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):
seq_mapper = []
if not rnn and not self_att:
raise NotImplementedError()
if rnn:
seq_mapper.extend([VariationalDropoutLayer(0.8), CudnnGru(model_rnn_dim, w_init=TruncatedNormal(stddev=0.05))])
if self_att:
seq_mapper.extend([VariationalDropoutLayer(0.8),
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 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())