def test_create_label_tensors(self):
tensorizer = LabelTensorizer(column="label")
init = tensorizer.initialize()
init.send(None) # kick
for row in self.data.train:
init.send(row)
init.close()
rows = [
{
"label": types.Label("weather/find")
},
{
"label": types.Label("alarm/set_alarm")
},
{
"label": types.Label("non/existent")
},
]
tensors = (tensorizer.numberize(row) for row in rows)
tensor = next(tensors)
self.assertEqual(6, tensor)
tensor = next(tensors)
self.assertEqual(1, tensor)
with self.assertRaises(Exception):
tensor = next(tensors)
def test_initialize_label_tensorizer(self):
tensorizer = LabelTensorizer(column="label")
init = tensorizer.initialize()
init.send(None) # kick
for row in self.data.train:
init.send(row)
init.close()
self.assertEqual(7, len(tensorizer.labels))
def test_create_word_tagging_output_layer(self):
tensorizer = LabelTensorizer()
tensorizer.vocab = Vocabulary(["foo", "bar"])
tensorizer.pad_idx = 0
layer = WordTaggingOutputLayer.from_config(
config=WordTaggingOutputLayer.Config(label_weights={"foo": 2.2}),
labels=tensorizer.vocab,
)
np.testing.assert_array_almost_equal(
np.array([2.2, 1]), layer.loss_fn.weight.detach().numpy()
)
def test_create_label_tensors_fails_with_unknown_label(self):
tensorizer = LabelTensorizer(column="label")
init = tensorizer.initialize()
init.send(None) # kick
for row in self.data.train:
init.send(row)
init.close()
batch = [
{"label": types.Label("non/existent")},
{"label": types.Label("alarm/set_alarm")},
]
with self.assertRaises(Exception):
tensorizer.create_training_tensors(batch)
def test_create_label_tensors(self):
tensorizer = LabelTensorizer(column="label")
init = tensorizer.initialize()
init.send(None) # kick
for row in self.data.train:
init.send(row)
init.close()
batch = [
{"label": types.Label("weather/find")},
{"label": types.Label("alarm/set_alarm")},
]
tensor = tensorizer.create_training_tensors(batch)
self.assertEqual((2,), tensor.size())
self.assertEqual([6, 1], tensor.tolist())
class InputConfig(ConfigBase):
right_tokens: RoBERTaTensorizer.Config = RoBERTaTensorizer.Config()
left_tokens: RoBERTaTensorizer.Config = RoBERTaTensorizer.Config()
right_dense: Optional[FloatListTensorizer.Config] = None
left_dense: Optional[FloatListTensorizer.Config] = None
labels: LabelTensorizer.Config = LabelTensorizer.Config()
class EncoderModelInput(BaseModel.Config.ModelInput):
tokens: Tensorizer.Config = Tensorizer.Config()
dense: Optional[FloatListTensorizer.Config] = None
labels: LabelTensorizer.Config = LabelTensorizer.Config()
# for metric reporter
num_tokens: NtokensTensorizer.Config = NtokensTensorizer.Config(
names=["tokens"], indexes=[2])
class EncoderPairwiseModelInput(ModelInputBase):
tokens1: Tensorizer.Config = Tensorizer.Config()
tokens2: Tensorizer.Config = Tensorizer.Config()
labels: LabelTensorizer.Config = LabelTensorizer.Config()
# for metric reporter
num_tokens: NtokensTensorizer.Config = NtokensTensorizer.Config(
names=["tokens1", "tokens2"], indexes=[2, 2])
def test_doc_classification_output_layer(self):
tensorizer = LabelTensorizer()
tensorizer.vocab = Vocabulary([SpecialTokens.PAD, "foo", "bar"])
layer = ClassificationOutputLayer.from_config(
config=ClassificationOutputLayer.Config(loss=CrossEntropyLoss.Config()),
labels=tensorizer.vocab,
)
self.assertEqual(layer.loss_fn.ignore_index, 0)
# use default pad
tensorizer.vocab = Vocabulary(["foo", "bar"])
layer = ClassificationOutputLayer.from_config(
config=ClassificationOutputLayer.Config(loss=CrossEntropyLoss.Config()),
labels=tensorizer.vocab,
)
self.assertEqual(layer.loss_fn.ignore_index, -1)
def test_batch_predict_caffe2_model(self):
with tempfile.NamedTemporaryFile() as snapshot_file, tempfile.NamedTemporaryFile() as caffe2_model_file:
train_data = tests_module.test_file("train_data_tiny.tsv")
eval_data = tests_module.test_file("test_data_tiny.tsv")
config = PyTextConfig(
task=DocumentClassificationTask.Config(
model=DocModel.Config(
inputs=DocModel.Config.ModelInput(
tokens=TokenTensorizer.Config(),
dense=FloatListTensorizer.Config(
column="dense", dim=1, error_check=True
),
labels=LabelTensorizer.Config(),
)
),
data=Data.Config(
source=TSVDataSource.Config(
train_filename=train_data,
eval_filename=eval_data,
test_filename=eval_data,
field_names=["label", "slots", "text", "dense"],
)
),
),
version=21,
save_snapshot_path=snapshot_file.name,
export_caffe2_path=caffe2_model_file.name,
)
task = create_task(config.task)
task.export(task.model, caffe2_model_file.name)
model = task.model
save(config, model, meta=None, tensorizers=task.data.tensorizers)
pt_results = task.predict(task.data.data_source.test)
def assert_caffe2_results_correct(caffe2_results):
for pt_res, res in zip(pt_results, caffe2_results):
np.testing.assert_array_almost_equal(
pt_res["score"].tolist()[0],
[score[0] for score in res.values()],
)
results = batch_predict_caffe2_model(
snapshot_file.name, caffe2_model_file.name
)
self.assertEqual(4, len(results))
assert_caffe2_results_correct(results)
results = batch_predict_caffe2_model(
snapshot_file.name, caffe2_model_file.name, cache_size=2
)
self.assertEqual(4, len(results))
assert_caffe2_results_correct(results)
results = batch_predict_caffe2_model(
snapshot_file.name, caffe2_model_file.name, cache_size=-1
)
self.assertEqual(4, len(results))
assert_caffe2_results_correct(results)
class ModelInput(Model.Config.ModelInput):
tokens: TokenTensorizer.Config = TokenTensorizer.Config()
word_labels: SlotLabelTensorizer.Config = SlotLabelTensorizer.Config(
allow_unknown=True)
doc_labels: LabelTensorizer.Config = LabelTensorizer.Config(
allow_unknown=True)
doc_weight: Optional[FloatTensorizer.Config] = None
word_weight: Optional[FloatTensorizer.Config] = None
class BertModelInput(BaseModel.Config.ModelInput):
tokens: BERTTensorizer.Config = BERTTensorizer.Config(max_seq_len=128)
dense: Optional[FloatListTensorizer.Config] = None
labels: LabelTensorizer.Config = LabelTensorizer.Config()
# for metric reporter
num_tokens: NtokensTensorizer.Config = NtokensTensorizer.Config(
names=["tokens"], indexes=[2]
)
class ModelInput(BasePairwiseModel.Config.ModelInput):
tokens1: TokenTensorizer.Config = TokenTensorizer.Config(column="text1")
tokens2: TokenTensorizer.Config = TokenTensorizer.Config(column="text2")
labels: LabelTensorizer.Config = LabelTensorizer.Config()
# for metric reporter
raw_text: JoinStringTensorizer.Config = JoinStringTensorizer.Config(
columns=["text1", "text2"]
)
def test_create_label_tensors(self):
tensorizer = LabelTensorizer(label_column="label")
self._initialize_tensorizer(tensorizer)
rows = [
{"label": "weather/find"},
{"label": "alarm/set_alarm"},
{"label": "non/existent"},
]
tensors = (tensorizer.numberize(row) for row in rows)
tensor = next(tensors)
self.assertEqual(6, tensor)
tensor = next(tensors)
self.assertEqual(1, tensor)
with self.assertRaises(Exception):
tensor = next(tensors)
class ModelInput(BaseModel.Config.ModelInput):
squad_input: Union[
SquadForBERTTensorizer.Config, SquadForRoBERTaTensorizer.Config
] = SquadForBERTTensorizer.Config(max_seq_len=256)
# is_impossible label
has_answer: LabelTensorizer.Config = LabelTensorizer.Config(
column="has_answer"
)
def test_initialize_tensorizers(self):
tensorizers = {
"tokens": WordTensorizer(column="text"),
"labels": LabelTensorizer(column="label"),
"chars": CharacterTensorizer(column="text"),
}
initialize_tensorizers(tensorizers, self.data.train)
self.assertEqual(49, len(tensorizers["tokens"].vocab))
self.assertEqual(7, len(tensorizers["labels"].labels))
def test_initialize_tensorizers(self):
tensorizers = {
"tokens": TokenTensorizer(text_column="text"),
"labels": LabelTensorizer(label_column="label"),
"chars": ByteTensorizer(text_column="text"),
}
initialize_tensorizers(tensorizers, self.data.train)
self.assertEqual(49, len(tensorizers["tokens"].vocab))
self.assertEqual(7, len(tensorizers["labels"].vocab))
class ModelInput(ModelInputBase):
tokens1: BERTTensorizerBase.Config = BERTTensorizer.Config(
columns=["text1"], max_seq_len=128)
tokens2: BERTTensorizerBase.Config = BERTTensorizer.Config(
columns=["text2"], max_seq_len=128)
labels: LabelTensorizer.Config = LabelTensorizer.Config()
# for metric reporter
num_tokens: NtokensTensorizer.Config = NtokensTensorizer.Config(
names=["tokens1", "tokens2"], indexes=[2, 2])
class ModelInput(Model.Config.ModelInput):
tokens: TokenTensorizer.Config = TokenTensorizer.Config()
word_labels: SlotLabelTensorizer.Config = SlotLabelTensorizer.Config(
allow_unknown=True)
doc_labels: LabelTensorizer.Config = LabelTensorizer.Config(
allow_unknown=True)
doc_weight: FloatTensorizer.Config = FloatTensorizer.Config(
column="doc_weight")
word_weight: FloatTensorizer.Config = FloatTensorizer.Config(
column="word_weight")
def test_data_initializes_tensorsizers(self):
tensorizers = {
"tokens": TokenTensorizer(text_column="text"),
"labels": LabelTensorizer(label_column="label"),
}
# verify TokenTensorizer isn't in an initialized state yet
assert tensorizers["tokens"].vocab is None
Data(self.data_source, tensorizers)
# Tensorizers should have been initialized
self.assertEqual(49, len(tensorizers["tokens"].vocab))
self.assertEqual(7, len(tensorizers["labels"].vocab))
def test_data_initializes_tensorsizers(self):
tensorizers = {
"tokens": WordTensorizer(column="text"),
"labels": LabelTensorizer(column="label"),
}
with self.assertRaises(AttributeError):
# verify WordTensorizer isn't in an initialized state yet
tensorizers["tokens"].vocab
Data(self.data_source, tensorizers)
# Tensorizers should have been initialized
self.assertEqual(49, len(tensorizers["tokens"].vocab))
self.assertEqual(7, len(tensorizers["labels"].labels))
def setUp(self):
self.data_source = TSVDataSource(
SafeFileWrapper(tests_module.test_file("train_dense_features_tiny.tsv")),
SafeFileWrapper(tests_module.test_file("test_dense_features_tiny.tsv")),
eval_file=None,
field_names=["label", "slots", "text", "dense"],
schema={"text": types.Text, "label": types.Label},
)
self.tensorizers = {
"tokens": WordTensorizer(column="text"),
"labels": LabelTensorizer(column="label", allow_unknown=True),
}
class ModelInput(BasePairwiseModel.Config.ModelInput):
tokens1: TokenTensorizer.Config = TokenTensorizer.Config(
column="text1")
tokens2: TokenTensorizer.Config = TokenTensorizer.Config(
column="text2")
labels: LabelTensorizer.Config = LabelTensorizer.Config()
class InputConfig(ConfigBase):
tokens: RoBERTaTensorizer.Config = RoBERTaTensorizer.Config()
right_dense: FloatListTensorizer.Config = None
left_dense: FloatListTensorizer.Config = None
labels: LabelTensorizer.Config = LabelTensorizer.Config()
class ModelInput(Model.Config.ModelInput):
tokens: WordTensorizer.Config = WordTensorizer.Config()
labels: LabelTensorizer.Config = LabelTensorizer.Config(
allow_unknown=True)
# for metric reporter
raw_text: MetaInput.Config = MetaInput.Config(column="text")
class ModelInput(Model.Config.ModelInput):
tokens: TokenTensorizer.Config = TokenTensorizer.Config()
labels: LabelTensorizer.Config = LabelTensorizer.Config()
class ModelInput(Model.Config.ModelInput):
tokens: TokenTensorizer.Config = TokenTensorizer.Config()
dense: Optional[FloatListTensorizer.Config] = None
labels: LabelTensorizer.Config = LabelTensorizer.Config()
def test_initialize_label_tensorizer(self):
tensorizer = LabelTensorizer(label_column="label")
self._initialize_tensorizer(tensorizer)
self.assertEqual(7, len(tensorizer.vocab))
class ModelInput(Model.Config.ModelInput):
tokens: TokenTensorizer.Config = TokenTensorizer.Config()
labels: LabelTensorizer.Config = LabelTensorizer.Config(allow_unknown=True)
# for metric reporter
raw_text: RawString.Config = RawString.Config(column="text")
class InputConfig(ConfigBase):
tokens: RoBERTaTensorizer.Config = RoBERTaTensorizer.Config()
labels: LabelTensorizer.Config = LabelTensorizer.Config()
