relu"))
return SequenceMapperSeq(FullyConnected(model_rnn_dim * 2, activation="relu"),
ResidualLayer(SequenceMapperSeq(
*seq_mapper
)))
def _apply_transposed(self, is_train, x, initial_states=None):
w_init = TruncatedNormal(stddev=0.05)
x_size = x.shape.as_list()[-1]
if x_size is None:
raise ValueError("Last dimension must be defined (have shape %s)" %
str(x.shape))
cell = cudnn_rnn_ops.CudnnLSTM(1,
self.n_out,
x_size,
input_mode="linear_input")
# We need to know the mapping of weights/baises -> CudnnLSTM parameter, so just
# build a `CudnnLSTM` and read its fields
c = cudnn_layers.CudnnLSTM(1, self.n_out)
c._input_size = x.shape.as_list()[-1]
w_shapes = c.canonical_weight_shapes
b_shapes = c.canonical_bias_shapes
weights = [w_init(s, tf.float32) for s in w_shapes]
biases = [tf.zeros(s, tf.float32) for s in b_shapes]
biases[1] = tf.constant(self.lstm_bias / 2.0, tf.float32, b_shapes[1])
biases[5] = tf.constant(self.lstm_bias / 2.0, tf.float32, b_shapes[5])
opaque_params_t = cell.canonical_to_params(weights, biases)
parameters = tf.get_variable("opaque_kernel",
initializer=opaque_params_t,
validate_shape=False)
p = 1.0 - self.dropout
if is_train and self.dropout > 0:
mult_bias = [tf.ones_like(x) for x in biases]
mult_w = [tf.ones_like(x) for x in weights]
bias_mask = tf.floor(tf.random_uniform(
(self.n_out, ), p, 1 + p)) / p
for j in range(4, 8):
mult_w[j] *= tf.expand_dims(bias_mask, 0)
mult_mask = cell.canonical_to_params(mult_w, mult_bias)
parameters = parameters * mult_mask
initial_state_h, initial_state_c = initial_states
out = cell(x, initial_state_h, initial_state_c, parameters, True)[0]
return out
def get_reread_merge_model(rnn_dim, use_elmo, keep_rate=0.8, res_rnn=True, res_self_att=False,
multiply_iteration_probs=False):
recurrent_layer = CudnnGru(rnn_dim, w_init=TruncatedNormal(stddev=0.05))
answer_encoder = IterativeAnswerEncoder()
embed_mapper = SequenceMapperSeq(VariationalDropoutLayer(keep_rate), recurrent_layer)
if res_rnn or res_self_att:
res_model = get_res_fc_seq_fc(model_rnn_dim=rnn_dim, rnn=res_rnn, self_att=res_self_att)
else:
res_model = FullyConnected(rnn_dim * 2, activation="relu")
attention = AttentionWithPostMapper(BiAttention(TriLinear(bias=True), True), post_mapper=res_model)
elmo_model = None
if use_elmo:
print("Using Elmo!")
elmo_model = get_hotpot_elmo()
lm_reduce = MapperSeq(
ElmoLayer(0, layer_norm=False, top_layer_only=False),
DropoutLayer(0.5),
)
embed_mapper = ElmoWrapper(input_append=True, output_append=False, rnn_layer=embed_mapper, lm_reduce=lm_reduce)
return IterativeContextReReadMergeModel(
encoder=QuestionsAndParagraphsEncoder(answer_encoder, use_sentence_segments=True),
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d", word_vec_init_scale=0, learn_unk=False, cpu=True),
char_embed=CharWordEmbedder(
LearnedCharEmbedder(word_size_th=14, char_th=50, char_dim=20, init_scale=0.05, force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True
),
elmo_model=elmo_model,
embed_mapper=embed_mapper,
sequence_encoder=MaxPool(map_layer=None, min_val=0, regular_reshape=True),
sentences_encoder=SentenceMaxEncoder(),
sentence_mapper=None,
post_merger=None,
merger=WithConcatOptions(sub=False, hadamard=True, dot=False, raw=True),
context_to_question_attention=attention,
question_to_context_attention=attention,
reread_merger=ConcatWithProduct(),
multiply_iteration_probs=multiply_iteration_probs,
max_batch_size=128
)
def get_model(rnn_dim, use_elmo, keep_rate=0.8):
recurrent_layer = CudnnGru(rnn_dim, w_init=TruncatedNormal(stddev=0.05))
answer_encoder = IterativeAnswerEncoder()
embed_mapper = SequenceMapperSeq(VariationalDropoutLayer(keep_rate), recurrent_layer)
elmo_model = None
if use_elmo:
print("Using Elmo!")
elmo_model = get_hotpot_elmo()
lm_reduce = MapperSeq(
ElmoLayer(0, layer_norm=False, top_layer_only=False),
DropoutLayer(0.5),
)
embed_mapper = ElmoWrapper(input_append=True, output_append=False, rnn_layer=embed_mapper, lm_reduce=lm_reduce)
reformulation = ProjectMapEncodeReformulation(project_layer=None,
sequence_mapper=None,
encoder=CudnnGruEncoder(rnn_dim, w_init=TruncatedNormal(stddev=0.05)))
# reformulation = WeightedSumThenProjectReformulation(rnn_dim*2, activation='relu')
# reformulation = ProjectThenWeightedSumReformulation(rnn_dim*2, activation='relu')
return IterativeContextMaxSentenceModel(
encoder=QuestionsAndParagraphsEncoder(answer_encoder, use_sentence_segments=True),
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d", word_vec_init_scale=0, learn_unk=False, cpu=True),
char_embed=CharWordEmbedder(
LearnedCharEmbedder(word_size_th=14, char_th=50, char_dim=20, init_scale=0.05, force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True
),
elmo_model=elmo_model,
embed_mapper=embed_mapper,
sequence_encoder=MaxPool(map_layer=None, min_val=0, regular_reshape=True),
sentences_encoder=SentenceMaxEncoder(),
sentence_mapper=recurrent_layer,
post_merger=None,
merger=WithConcatOptions(sub=False, hadamard=True, dot=True, raw=True),
reformulation_layer=reformulation,
max_batch_size=128
)
def get_contexts_to_question_model(rnn_dim, post_merge):
recurrent_layer = CudnnGru(rnn_dim, w_init=TruncatedNormal(stddev=0.05))
answer_encoder = BinaryAnswerEncoder()
if post_merge == 'res_rnn_self_att':
post_map_layer = SequenceMapperSeq(
FullyConnected(rnn_dim * 2, activation="relu"),
ResidualLayer(
SequenceMapperSeq(
VariationalDropoutLayer(0.8),
recurrent_layer,
VariationalDropoutLayer(0.8),
StaticAttentionSelf(TriLinear(bias=True),
ConcatWithProduct()),
FullyConnected(rnn_dim * 2, activation="relu"),
)))
elif post_merge == 'res_rnn':
post_map_layer = SequenceMapperSeq(
FullyConnected(rnn_dim * 2, activation="relu"),
ResidualLayer(
SequenceMapperSeq(
VariationalDropoutLayer(0.8),
recurrent_layer,
FullyConnected(rnn_dim * 2, activation="relu"),
)))
elif post_merge == 'res_self_att':
post_map_layer = SequenceMapperSeq(
FullyConnected(rnn_dim * 2, activation="relu"),
ResidualLayer(
SequenceMapperSeq(
VariationalDropoutLayer(0.8),
StaticAttentionSelf(TriLinear(bias=True),
ConcatWithProduct()),
FullyConnected(rnn_dim * 2, activation="relu"),
)))
else:
raise NotImplementedError()
return ContextsToQuestionModel(
encoder=QuestionsAndParagraphsEncoder(answer_encoder),
word_embed=FixedWordEmbedder(vec_name="glove.840B.300d",
word_vec_init_scale=0,
learn_unk=False,
cpu=True),
char_embed=CharWordEmbedder(LearnedCharEmbedder(word_size_th=14,
char_th=50,
char_dim=20,
init_scale=0.05,
force_cpu=True),
MaxPool(Conv1d(100, 5, 0.8)),
shared_parameters=True),
embed_mapper=SequenceMapperSeq(
VariationalDropoutLayer(0.8),
recurrent_layer,
VariationalDropoutLayer(0.8),
),
attention_merger=MaxMerge(pre_map_layer=None,
post_map_layer=post_map_layer),
context_to_question_attention=BiAttention(TriLinear(bias=True), True),
sequence_encoder=MaxPool(map_layer=None,
min_val=0,
regular_reshape=True),
predictor=BinaryFixedPredictor())