def test_init_checks_dropout_consistency(self):
with pytest.raises(ConfigurationError):
Maxout(input_dim=2,
num_layers=3,
output_dims=5,
pool_sizes=4,
dropout=[0.2, 0.3])
def from_params(cls, vocab: Vocabulary, params: Params) -> 'BiattentiveClassificationNetwork':
embedder_params = params.pop("text_field_embedder")
text_field_embedder = TextFieldEmbedder.from_params(vocab, embedder_params)
embedding_dropout = params.pop("embedding_dropout")
pre_encode_feedforward = FeedForward.from_params(params.pop("pre_encode_feedforward"))
encoder = Seq2SeqEncoder.from_params(params.pop("encoder"))
integrator = Seq2SeqEncoder.from_params(params.pop("integrator"))
integrator_dropout = params.pop("integrator_dropout")
output_layer_params = params.pop("output_layer")
if "activations" in output_layer_params:
output_layer = FeedForward.from_params(output_layer_params)
else:
output_layer = Maxout.from_params(output_layer_params)
initializer = InitializerApplicator.from_params(params.pop('initializer', []))
regularizer = RegularizerApplicator.from_params(params.pop('regularizer', []))
return cls(vocab=vocab,
text_field_embedder=text_field_embedder,
embedding_dropout=embedding_dropout,
pre_encode_feedforward=pre_encode_feedforward,
encoder=encoder,
integrator=integrator,
integrator_dropout=integrator_dropout,
output_layer=output_layer,
initializer=initializer,
regularizer=regularizer)
def test_forward_gives_correct_output(self):
params = Params({
'input_dim': 2,
'output_dims': 3,
'pool_sizes': 4,
'dropout': 0.0,
'num_layers': 2
})
maxout = Maxout.from_params(params)
constant_init = Initializer.from_params(
Params({
"type": "constant",
"val": 1.
}))
initializer = InitializerApplicator([(".*", constant_init)])
initializer(maxout)
input_tensor = torch.FloatTensor([[-3, 1]])
output = maxout(input_tensor).data.numpy()
assert output.shape == (1, 3)
# This output was checked by hand
# The output of the first maxout layer is [-1, -1, -1], since the
# matrix multiply gives us [-2]*12. Reshaping and maxing
# produces [-2, -2, -2] and the bias increments these values.
# The second layer output is [-2, -2, -2], since the matrix
# matrix multiply gives us [-3]*12. Reshaping and maxing
# produces [-3, -3, -3] and the bias increments these values.
assert_almost_equal(output, [[-2, -2, -2]])
def test_forward_gives_correct_output(self):
params = Params({
'input_dim': 2,
'output_dims': 3,
'pool_sizes': 4,
'dropout': 0.0,
'num_layers': 2
})
maxout = Maxout.from_params(params)
constant_init = Initializer.from_params(Params({"type": "constant", "val": 1.}))
initializer = InitializerApplicator([(".*", constant_init)])
initializer(maxout)
input_tensor = torch.FloatTensor([[-3, 1]])
output = maxout(input_tensor).data.numpy()
assert output.shape == (1, 3)
# This output was checked by hand
# The output of the first maxout layer is [-1, -1, -1], since the
# matrix multiply gives us [-2]*12. Reshaping and maxing
# produces [-2, -2, -2] and the bias increments these values.
# The second layer output is [-2, -2, -2], since the matrix
# matrix multiply gives us [-3]*12. Reshaping and maxing
# produces [-3, -3, -3] and the bias increments these values.
assert_almost_equal(output, [[-2, -2, -2]])
def test_init_checks_pool_sizes_consistency(self):
with pytest.raises(ConfigurationError):
Maxout(input_dim=2,
num_layers=2,
output_dims=5,
pool_sizes=[4, 5, 2],
dropout=0.0)
def __init__(self,
vocab: Vocabulary,
embedder: TextFieldEmbedder,
question_encoder: Seq2SeqEncoder,
passage_encoder: Seq2SeqEncoder,
feed_forward: FeedForward,
dropout: float = 0.1,
num_decoding_steps: int = 40,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(AnswerSynthesis, self).__init__(vocab, regularizer)
self._vocab = vocab
self._vocab_size = vocab.get_vocab_size() # default: tokens
self._num_decoding_steps = num_decoding_steps
self._start_token_index = self._vocab.get_token_index(START_SYMBOL)
self._end_token_index = self._vocab.get_token_index(END_SYMBOL)
self._embedder = embedder
self._question_encoder = question_encoder
self._passage_encoder = passage_encoder
encoding_dim = question_encoder.get_output_dim()
embedding_dim = embedder.get_output_dim()
self._span_start_embedding = nn.Embedding(2, 50)
self._span_end_embedding = nn.Embedding(2, 50)
self._gru_decoder = nn.GRUCell(encoding_dim + embedding_dim,
encoding_dim)
self._feed_forward = feed_forward
self._attention = Attention(NonlinearSimilarity(encoding_dim))
self._W_r = nn.Linear(embedding_dim, encoding_dim, bias=False)
self._U_r = nn.Linear(encoding_dim, encoding_dim, bias=False)
self._V_r = nn.Linear(encoding_dim, encoding_dim, bias=False)
self._max_out = Maxout(encoding_dim,
num_layers=1,
output_dims=int(encoding_dim / 2),
pool_sizes=2)
self._W_o = nn.Linear(int(encoding_dim / 2),
self._vocab_size,
bias=False)
self._squad_metrics = SquadEmAndF1()
#self._predict_acc = CategoricalAccuracy()
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
initializer(self)
self._num_iter = 0
def from_params( # type: ignore
cls, vocab: Vocabulary,
params: Params) -> "BiattentiveClassificationNetwork":
embedder_params = params.pop("text_field_embedder")
text_field_embedder = TextFieldEmbedder.from_params(
vocab=vocab, params=embedder_params)
embedding_dropout = params.pop("embedding_dropout")
pre_encode_feedforward = FeedForward.from_params(
params.pop("pre_encode_feedforward"))
encoder = Seq2SeqEncoder.from_params(params.pop("encoder"))
integrator = Seq2SeqEncoder.from_params(params.pop("integrator"))
integrator_dropout = params.pop("integrator_dropout")
output_layer_params = params.pop("output_layer")
if "activations" in output_layer_params:
output_layer = FeedForward.from_params(output_layer_params)
else:
output_layer = Maxout.from_params(output_layer_params)
elmo = params.pop("elmo", None)
if elmo is not None:
elmo = Elmo.from_params(elmo)
use_input_elmo = params.pop_bool("use_input_elmo", False)
use_integrator_output_elmo = params.pop_bool(
"use_integrator_output_elmo", False)
initializer = InitializerApplicator.from_params(
params.pop("initializer", []))
regularizer = RegularizerApplicator.from_params(
params.pop("regularizer", []))
params.assert_empty(cls.__name__)
return cls(
vocab=vocab,
text_field_embedder=text_field_embedder,
embedding_dropout=embedding_dropout,
pre_encode_feedforward=pre_encode_feedforward,
encoder=encoder,
integrator=integrator,
integrator_dropout=integrator_dropout,
output_layer=output_layer,
elmo=elmo,
use_input_elmo=use_input_elmo,
use_integrator_output_elmo=use_integrator_output_elmo,
initializer=initializer,
regularizer=regularizer,
)
def from_params(
cls, vocab: Vocabulary,
params: Params) -> 'SemEvalClassifierAttention': # type: ignore
# pylint: disable=arguments-differ
embedder_params = params.pop("text_field_embedder")
text_field_embedder = TextFieldEmbedder.from_params(
vocab=vocab, params=embedder_params)
embedding_dropout = params.pop("embedding_dropout")
encoder = Seq2SeqEncoder.from_params(params.pop("encoder"))
integrator = Seq2SeqEncoder.from_params(params.pop("integrator"))
integrator_dropout = params.pop("integrator_dropout")
output_layer_params = params.pop("output_layer")
if "activations" in output_layer_params:
output_layer = FeedForward.from_params(output_layer_params)
else:
output_layer = Maxout.from_params(output_layer_params)
elmo = params.pop("elmo", None)
if elmo is not None:
elmo = Elmo.from_params(elmo)
use_input_elmo = params.pop_bool("use_input_elmo", False)
use_integrator_output_elmo = params.pop_bool(
"use_integrator_output_elmo", False)
initializer = InitializerApplicator.from_params(
params.pop('initializer', []))
regularizer = RegularizerApplicator.from_params(
params.pop('regularizer', []))
params.assert_empty(cls.__name__)
return cls(vocab=vocab,
text_field_embedder=text_field_embedder,
embedding_dropout=embedding_dropout,
encoder=encoder,
integrator=integrator,
integrator_dropout=integrator_dropout,
output_layer=output_layer,
initializer=initializer,
regularizer=regularizer)
def from_params(cls, vocab: Vocabulary, params: Params) -> 'BiattentiveClassificationNetwork': # type: ignore
# pylint: disable=arguments-differ
embedder_params = params.pop("text_field_embedder")
text_field_embedder = TextFieldEmbedder.from_params(vocab=vocab, params=embedder_params)
embedding_dropout = params.pop("embedding_dropout")
pre_encode_feedforward = FeedForward.from_params(params.pop("pre_encode_feedforward"))
encoder = Seq2SeqEncoder.from_params(params.pop("encoder"))
integrator = Seq2SeqEncoder.from_params(params.pop("integrator"))
integrator_dropout = params.pop("integrator_dropout")
output_layer_params = params.pop("output_layer")
if "activations" in output_layer_params:
output_layer = FeedForward.from_params(output_layer_params)
else:
output_layer = Maxout.from_params(output_layer_params)
elmo = params.pop("elmo", None)
if elmo is not None:
elmo = Elmo.from_params(elmo)
use_input_elmo = params.pop_bool("use_input_elmo", False)
use_integrator_output_elmo = params.pop_bool("use_integrator_output_elmo", False)
initializer = InitializerApplicator.from_params(params.pop('initializer', []))
regularizer = RegularizerApplicator.from_params(params.pop('regularizer', []))
params.assert_empty(cls.__name__)
return cls(vocab=vocab,
text_field_embedder=text_field_embedder,
embedding_dropout=embedding_dropout,
pre_encode_feedforward=pre_encode_feedforward,
encoder=encoder,
integrator=integrator,
integrator_dropout=integrator_dropout,
output_layer=output_layer,
elmo=elmo,
use_input_elmo=use_input_elmo,
use_integrator_output_elmo=use_integrator_output_elmo,
initializer=initializer,
regularizer=regularizer)
def from_params(cls, vocab , params ) : # type: ignore
# pylint: disable=arguments-differ
embedder_params = params.pop(u"text_field_embedder")
text_field_embedder = TextFieldEmbedder.from_params(vocab=vocab, params=embedder_params)
embedding_dropout = params.pop(u"embedding_dropout")
pre_encode_feedforward = FeedForward.from_params(params.pop(u"pre_encode_feedforward"))
encoder = Seq2SeqEncoder.from_params(params.pop(u"encoder"))
integrator = Seq2SeqEncoder.from_params(params.pop(u"integrator"))
integrator_dropout = params.pop(u"integrator_dropout")
output_layer_params = params.pop(u"output_layer")
if u"activations" in output_layer_params:
output_layer = FeedForward.from_params(output_layer_params)
else:
output_layer = Maxout.from_params(output_layer_params)
elmo = params.pop(u"elmo", None)
if elmo is not None:
elmo = Elmo.from_params(elmo)
use_input_elmo = params.pop_bool(u"use_input_elmo", False)
use_integrator_output_elmo = params.pop_bool(u"use_integrator_output_elmo", False)
initializer = InitializerApplicator.from_params(params.pop(u'initializer', []))
regularizer = RegularizerApplicator.from_params(params.pop(u'regularizer', []))
params.assert_empty(cls.__name__)
return cls(vocab=vocab,
text_field_embedder=text_field_embedder,
embedding_dropout=embedding_dropout,
pre_encode_feedforward=pre_encode_feedforward,
encoder=encoder,
integrator=integrator,
integrator_dropout=integrator_dropout,
output_layer=output_layer,
elmo=elmo,
use_input_elmo=use_input_elmo,
use_integrator_output_elmo=use_integrator_output_elmo,
initializer=initializer,
regularizer=regularizer)
def __init__(self, embedder: TextFieldEmbedder,
embedding_dropout_prob: float, embedding_dim: int,
use_input_elmo: bool, pre_encode_dim: Union[int, List[int]],
pre_encode_layer_dropout_prob: Union[float, List[float]],
encoder_dim: int, integrtator_dim: int,
integrator_dropout: float, use_integrator_output_elmo: bool,
output_dim: Union[int, List[int]], output_pool_size: int,
output_dropout_prob: Union[float, List[float]], comp: int,
elmo: Elmo):
super(LABiattentiveClassificationNetwork, self).__init__()
# pre_encode_feedforward
self.text_field_embedder = embedder
self.embedding_dropout = nn.Dropout(embedding_dropout_prob)
self.use_input_elmo = use_input_elmo
embedding_dim += ELMO_DIM if self.use_input_elmo else 0
if isinstance(pre_encode_dim, int):
pre_encode_layer_num = 1
pre_encode_dim = [pre_encode_dim]
pre_encode_layer_dropout_prob = [pre_encode_layer_dropout_prob]
else:
pre_encode_layer_num = len(pre_encode_dim)
self.pre_encode_feedforward = FeedForward(
input_dim=embedding_dim,
num_layers=pre_encode_layer_num,
hidden_dims=pre_encode_dim,
activations=[nn.ReLU()] * pre_encode_layer_num,
dropout=pre_encode_layer_dropout_prob)
pytorch_encoder = nn.LSTM(input_size=pre_encode_dim[-1],
hidden_size=encoder_dim,
num_layers=1,
bidirectional=True,
batch_first=True)
self.encoder = PytorchSeq2SeqWrapper(pytorch_encoder)
pytorch_integrator = nn.LSTM(input_size=6 * encoder_dim,
hidden_size=integrtator_dim,
num_layers=1,
bidirectional=True,
batch_first=True)
self.integrator = PytorchSeq2SeqWrapper(pytorch_integrator)
self.integrator_dropout = nn.Dropout(p=integrator_dropout)
self.use_integrator_output_elmo = use_integrator_output_elmo
if self.use_integrator_output_elmo:
self.combined_integrator_output_dim = (
self.integrator.get_output_dim() + self.elmo.get_output_dim())
else:
self.combined_integrator_output_dim = self.integrator.get_output_dim(
)
self.self_attentive_pooling_projection = nn.Linear(
self.combined_integrator_output_dim, 1)
if isinstance(output_dim, int):
output_layer_num = 1
la_output_dim = [output_dim]
output_dropout_prob = [output_dropout_prob]
else:
la_output_dim = copy(output_dim)
output_layer_num = len(output_dim)
la_output_dim[-1] = la_output_dim[-1] * comp
self.la_output_layer = Maxout(input_dim=integrtator_dim * 8,
num_layers=output_layer_num,
output_dims=la_output_dim,
pool_sizes=output_pool_size,
dropout=output_dropout_prob)
initializer = InitializerApplicator()
self.elmo = elmo
initializer(self)
