Zero shot classification pipeline (#5760)
* add initial zero-shot pipeline * change default args * update default template * add label string splitting * add str labels support, remove nli from name * style * add input validation and working tf defaults * tests * quality check * add docstring to __call__ * add slow tests * Change truncation to only_first also lower precision on tests for readibility * style
This commit is contained in:
Joe Davison
@@ -318,6 +318,138 @@ class MonoColumnInputTestCase(unittest.TestCase):
|
||||
QA_FINETUNED_MODELS = ["sshleifer/tiny-distilbert-base-cased-distilled-squad"]
|
||||
|
||||
|
||||
class ZeroShotClassificationPipelineTests(unittest.TestCase):
|
||||
def _test_scores_sum_to_one(self, result):
|
||||
sum = 0.0
|
||||
for score in result["scores"]:
|
||||
sum += score
|
||||
self.assertAlmostEqual(sum, 1.0)
|
||||
|
||||
def _test_zero_shot_pipeline(self, nlp):
|
||||
output_keys = {"sequence", "labels", "scores"}
|
||||
valid_mono_inputs = [
|
||||
{"sequences": "Who are you voting for in 2020?", "candidate_labels": "politics"},
|
||||
{"sequences": "Who are you voting for in 2020?", "candidate_labels": ["politics"]},
|
||||
{"sequences": "Who are you voting for in 2020?", "candidate_labels": "politics, public health"},
|
||||
{"sequences": "Who are you voting for in 2020?", "candidate_labels": ["politics", "public health"]},
|
||||
{"sequences": ["Who are you voting for in 2020?"], "candidate_labels": "politics"},
|
||||
{
|
||||
"sequences": "Who are you voting for in 2020?",
|
||||
"candidate_labels": "politics",
|
||||
"hypothesis_template": "This text is about {}",
|
||||
},
|
||||
]
|
||||
valid_multi_input = {
|
||||
"sequences": ["Who are you voting for in 2020?", "What is the capital of Spain?"],
|
||||
"candidate_labels": "politics",
|
||||
}
|
||||
invalid_inputs = [
|
||||
{"sequences": None, "candidate_labels": "politics"},
|
||||
{"sequences": "", "candidate_labels": "politics"},
|
||||
{"sequences": "Who are you voting for in 2020?", "candidate_labels": None},
|
||||
{"sequences": "Who are you voting for in 2020?", "candidate_labels": ""},
|
||||
{
|
||||
"sequences": "Who are you voting for in 2020?",
|
||||
"candidate_labels": "politics",
|
||||
"hypothesis_template": None,
|
||||
},
|
||||
{
|
||||
"sequences": "Who are you voting for in 2020?",
|
||||
"candidate_labels": "politics",
|
||||
"hypothesis_template": "",
|
||||
},
|
||||
{
|
||||
"sequences": "Who are you voting for in 2020?",
|
||||
"candidate_labels": "politics",
|
||||
"hypothesis_template": "Template without formatting syntax.",
|
||||
},
|
||||
]
|
||||
self.assertIsNotNone(nlp)
|
||||
|
||||
for mono_input in valid_mono_inputs:
|
||||
mono_result = nlp(**mono_input)
|
||||
self.assertIsInstance(mono_result, dict)
|
||||
if len(mono_result["labels"]) > 1:
|
||||
self._test_scores_sum_to_one(mono_result)
|
||||
|
||||
for key in output_keys:
|
||||
self.assertIn(key, mono_result)
|
||||
|
||||
multi_result = nlp(**valid_multi_input)
|
||||
self.assertIsInstance(multi_result, list)
|
||||
self.assertIsInstance(multi_result[0], dict)
|
||||
self.assertEqual(len(multi_result), len(valid_multi_input["sequences"]))
|
||||
|
||||
for result in multi_result:
|
||||
for key in output_keys:
|
||||
self.assertIn(key, result)
|
||||
|
||||
if len(result["labels"]) > 1:
|
||||
self._test_scores_sum_to_one(result)
|
||||
|
||||
for bad_input in invalid_inputs:
|
||||
self.assertRaises(Exception, nlp, **bad_input)
|
||||
|
||||
def _test_zero_shot_pipeline_outputs(self, nlp):
|
||||
inputs = [
|
||||
{
|
||||
"sequences": "Who are you voting for in 2020?",
|
||||
"candidate_labels": ["politics", "public health", "science"],
|
||||
},
|
||||
{
|
||||
"sequences": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.",
|
||||
"candidate_labels": ["machine learning", "statistics", "translation", "vision"],
|
||||
"multi_class": True,
|
||||
},
|
||||
]
|
||||
|
||||
expected_outputs = [
|
||||
{
|
||||
"sequence": "Who are you voting for in 2020?",
|
||||
"labels": ["politics", "public health", "science"],
|
||||
"scores": [0.975, 0.015, 0.008],
|
||||
},
|
||||
{
|
||||
"sequence": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.",
|
||||
"labels": ["translation", "machine learning", "vision", "statistics"],
|
||||
"scores": [0.817, 0.712, 0.018, 0.017],
|
||||
},
|
||||
]
|
||||
|
||||
for input, expected_output in zip(inputs, expected_outputs):
|
||||
output = nlp(**input)
|
||||
for key in output:
|
||||
if key == "scores":
|
||||
for output_score, expected_score in zip(output[key], expected_output[key]):
|
||||
self.assertAlmostEqual(output_score, expected_score, places=2)
|
||||
else:
|
||||
self.assertEqual(output[key], expected_output[key])
|
||||
|
||||
@require_torch
|
||||
def test_torch_zero_shot_classification(self):
|
||||
for model_name in TEXT_CLASSIF_FINETUNED_MODELS:
|
||||
nlp = pipeline(task="zero-shot-classification", model=model_name, tokenizer=model_name)
|
||||
self._test_zero_shot_pipeline(nlp)
|
||||
|
||||
@require_tf
|
||||
def test_tf_zero_shot_classification(self):
|
||||
for model_name in TEXT_CLASSIF_FINETUNED_MODELS:
|
||||
nlp = pipeline(task="zero-shot-classification", model=model_name, tokenizer=model_name, framework="tf")
|
||||
self._test_zero_shot_pipeline(nlp)
|
||||
|
||||
@slow
|
||||
@require_torch
|
||||
def test_torch_zero_shot_outputs(self):
|
||||
nlp = pipeline(task="zero-shot-classification", model="roberta-large-mnli")
|
||||
self._test_zero_shot_pipeline_outputs(nlp)
|
||||
|
||||
@slow
|
||||
@require_tf
|
||||
def test_tf_zero_shot_outputs(self):
|
||||
nlp = pipeline(task="zero-shot-classification", model="roberta-large-mnli", framework="tf")
|
||||
self._test_zero_shot_pipeline_outputs(nlp)
|
||||
|
||||
|
||||
class QAPipelineTests(unittest.TestCase):
|
||||
def _test_qa_pipeline(self, nlp):
|
||||
output_keys = {"score", "answer", "start", "end"}
|
||||
|
||||
Reference in New Issue
Block a user