def ensure_batch_predictions_are_consistent(self):
self.model.eval()
single_predictions = []
for i, instance in enumerate(self.instances):
dataset = Batch([instance])
tensors = dataset.as_tensor_dict(dataset.get_padding_lengths())
result = self.model(**tensors)
single_predictions.append(result)
full_dataset = Batch(self.instances)
batch_tensors = full_dataset.as_tensor_dict(
full_dataset.get_padding_lengths())
batch_predictions = self.model(**batch_tensors)
for i, instance_predictions in enumerate(single_predictions):
for key, single_predicted in instance_predictions.items():
tolerance = 1e-6
if 'loss' in key:
# Loss is particularly unstable; we'll just be satisfied if everything else is
# close.
continue
single_predicted = single_predicted[0]
batch_predicted = batch_predictions[key][i]
if isinstance(single_predicted, torch.Tensor):
if single_predicted.size() != batch_predicted.size():
slices = tuple(
slice(0, size) for size in single_predicted.size())
batch_predicted = batch_predicted[slices]
assert_allclose(single_predicted.data.numpy(),
batch_predicted.data.numpy(),
atol=tolerance,
err_msg=key)
else:
assert single_predicted == batch_predicted, key
def test_forward_pass_runs_correctly(self):
"""
Check to make sure a forward pass on an ensemble of two identical copies of a model yields the same
results as the model itself.
"""
bidaf_ensemble = BidafEnsemble([self.model, self.model])
batch = Batch(self.instances)
batch.index_instances(self.vocab)
training_tensors = batch.as_tensor_dict()
bidaf_output_dict = self.model(**training_tensors)
ensemble_output_dict = bidaf_ensemble(**training_tensors)
metrics = self.model.get_metrics(reset=True)
# We've set up the data such that there's a fake answer that consists of the whole
# paragraph. _Any_ valid prediction for that question should produce an F1 of greater than
# zero, while if we somehow haven't been able to load the evaluation data, or there was an
# error with using the evaluation script, this will fail. This makes sure that we've
# loaded the evaluation data correctly and have hooked things up to the official evaluation
# script.
assert metrics['f1'] > 0
assert torch.equal(ensemble_output_dict['best_span'],
bidaf_output_dict['best_span'])
assert ensemble_output_dict['best_span_str'] == bidaf_output_dict[
'best_span_str']
def batch_to_ids(batch: List[List[str]]) -> torch.Tensor:
"""
Converts a batch of tokenized sentences to a tensor representing the sentences with encoded characters
(len(batch), max sentence length, max word length).
Parameters
----------
batch : ``List[List[str]]``, required
A list of tokenized sentences.
Returns
-------
A tensor of padded character ids.
"""
instances = []
indexer = ELMoTokenCharactersIndexer()
for sentence in batch:
tokens = [Token(token) for token in sentence]
field = TextField(tokens, {'character_ids': indexer})
instance = Instance({"elmo": field})
instances.append(instance)
dataset = Batch(instances)
vocab = Vocabulary()
dataset.index_instances(vocab)
return dataset.as_tensor_dict()['elmo']['character_ids']
def forward_on_instances(
self, instances: List[Instance]) -> List[Dict[str, numpy.ndarray]]:
"""
Takes a list of :class:`~allennlp.data.instance.Instance`s, converts that text into
arrays using this model's :class:`Vocabulary`, passes those arrays through
:func:`self.forward()` and :func:`self.decode()` (which by default does nothing)
and returns the result. Before returning the result, we convert any
``torch.Tensors`` into numpy arrays and separate the
batched output into a list of individual dicts per instance. Note that typically
this will be faster on a GPU (and conditionally, on a CPU) than repeated calls to
:func:`forward_on_instance`.
Parameters
----------
instances : List[Instance], required
The instances to run the model on.
cuda_device : int, required
The GPU device to use. -1 means use the CPU.
Returns
-------
A list of the models output for each instance.
"""
batch_size = len(instances)
with torch.no_grad():
cuda_device = self._get_prediction_device()
dataset = Batch(instances)
dataset.index_instances(self.vocab)
model_input = dataset.as_tensor_dict(cuda_device=cuda_device)
outputs = self.decode(self(**model_input))
instance_separated_output: List[Dict[str, numpy.ndarray]] = [
{} for _ in dataset.instances
]
for name, output in list(outputs.items()):
if isinstance(output, torch.Tensor):
# NOTE(markn): This is a hack because 0-dim pytorch tensors are not iterable.
# This occurs with batch size 1, because we still want to include the loss in that case.
if output.dim() == 0:
output = output.unsqueeze(0)
if output.size(0) != batch_size:
self._maybe_warn_for_unseparable_batches(name)
continue
output = output.detach().cpu().numpy()
elif len(output) != batch_size:
self._maybe_warn_for_unseparable_batches(name)
continue
outputs[name] = output
for instance_output, batch_element in zip(
instance_separated_output, output):
instance_output[name] = batch_element
return instance_separated_output
def test_as_tensor_dict(self):
dataset = Batch(self.instances)
dataset.index_instances(self.vocab)
padding_lengths = dataset.get_padding_lengths()
tensors = dataset.as_tensor_dict(padding_lengths)
text1 = tensors["text1"]["tokens"].detach().cpu().numpy()
text2 = tensors["text2"]["tokens"].detach().cpu().numpy()
numpy.testing.assert_array_almost_equal(text1, numpy.array([[2, 3, 4, 5, 6],
[1, 3, 4, 5, 6]]))
numpy.testing.assert_array_almost_equal(text2, numpy.array([[2, 3, 4, 1, 5, 6],
[2, 3, 1, 0, 0, 0]]))
def test_tagger_with_elmo_token_embedder_forward_pass_runs_correctly(self):
dataset = Batch(self.instances)
dataset.index_instances(self.vocab)
training_tensors = dataset.as_tensor_dict()
output_dict = self.model(**training_tensors)
tags = output_dict['tags']
assert len(tags) == 2
assert len(tags[0]) == 7
assert len(tags[1]) == 7
for example_tags in tags:
for tag_id in example_tags:
tag = self.model.vocab.get_token_from_index(tag_id,
namespace="labels")
assert tag in {'O', 'I-ORG', 'I-PER', 'I-LOC'}
def _sentences_to_ids(self, sentences):
indexer = ELMoTokenCharactersIndexer()
# For each sentence, first create a TextField, then create an instance
instances = []
for sentence in sentences:
tokens = [Token(token) for token in sentence]
field = TextField(tokens, {'character_ids': indexer})
instance = Instance({'elmo': field})
instances.append(instance)
dataset = Batch(instances)
vocab = Vocabulary()
dataset.index_instances(vocab)
return dataset.as_tensor_dict()['elmo']['character_ids']
def get_vocab_and_both_elmo_indexed_ids(batch: List[List[str]]):
instances = []
indexer = ELMoTokenCharactersIndexer()
indexer2 = SingleIdTokenIndexer()
for sentence in batch:
tokens = [Token(token) for token in sentence]
field = TextField(tokens, {
'character_ids': indexer,
'tokens': indexer2
})
instance = Instance({"elmo": field})
instances.append(instance)
dataset = Batch(instances)
vocab = Vocabulary.from_instances(instances)
dataset.index_instances(vocab)
return vocab, dataset.as_tensor_dict()["elmo"]
def test_forward_pass_runs_correctly(self):
batch = Batch(self.instances)
batch.index_instances(self.vocab)
training_tensors = batch.as_tensor_dict()
output_dict = self.model(**training_tensors)
metrics = self.model.get_metrics(reset=True)
# We've set up the data such that there's a fake answer that consists of the whole
# paragraph. _Any_ valid prediction for that question should produce an F1 of greater than
# zero, while if we somehow haven't been able to load the evaluation data, or there was an
# error with using the evaluation script, this will fail. This makes sure that we've
# loaded the evaluation data correctly and have hooked things up to the official evaluation
# script.
assert metrics['f1'] > 0
span_start_probs = output_dict['span_start_probs'][0].data.numpy()
span_end_probs = output_dict['span_start_probs'][0].data.numpy()
assert_almost_equal(numpy.sum(span_start_probs, -1), 1, decimal=6)
assert_almost_equal(numpy.sum(span_end_probs, -1), 1, decimal=6)
span_start, span_end = tuple(output_dict['best_span'][0].data.numpy())
assert span_start >= 0
assert span_start <= span_end
assert span_end < self.instances[0].fields['passage'].sequence_length()
assert isinstance(output_dict['best_span_str'][0], str)