Adding `batch_size` support for (almost) all pipelines (#13724)

* Tentative enabling of `batch_size` for pipelines.

* Add systematic test for pipeline batching.

* Enabling batch_size on almost all pipelines

- Not `zero-shot` (it's already passing stuff as batched so trickier)
- Not `QA` (preprocess uses squad features, we need to switch to real
tensors at this boundary.

* Adding `min_length_for_response` for conversational.

* Making CTC, speech mappings avaiable regardless of framework.

* Attempt at fixing automatic tests (ffmpeg not enabled for fast tests)

* Removing ffmpeg dependency in tests.

* Small fixes.

* Slight cleanup.

* Adding docs

and adressing comments.

* Quality.

* Update docs/source/main_classes/pipelines.rst
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/question_answering.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/zero_shot_classification.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Improving docs.

* Update docs/source/main_classes/pipelines.rst
Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com>

* N -> oberved_batch_size

softmax trick.

* Follow `padding_side`.

* Supporting image pipeline batching (and padding).

* Rename `unbatch` -> `loader_batch`.

* unbatch_size forgot.

* Custom padding for offset mappings.

* Attempt to remove librosa.

* Adding require_audio.

* torchaudio.

* Back to using datasets librosa.

* Adding help to set a pad_token on the tokenizer.

* Update src/transformers/pipelines/base.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/base.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/base.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Quality.
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com>

tests/test_pipelines_summarization.py

@@ -36,8 +36,12 @@ class SummarizationPipelineTests(unittest.TestCase, metaclass=PipelineTestCaseMe
     model_mapping = MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
     tf_model_mapping = TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
 
-    def run_pipeline_test(self, model, tokenizer, feature_extractor):
+    def get_test_pipeline(self, model, tokenizer, feature_extractor):
         summarizer = SummarizationPipeline(model=model, tokenizer=tokenizer)
+        return summarizer, ["(CNN)The Palestinian Authority officially became", "Some other text"]
+
+    def run_pipeline_test(self, summarizer, _):
+        model = summarizer.model
 
         outputs = summarizer("(CNN)The Palestinian Authority officially became")
         self.assertEqual(outputs, [{"summary_text": ANY(str)}])

tests/test_pipelines_text2text_generation.py

@@ -30,9 +30,11 @@ class Text2TextGenerationPipelineTests(unittest.TestCase, metaclass=PipelineTest
     model_mapping = MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
     tf_model_mapping = TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
 
-    def run_pipeline_test(self, model, tokenizer, feature_extractor):
+    def get_test_pipeline(self, model, tokenizer, feature_extractor):
         generator = Text2TextGenerationPipeline(model=model, tokenizer=tokenizer)
+        return generator, ["Something to write", "Something else"]
 
+    def run_pipeline_test(self, generator, _):
         outputs = generator("Something there")
         self.assertEqual(outputs, [{"generated_text": ANY(str)}])
         # These are encoder decoder, they don't just append to incoming string

tests/test_pipelines_text_classification.py

@@ -72,9 +72,12 @@ class TextClassificationPipelineTests(unittest.TestCase, metaclass=PipelineTestC
         outputs = text_classifier("Birds are a type of animal")
         self.assertEqual(nested_simplify(outputs), [{"label": "POSITIVE", "score": 0.988}])
 
-    def run_pipeline_test(self, model, tokenizer, feature_extractor):
+    def get_test_pipeline(self, model, tokenizer, feature_extractor):
         text_classifier = TextClassificationPipeline(model=model, tokenizer=tokenizer)
+        return text_classifier, ["HuggingFace is in", "This is another test"]
 
+    def run_pipeline_test(self, text_classifier, _):
+        model = text_classifier.model
         # Small inputs because BartTokenizer tiny has maximum position embeddings = 22
         valid_inputs = "HuggingFace is in"
         outputs = text_classifier(valid_inputs)

tests/test_pipelines_text_generation.py

@@ -88,8 +88,14 @@ class TextGenerationPipelineTests(unittest.TestCase, metaclass=PipelineTestCaseM
             ],
         )
 
-    def run_pipeline_test(self, model, tokenizer, feature_extractor):
+    def get_test_pipeline(self, model, tokenizer, feature_extractor):
         text_generator = TextGenerationPipeline(model=model, tokenizer=tokenizer)
+        return text_generator, ["This is a test", "Another test"]
+
+    def run_pipeline_test(self, text_generator, _):
+        model = text_generator.model
+        tokenizer = text_generator.tokenizer
+
         outputs = text_generator("This is a test")
         self.assertEqual(outputs, [{"generated_text": ANY(str)}])
         self.assertTrue(outputs[0]["generated_text"].startswith("This is a test"))

tests/test_pipelines_token_classification.py

@@ -45,8 +45,13 @@ class TokenClassificationPipelineTests(unittest.TestCase, metaclass=PipelineTest
     model_mapping = MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING
     tf_model_mapping = TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING
 
-    def run_pipeline_test(self, model, tokenizer, feature_extractor):
+    def get_test_pipeline(self, model, tokenizer, feature_extractor):
         token_classifier = TokenClassificationPipeline(model=model, tokenizer=tokenizer)
+        return token_classifier, ["A simple string", "A simple string that is quite a bit longer"]
+
+    def run_pipeline_test(self, token_classifier, _):
+        model = token_classifier.model
+        tokenizer = token_classifier.tokenizer
 
         outputs = token_classifier("A simple string")
         self.assertIsInstance(outputs, list)

tests/test_pipelines_translation.py

@@ -20,6 +20,7 @@ from transformers import (
     MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
     TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
     MBart50TokenizerFast,
+    MBartConfig,
     MBartForConditionalGeneration,
     TranslationPipeline,
     pipeline,
@@ -34,14 +35,16 @@ class TranslationPipelineTests(unittest.TestCase, metaclass=PipelineTestCaseMeta
     model_mapping = MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
     tf_model_mapping = TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING
 
-    def run_pipeline_test(self, model, tokenizer, feature_extractor):
-        translator = TranslationPipeline(model=model, tokenizer=tokenizer)
-        try:
-            outputs = translator("Some string")
-        except ValueError:
-            # Triggered by m2m langages
-            src_lang, tgt_lang = list(translator.tokenizer.lang_code_to_id.keys())[:2]
-            outputs = translator("Some string", src_lang=src_lang, tgt_lang=tgt_lang)
+    def get_test_pipeline(self, model, tokenizer, feature_extractor):
+        if isinstance(model.config, MBartConfig):
+            src_lang, tgt_lang = list(tokenizer.lang_code_to_id.keys())[:2]
+            translator = TranslationPipeline(model=model, tokenizer=tokenizer, src_lang=src_lang, tgt_lang=tgt_lang)
+        else:
+            translator = TranslationPipeline(model=model, tokenizer=tokenizer)
+        return translator, ["Some string", "Some other text"]
+
+    def run_pipeline_test(self, translator, _):
+        outputs = translator("Some string")
         self.assertEqual(outputs, [{"translation_text": ANY(str)}])
 
     @require_torch

tests/test_pipelines_zero_shot.py

@@ -31,9 +31,13 @@ class ZeroShotClassificationPipelineTests(unittest.TestCase, metaclass=PipelineT
     model_mapping = MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING
     tf_model_mapping = TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING
 
-    def run_pipeline_test(self, model, tokenizer, feature_extractor):
-        classifier = ZeroShotClassificationPipeline(model=model, tokenizer=tokenizer)
+    def get_test_pipeline(self, model, tokenizer, feature_extractor):
+        classifier = ZeroShotClassificationPipeline(
+            model=model, tokenizer=tokenizer, candidate_labels=["polics", "health"]
+        )
+        return classifier, ["Who are you voting for in 2020?", "My stomach hurts."]
 
+    def run_pipeline_test(self, classifier, _):
         outputs = classifier("Who are you voting for in 2020?", candidate_labels="politics")
         self.assertEqual(outputs, {"sequence": ANY(str), "labels": [ANY(str)], "scores": [ANY(float)]})