def test_get_best_span(self):
# pylint: disable=protected-access
span_begin_probs = torch.FloatTensor([[0.1, 0.3, 0.05, 0.3,
0.25]]).log()
span_end_probs = torch.FloatTensor([[0.65, 0.05, 0.2, 0.05,
0.05]]).log()
begin_end_idxs = BidirectionalAttentionFlow.get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[0, 0]])
# When we were using exclusive span ends, this was an edge case of the dynamic program.
# We're keeping the test to make sure we get it right now, after the switch in inclusive
# span end. The best answer is (1, 1).
span_begin_probs = torch.FloatTensor([[0.4, 0.5, 0.1]]).log()
span_end_probs = torch.FloatTensor([[0.3, 0.6, 0.1]]).log()
begin_end_idxs = BidirectionalAttentionFlow.get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[1, 1]])
# Another instance that used to be an edge case.
span_begin_probs = torch.FloatTensor([[0.8, 0.1, 0.1]]).log()
span_end_probs = torch.FloatTensor([[0.8, 0.1, 0.1]]).log()
begin_end_idxs = BidirectionalAttentionFlow.get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[0, 0]])
span_begin_probs = torch.FloatTensor([[0.1, 0.2, 0.05, 0.3,
0.25]]).log()
span_end_probs = torch.FloatTensor([[0.1, 0.2, 0.5, 0.05, 0.15]]).log()
begin_end_idxs = BidirectionalAttentionFlow.get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[1, 2]])
def test_get_best_span(self):
# pylint: disable=protected-access
span_begin_probs = Variable(torch.FloatTensor([[0.1, 0.3, 0.05, 0.3, 0.25]])).log()
span_end_probs = Variable(torch.FloatTensor([[0.65, 0.05, 0.2, 0.05, 0.05]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[0, 0]])
# When we were using exlcusive span ends, this was an edge case of the dynamic program.
# We're keeping the test to make sure we get it right now, after the switch in inclusive
# span end. The best answer is (1, 1).
span_begin_probs = Variable(torch.FloatTensor([[0.4, 0.5, 0.1]])).log()
span_end_probs = Variable(torch.FloatTensor([[0.3, 0.6, 0.1]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[1, 1]])
# Another instance that used to be an edge case.
span_begin_probs = Variable(torch.FloatTensor([[0.8, 0.1, 0.1]])).log()
span_end_probs = Variable(torch.FloatTensor([[0.8, 0.1, 0.1]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[0, 0]])
span_begin_probs = Variable(torch.FloatTensor([[0.1, 0.2, 0.05, 0.3, 0.25]])).log()
span_end_probs = Variable(torch.FloatTensor([[0.1, 0.2, 0.5, 0.05, 0.15]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[1, 2]])
def __init__(self, nlp_toolkit: NLPToolkit):
self.nlp_toolkit = nlp_toolkit
archive_loader = ArchiveLoader(_MODEL_ARCHIVE)
config = archive_loader.get_config()
model_params = config.get('model')
vocabulary = archive_loader.get_vocabulary()
self.bidaf_model = BidirectionalAttentionFlow.from_params(
vocabulary, model_params)
model_state = archive_loader.get_model_state()
self.bidaf_model.load_state_dict(model_state)
self.vocab_reader = archive_loader.get_vocab_reader()
def test_get_best_span(self):
# pylint: disable=protected-access
# Note that the best span cannot be (1, 0) since even though 0.3 * 0.5 is the greatest
# value, the end span index is constrained to occur after the begin span index.
span_begin_probs = Variable(
torch.FloatTensor([[0.1, 0.3, 0.05, 0.3, 0.25]])).log()
span_end_probs = Variable(
torch.FloatTensor([[0.5, 0.1, 0.2, 0.05, 0.15]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[1, 2]])
# Testing an edge case of the dynamic program here, for the order of when you update the
# best previous span position. We should not get (1, 1), because that's an empty span.
span_begin_probs = Variable(torch.FloatTensor([[0.4, 0.5, 0.1]])).log()
span_end_probs = Variable(torch.FloatTensor([[0.3, 0.6, 0.1]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[0, 1]])
# Testing another edge case of the dynamic program here, where (0, 0) is the best solution
# without constraints.
span_begin_probs = Variable(torch.FloatTensor([[0.8, 0.1, 0.1]])).log()
span_end_probs = Variable(torch.FloatTensor([[0.8, 0.1, 0.1]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[0, 1]])
# test higher-order input
# Note that the best span cannot be (1, 1) since even though 0.3 * 0.5 is the greatest
# value, the end span index is constrained to occur after the begin span index.
span_begin_probs = Variable(
torch.FloatTensor([[0.1, 0.3, 0.05, 0.3, 0.25]])).log()
span_end_probs = Variable(
torch.FloatTensor([[0.1, 0.5, 0.2, 0.05, 0.15]])).log()
begin_end_idxs = BidirectionalAttentionFlow._get_best_span(
span_begin_probs, span_end_probs)
assert_almost_equal(begin_end_idxs.data.numpy(), [[1, 2]])
def setUp(self):
super(BidirectionalAttentionFlowTest, self).setUp()
constants.GLOVE_PATH = 'tests/fixtures/glove.6B.100d.sample.txt.gz'
reader_params = Params({
'token_indexers': {
'tokens': {
'type': 'single_id'
},
'token_characters': {
'type': 'characters'
}
}
})
dataset = SquadReader.from_params(reader_params).read(
'tests/fixtures/data/squad.json')
vocab = Vocabulary.from_dataset(dataset)
self.vocab = vocab
dataset.index_instances(vocab)
self.dataset = dataset
self.token_indexers = {
'tokens': SingleIdTokenIndexer(),
'token_characters': TokenCharactersIndexer()
}
self.model = BidirectionalAttentionFlow.from_params(
self.vocab, Params({}))
small_params = Params({
'text_field_embedder': {
'tokens': {
'type': 'embedding',
'pretrained_file': constants.GLOVE_PATH,
'trainable': False,
'projection_dim': 4
},
'token_characters': {
'type': 'character_encoding',
'embedding': {
'embedding_dim': 8
},
'encoder': {
'type': 'cnn',
'embedding_dim': 8,
'num_filters': 4,
'ngram_filter_sizes': [5]
}
}
},
'phrase_layer': {
'type': 'lstm',
'bidirectional': True,
'input_size': 8,
'hidden_size': 4,
'num_layers': 1,
},
'similarity_function': {
'type': 'linear',
'combination': 'x,y,x*y',
'tensor_1_dim': 8,
'tensor_2_dim': 8
},
'modeling_layer': {
'type': 'lstm',
'bidirectional': True,
'input_size': 32,
'hidden_size': 4,
'num_layers': 1,
},
'span_end_encoder': {
'type': 'lstm',
'bidirectional': True,
'input_size': 56,
'hidden_size': 4,
'num_layers': 1,
},
})
self.small_model = BidirectionalAttentionFlow.from_params(
self.vocab, small_params)
