def forward_on_instances(
self, instances: List[Instance],
cuda_device: int) -> 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.autograd.Variables`` or ``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`.
"""
dataset = Dataset(instances)
dataset.index_instances(self.vocab)
model_input = dataset.as_tensor_dict(cuda_device=cuda_device,
for_training=False)
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.autograd.Variable):
output = output.data.cpu().numpy()
outputs[name] = output
for instance_output, batch_element in zip(
instance_separated_output, output):
instance_output[name] = batch_element
return instance_separated_output
def get_answer():
# Take user input and convert to Instance
user_context = request.args.get("context", "", type=str)
user_question = request.args.get("question", "", type=str)
input_instance = squad_reader.text_to_instance(
question_text=user_question, passage_text=user_context)
# Make a dataset from the instance
dataset = Dataset([input_instance])
dataset.index_instances(train_vocab)
batch = dataset.as_tensor_dict(cuda_device=0 if cuda else -1,
for_training=False)
# Extract relevant data from batch.
passage = batch["passage"]["tokens"]
question = batch["question"]["tokens"]
metadata = batch.get("metadata", {})
# Run data through model to get start and end logits.
output_dict = model(passage, question)
start_logits = output_dict["start_logits"]
end_logits = output_dict["end_logits"]
# Compute the best span
best_span = get_best_span(start_logits, end_logits)
# Get the string corresponding to the best span
passage_str = metadata[0]['original_passage']
offsets = metadata[0]['token_offsets']
predicted_span = tuple(best_span[0].data.cpu().numpy())
start_offset = offsets[predicted_span[0]][0]
end_offset = offsets[predicted_span[1]][1]
best_span_string = passage_str[start_offset:end_offset]
# Return the best string back to the GUI
return jsonify(answer=best_span_string)
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 = Dataset(instances)
vocab = Vocabulary()
dataset.index_instances(vocab)
return dataset.as_tensor_dict()['elmo']['character_ids']
def ensure_batch_predictions_are_consistent(self):
self.model.eval()
single_predictions = []
for i, instance in enumerate(self.dataset.instances):
dataset = Dataset([instance])
tensors = dataset.as_tensor_dict(dataset.get_padding_lengths(),
for_training=False)
result = self.model(**tensors)
single_predictions.append(result)
batch_tensors = self.dataset.as_tensor_dict(
self.dataset.get_padding_lengths(), for_training=False)
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 key == 'loss':
# 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.autograd.Variable):
if single_predicted.size() != batch_predicted.size():
# This is probably a sequence model, and our output shape has some padded
# elements in the batched case. Fixing this in general is complicated;
# we'll just fix some easy cases that we actually have, for now.
num_tokens = single_predicted.size(0)
if batch_predicted.dim() == 1:
batch_predicted = batch_predicted[:num_tokens]
elif batch_predicted.dim() == 2:
batch_predicted = batch_predicted[:num_tokens, :]
else:
raise NotImplementedError
assert_allclose(single_predicted.data.numpy(),
batch_predicted.data.numpy(),
atol=tolerance,
err_msg=key)
else:
assert single_predicted == batch_predicted, key