[tests|tokenizers] Refactoring pipelines test backbone - Small tokenizers...

[tests|tokenizers] Refactoring pipelines test backbone - Small tokenizers improvements - General tests speedups (#7970)

* WIP refactoring pipeline tests - switching to fast tokenizers

* fix dialog pipeline and fill-mask

* refactoring pipeline tests backbone

* make large tests slow

* fix tests (tf Bart inactive for now)

* fix doc...

* clean up for merge

* fixing tests - remove bart from summarization until there is TF

* fix quality and RAG

* Add new translation pipeline tests - fix JAX tests

* only slow for dialog

* Fixing the missing TF-BART imports in modeling_tf_auto

* spin out pipeline tests in separate CI job

* adding pipeline test to CI YAML

* add slow pipeline tests

* speed up tf and pt join test to avoid redoing all the standalone pt and tf tests

* Update src/transformers/tokenization_utils_base.py
Co-authored-by: Sam Shleifer <sshleifer@gmail.com>

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

* Update src/transformers/pipelines.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

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

* add require_torch and require_tf in is_pt_tf_cross_test
Co-authored-by: Sam Shleifer <sshleifer@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

.circleci/config.yml

@@ -84,7 +84,7 @@ jobs:
                 key: v0.3-{{ checksum "setup.py" }}
                 paths:
                     - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ --cov --durations=0 | tee output.txt
+            - run: RUN_PT_TF_CROSS_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pt_tf_cross_test --cov --durations=0 | tee output.txt
             - run: codecov
             - store_artifacts:
                   path: ~/transformers/output.txt
@@ -164,6 +164,56 @@ jobs:
             - store_artifacts:
                   path: ~/transformers/output.txt
                   destination: test_output.txt
+    run_tests_pipelines_torch:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - restore_cache:
+                  keys:
+                      - v0.3-torch-{{ checksum "setup.py" }}
+                      - v0.3-{{ checksum "setup.py" }}
+            - run: pip install --upgrade pip
+            - run: pip install git+https://github.com/huggingface/datasets
+            - run: pip install .[sklearn,torch,testing]
+            - save_cache:
+                  key: v0.3-torch-{{ checksum "setup.py" }}
+                  paths:
+                      - '~/.cache/pip'
+            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pipeline_test | tee output.txt
+            - store_artifacts:
+                  path: ~/transformers/output.txt
+                  destination: test_output.txt
+    run_tests_pipelines_tf:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - restore_cache:
+                  keys:
+                      - v0.3-tf-{{ checksum "setup.py" }}
+                      - v0.3-{{ checksum "setup.py" }}
+            - run: pip install --upgrade pip
+            - run: pip install git+https://github.com/huggingface/datasets
+            - run: pip install .[sklearn,tf-cpu,testing]
+            - save_cache:
+                  key: v0.3-tf-{{ checksum "setup.py" }}
+                  paths:
+                      - '~/.cache/pip'
+            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pipeline_test | tee output.txt
+            - store_artifacts:
+               path: ~/transformers/output.txt
+               destination: test_output.txt
     run_tests_custom_tokenizers:
         working_directory: ~/transformers
         docker:
@@ -331,6 +381,8 @@ workflows:
             - run_tests_torch
             - run_tests_tf
             - run_tests_flax
+            - run_tests_pipelines_torch
+            - run_tests_pipelines_tf
             - build_doc
             - deploy_doc: *workflow_filters
     tpu_testing_jobs:

.github/workflows/self-push.yml

@@ -16,52 +16,52 @@ jobs:
   run_tests_torch_and_tf_gpu:
     runs-on: [self-hosted, single-gpu]
     steps:
-    - uses: actions/checkout@v2
-    - name: Python version
-      run: |
-        which python
-        python --version
-        pip --version
-    - name: Current dir
-      run: pwd
-    - run: nvidia-smi
+      - uses: actions/checkout@v2
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
 
-    - name: Loading cache.
-      uses: actions/cache@v2
-      id: cache
-      with:
-        path: .env
-        key: v0-tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v0-tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
 
-    - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
-      run: |
-        python -m venv .env
-        source .env/bin/activate
-        which python
-        python --version
-        pip --version
-    - name: Install dependencies
-      run: |
-        source .env/bin/activate
-        pip install --upgrade pip
-        pip install torch!=1.6.0
-        pip install .[sklearn,testing,onnxruntime]
-        pip install git+https://github.com/huggingface/datasets
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install torch!=1.6.0
+          pip install .[sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
 
-    - name: Are GPUs recognized by our DL frameworks
-      run: |
-        source .env/bin/activate
-        python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
-        python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
 
-    - name: Run all non-slow tests on GPU
-      env:
-        TF_FORCE_GPU_ALLOW_GROWTH: "true"
-        # TF_GPU_MEMORY_LIMIT: 4096
-        OMP_NUM_THREADS: 1
-      run: |
-        source .env/bin/activate
-        python -m pytest -n 2 --dist=loadfile -s ./tests/
+      - name: Run all non-slow tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          # TF_GPU_MEMORY_LIMIT: 4096
+          OMP_NUM_THREADS: 1
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 2 --dist=loadfile -s ./tests/
 
   run_tests_torch_and_tf_multiple_gpu:
     runs-on: [self-hosted, multi-gpu]

.github/workflows/self-scheduled.yml

@@ -12,64 +12,75 @@ jobs:
   run_all_tests_torch_and_tf_gpu:
     runs-on: [self-hosted, single-gpu]
     steps:
-    - uses: actions/checkout@v2
-
-    - name: Loading cache.
-      uses: actions/cache@v2
-      id: cache
-      with:
-        path: .env
-        key: v0-slow_tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
-
-    - name: Python version
-      run: |
-        which python
-        python --version
-        pip --version
-    - name: Current dir
-      run: pwd
-    - run: nvidia-smi
-    - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
-      if: steps.cache.outputs.cache-hit != 'true'
-      run: |
-        python -m venv .env
-        source .env/bin/activate
-        which python
-        python --version
-        pip --version
-    - name: Install dependencies
-      run: |
-        source .env/bin/activate
-        pip install --upgrade pip
-        pip install torch!=1.6.0
-        pip install .[sklearn,testing,onnxruntime]
-        pip install git+https://github.com/huggingface/datasets
-
-    - name: Are GPUs recognized by our DL frameworks
-      run: |
-        source .env/bin/activate
-        python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
-        python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
-
-
-    - name: Run all tests on GPU
-      env:
-        TF_FORCE_GPU_ALLOW_GROWTH: "true"
-        OMP_NUM_THREADS: 1
-        RUN_SLOW: yes
-      run: |
-        source .env/bin/activate
-        python -m pytest -n 1 --dist=loadfile -s ./tests/ --durations=50
-
-    - name: Run examples tests on GPU
-      env:
-        TF_FORCE_GPU_ALLOW_GROWTH: "true"
-        OMP_NUM_THREADS: 1
-        RUN_SLOW: yes
-      run: |
-        source .env/bin/activate
-        pip install -r examples/requirements.txt
-        python -m pytest -n 1 --dist=loadfile -s examples --durations=50
+      - uses: actions/checkout@v2
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v0-slow_tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install torch!=1.6.0
+          pip install .[sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+
+
+      - name: Run all tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s ./tests/ --durations=50
+
+      - name: Run examples tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+        run: |
+          source .env/bin/activate
+          pip install -r examples/requirements.txt
+          python -m pytest -n 1 --dist=loadfile -s examples --durations=50
+
+      - name: Run all pipeline tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+          RUN_PIPELINE_TESTS: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s ./tests/ -m is_pipeline_test --durations=50
+
 
   run_all_tests_torch_and_tf_multiple_gpu:
     runs-on: [self-hosted, multi-gpu]
@@ -131,3 +142,13 @@ jobs:
           source .env/bin/activate
           pip install -r examples/requirements.txt
           python -m pytest -n 1 --dist=loadfile -s examples --durations=50
+
+      - name: Run all pipeline tests on GPU
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+          RUN_PIPELINE_TESTS: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s ./tests/ -m is_pipeline_test --durations=50

src/transformers/convert_slow_tokenizers_checkpoints_to_fast.py

@@ -78,15 +78,16 @@ def convert_slow_checkpoint_to_fast(tokenizer_name, checkpoint_name, dump_path,
                 "=> {} with prefix {}, add_prefix {}".format(dump_path_full, checkpoint_prefix_name, add_prefix)
             )
 
-            file_path = list(tokenizer.pretrained_vocab_files_map.values())[0][checkpoint]
-            next_char = file_path.split(checkpoint)[-1][0]
-            if next_char == "/":
-                dump_path_full = os.path.join(dump_path_full, checkpoint_prefix_name)
-                checkpoint_prefix_name = None
-
-            logger.info(
-                "=> {} with prefix {}, add_prefix {}".format(dump_path_full, checkpoint_prefix_name, add_prefix)
-            )
+            if checkpoint in list(tokenizer.pretrained_vocab_files_map.values())[0]:
+                file_path = list(tokenizer.pretrained_vocab_files_map.values())[0][checkpoint]
+                next_char = file_path.split(checkpoint)[-1][0]
+                if next_char == "/":
+                    dump_path_full = os.path.join(dump_path_full, checkpoint_prefix_name)
+                    checkpoint_prefix_name = None
+
+                logger.info(
+                    "=> {} with prefix {}, add_prefix {}".format(dump_path_full, checkpoint_prefix_name, add_prefix)
+                )
 
             file_names = tokenizer.save_pretrained(
                 dump_path_full, legacy_format=False, filename_prefix=checkpoint_prefix_name

src/transformers/data/processors/squad.py

@@ -7,11 +7,8 @@ import numpy as np
 from tqdm import tqdm
 
 from ...file_utils import is_tf_available, is_torch_available
-from ...tokenization_bart import BartTokenizer
 from ...tokenization_bert import whitespace_tokenize
-from ...tokenization_longformer import LongformerTokenizer
-from ...tokenization_roberta import RobertaTokenizer
-from ...tokenization_utils_base import TruncationStrategy
+from ...tokenization_utils_base import PreTrainedTokenizerBase, TruncationStrategy
 from ...utils import logging
 from .utils import DataProcessor
 
@@ -112,7 +109,14 @@ def squad_convert_example_to_features(
     all_doc_tokens = []
     for (i, token) in enumerate(example.doc_tokens):
         orig_to_tok_index.append(len(all_doc_tokens))
-        if isinstance(tokenizer, (RobertaTokenizer, LongformerTokenizer, BartTokenizer)):
+        if tokenizer.__class__.__name__ in [
+            "RobertaTokenizer",
+            "LongformerTokenizer",
+            "BartTokenizer",
+            "RobertaTokenizerFast",
+            "LongformerTokenizerFast",
+            "BartTokenizerFast",
+        ]:
             sub_tokens = tokenizer.tokenize(token, add_prefix_space=True)
         else:
             sub_tokens = tokenizer.tokenize(token)
@@ -292,7 +296,7 @@ def squad_convert_example_to_features(
     return features
 
 
-def squad_convert_example_to_features_init(tokenizer_for_convert):
+def squad_convert_example_to_features_init(tokenizer_for_convert: PreTrainedTokenizerBase):
     global tokenizer
     tokenizer = tokenizer_for_convert
 
@@ -344,9 +348,9 @@ def squad_convert_examples_to_features(
             is_training=not evaluate,
         )
     """
-
     # Defining helper methods
     features = []
+
     threads = min(threads, cpu_count())
     with Pool(threads, initializer=squad_convert_example_to_features_init, initargs=(tokenizer,)) as p:
         annotate_ = partial(
@@ -365,6 +369,7 @@ def squad_convert_examples_to_features(
                 disable=not tqdm_enabled,
             )
         )
+
     new_features = []
     unique_id = 1000000000
     example_index = 0

src/transformers/modeling_tf_auto.py

@@ -52,7 +52,7 @@ from .modeling_tf_albert import (
     TFAlbertForTokenClassification,
     TFAlbertModel,
 )
-from .modeling_tf_bart import TFBartForConditionalGeneration
+from .modeling_tf_bart import TFBartForConditionalGeneration, TFBartModel
 from .modeling_tf_bert import (
     TFBertForMaskedLM,
     TFBertForMultipleChoice,
@@ -163,6 +163,7 @@ TF_MODEL_MAPPING = OrderedDict(
         (T5Config, TFT5Model),
         (DistilBertConfig, TFDistilBertModel),
         (AlbertConfig, TFAlbertModel),
+        (BartConfig, TFBartModel),
         (CamembertConfig, TFCamembertModel),
         (XLMRobertaConfig, TFXLMRobertaModel),
         (LongformerConfig, TFLongformerModel),
@@ -186,6 +187,7 @@ TF_MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
         (T5Config, TFT5ForConditionalGeneration),
         (DistilBertConfig, TFDistilBertForMaskedLM),
         (AlbertConfig, TFAlbertForPreTraining),
+        (BartConfig, TFBartForConditionalGeneration),
         (CamembertConfig, TFCamembertForMaskedLM),
         (XLMRobertaConfig, TFXLMRobertaForMaskedLM),
         (RobertaConfig, TFRobertaForMaskedLM),

src/transformers/modeling_tf_utils.py

@@ -640,12 +640,12 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin):
         # Load model
         if pretrained_model_name_or_path is not None:
             if os.path.isdir(pretrained_model_name_or_path):
-                if os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
+                if from_pt and os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
+                    # Load from a PyTorch checkpoint in priority if from_pt
+                    archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
+                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
                     # Load from a TF 2.0 checkpoint
                     archive_file = os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)
-                elif from_pt and os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
-                    # Load from a PyTorch checkpoint
-                    archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
                 else:
                     raise EnvironmentError(
                         "Error no file named {} found in directory {} or `from_pt` set to False".format(

src/transformers/modeling_utils.py

@@ -882,10 +882,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):
         if pretrained_model_name_or_path is not None:
             if os.path.isdir(pretrained_model_name_or_path):
                 if from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")):
-                    # Load from a TF 1.0 checkpoint
+                    # Load from a TF 1.0 checkpoint in priority if from_tf
                     archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")
                 elif from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
-                    # Load from a TF 2.0 checkpoint
+                    # Load from a TF 2.0 checkpoint in priority if from_tf
                     archive_file = os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)
                 elif os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
                     # Load from a PyTorch checkpoint
@@ -951,7 +951,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):
                 state_dict = torch.load(resolved_archive_file, map_location="cpu")
             except Exception:
                 raise OSError(
-                    "Unable to load weights from pytorch checkpoint file. "
+                    f"Unable to load weights from pytorch checkpoint file for '{pretrained_model_name_or_path}' "
+                    f"at '{resolved_archive_file}'"
                     "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. "
                 )
 

src/transformers/pipelines.py

@@ -38,7 +38,7 @@ from .modelcard import ModelCard
 from .tokenization_auto import AutoTokenizer
 from .tokenization_bert import BasicTokenizer
 from .tokenization_utils import PreTrainedTokenizer
-from .tokenization_utils_base import BatchEncoding, PaddingStrategy
+from .tokenization_utils_base import PaddingStrategy
 from .utils import logging
 
 
@@ -2396,11 +2396,12 @@ class ConversationalPipeline(Pipeline):
 
     def __init__(self, min_length_for_response=32, *args, **kwargs):
         super().__init__(*args, **kwargs)
+
+        # We need at least an eos_token
         assert self.tokenizer.eos_token_id is not None, "DialoguePipeline tokenizer should have an EOS token set"
-        if self.tokenizer.pad_token_id is not None:
-            self.pad_token_id = self.tokenizer.pad_token_id
-        else:
-            self.pad_token_id = self.tokenizer.eos_token_id
+        if self.tokenizer.pad_token_id is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+
         self.min_length_for_response = min_length_for_response
 
     def __call__(
@@ -2496,7 +2497,7 @@ class ConversationalPipeline(Pipeline):
         """
         # Parse arguments
         inputs = self._args_parser(*args, **kwargs)
-        inputs = self.tokenizer.batch_encode_plus(inputs, add_special_tokens=False, padding=False).get("input_ids", [])
+        inputs = self.tokenizer(inputs, add_special_tokens=False, padding=False).get("input_ids", [])
         for input in inputs:
             input.append(self.tokenizer.eos_token_id)
         return inputs
@@ -2516,7 +2517,7 @@ class ConversationalPipeline(Pipeline):
             sequence_tokens = []
             is_previous_pad = False
             for token in sequence:
-                if token == self.pad_token_id:
+                if token == self.tokenizer.pad_token_id:
                     if is_previous_pad:
                         continue
                     else:
@@ -2550,13 +2551,10 @@ class ConversationalPipeline(Pipeline):
                     else:
                         new_input = new_input[cutoff_eos_index + 1 :]
             outputs.append(new_input)
-        max_len = max([len(item) for item in outputs])
-        outputs = [output + [self.pad_token_id] * (max_len - len(output)) for output in outputs]
-        outputs = BatchEncoding(
-            {"input_ids": outputs, "attention_mask": [[1] * len(outputs)]},
-            tensor_type=self.framework,
+        padded_outputs = self.tokenizer.pad(
+            {"input_ids": outputs}, padding="longest", return_attention_mask=True, return_tensors=self.framework
         )
-        return outputs
+        return padded_outputs
 
 
 # Register all the supported tasks here
@@ -2700,6 +2698,7 @@ def pipeline(
     config: Optional[Union[str, PretrainedConfig]] = None,
     tokenizer: Optional[Union[str, PreTrainedTokenizer]] = None,
     framework: Optional[str] = None,
+    use_fast: bool = False,
     **kwargs
 ) -> Pipeline:
     """
@@ -2749,6 +2748,8 @@ def pipeline(
             If no framework is specified, will default to the one currently installed. If no framework is specified
             and both frameworks are installed, will default to the framework of the :obj:`model`, or to PyTorch if no
             model is provided.
+        use_fast (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to use a Fast tokenizer if possible (a :class:`~transformers.PreTrainedTokenizerFast`).
         kwargs:
             Additional keyword arguments passed along to the specific pipeline init (see the documentation for the
             corresponding pipeline class for possible values).
@@ -2807,9 +2808,10 @@ def pipeline(
     if isinstance(tokenizer, (str, tuple)):
         if isinstance(tokenizer, tuple):
             # For tuple we have (tokenizer name, {kwargs})
-            tokenizer = AutoTokenizer.from_pretrained(tokenizer[0], **tokenizer[1])
+            use_fast = tokenizer[1].pop("use_fast", use_fast)
+            tokenizer = AutoTokenizer.from_pretrained(tokenizer[0], use_fast=use_fast, **tokenizer[1])
         else:
-            tokenizer = AutoTokenizer.from_pretrained(tokenizer)
+            tokenizer = AutoTokenizer.from_pretrained(tokenizer, use_fast=use_fast)
 
     # Instantiate config if needed
     if isinstance(config, str):

src/transformers/testing_utils.py

@@ -59,10 +59,50 @@ def parse_int_from_env(key, default=None):
 
 
 _run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False)
+_run_pt_tf_cross_tests = parse_flag_from_env("RUN_PT_TF_CROSS_TESTS", default=False)
 _run_custom_tokenizers = parse_flag_from_env("RUN_CUSTOM_TOKENIZERS", default=False)
+_run_pipeline_tests = parse_flag_from_env("RUN_PIPELINE_TESTS", default=False)
 _tf_gpu_memory_limit = parse_int_from_env("TF_GPU_MEMORY_LIMIT", default=None)
 
 
+def is_pt_tf_cross_test(test_case):
+    """
+    Decorator marking a test as a test that control interactions between PyTorch and TensorFlow.
+
+    PT+TF tests are skipped by default and we can run only them by setting RUN_PT_TF_CROSS_TESTS environment variable
+    to a truthy value and selecting the is_pt_tf_cross_test pytest mark.
+
+    """
+    if not _run_pt_tf_cross_tests or not _torch_available or not _tf_available:
+        return unittest.skip("test is PT+TF test")(test_case)
+    else:
+        try:
+            import pytest  # We don't need a hard dependency on pytest in the main library
+        except ImportError:
+            return test_case
+        else:
+            return pytest.mark.is_pt_tf_cross_test()(test_case)
+
+
+def is_pipeline_test(test_case):
+    """
+    Decorator marking a test as a pipeline test.
+
+    Pipeline tests are skipped by default and we can run only them by setting RUN_PIPELINE_TEST environment variable
+    to a truthy value and selecting the is_pipeline_test pytest mark.
+
+    """
+    if not _run_pipeline_tests:
+        return unittest.skip("test is pipeline test")(test_case)
+    else:
+        try:
+            import pytest  # We don't need a hard dependency on pytest in the main library
+        except ImportError:
+            return test_case
+        else:
+            return pytest.mark.is_pipeline_test()(test_case)
+
+
 def slow(test_case):
     """
     Decorator marking a test as slow.

src/transformers/tokenization_utils.py

@@ -136,18 +136,6 @@ class PreTrainedTokenizer(PreTrainedTokenizerBase):
         """
         raise NotImplementedError
 
-    def get_vocab(self) -> Dict[str, int]:
-        """
-        Returns the vocabulary as a dictionary of token to index.
-
-        :obj:`tokenizer.get_vocab()[token]` is equivalent to :obj:`tokenizer.convert_tokens_to_ids(token)` when
-        :obj:`token` is in the vocab.
-
-        Returns:
-            :obj:`Dict[str, int]`: The vocabulary.
-        """
-        raise NotImplementedError()
-
     def get_added_vocab(self) -> Dict[str, int]:
         """
         Returns the added tokens in the vocabulary as a dictionary of token to index.
@@ -733,47 +721,15 @@ class PreTrainedTokenizer(PreTrainedTokenizerBase):
         raise NotImplementedError
 
     def convert_tokens_to_string(self, tokens: List[str]) -> str:
-        """
-        Converts a sequence of token ids in a single string.
-
-        The most simple way to do it is ``" ".join(tokens)`` but we often want to remove
-        sub-word tokenization artifacts at the same time.
-
-        Args:
-            tokens (:obj:`List[str]`): The token to join in a string.
-
-        Return: The joined tokens.
-        """
         return " ".join(tokens)
 
-    def decode(
+    def _decode(
         self,
         token_ids: List[int],
         skip_special_tokens: bool = False,
         clean_up_tokenization_spaces: bool = True,
         spaces_between_special_tokens: bool = True,
     ) -> str:
-        """
-        Converts a sequence of ids in a string, using the tokenizer and vocabulary
-        with options to remove special tokens and clean up tokenization spaces.
-
-        Similar to doing ``self.convert_tokens_to_string(self.convert_ids_to_tokens(token_ids))``.
-
-        Args:
-            token_ids (:obj:`List[int]`):
-                List of tokenized input ids. Can be obtained using the ``__call__`` method.
-            skip_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
-                Whether or not to remove special tokens in the decoding.
-            clean_up_tokenization_spaces (:obj:`bool`, `optional`, defaults to :obj:`True`):
-                Whether or not to clean up the tokenization spaces.
-            spaces_between_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`True`):
-                Whether or not to add spaces around special tokens.
-                The behavior of Fast tokenizers is to have this to :obj:`False`.
-                This is setup to :obj:`True` in slow tokenizers for backward compatibility.
-
-        Returns:
-            :obj:`str`: The decoded sentence.
-        """
         filtered_tokens = self.convert_ids_to_tokens(token_ids, skip_special_tokens=skip_special_tokens)
 
         # To avoid mixing byte-level and unicode for byte-level BPT

src/transformers/tokenization_utils_base.py

@@ -175,6 +175,23 @@ class TokenSpan(NamedTuple):
     end: int
 
 
+def to_py_obj(obj):
+    """
+    Convert a TensorFlow tensor, PyTorch tensor, Numpy array or python list
+    to a python list.
+    """
+    if isinstance(obj, (list, tuple)):
+        return [to_py_obj(o) for o in obj]
+    elif is_tf_available() and isinstance(obj, tf.Tensor):
+        return obj.numpy().tolist()
+    elif is_torch_available() and isinstance(obj, torch.Tensor):
+        return obj.detach().cpu().tolist()
+    elif isinstance(obj, np.ndarray):
+        return obj.tolist()
+    else:
+        return obj
+
+
 class BatchEncoding(UserDict):
     """
     Holds the output of the :meth:`~transformers.tokenization_utils_base.PreTrainedTokenizerBase.encode_plus`
@@ -1025,6 +1042,38 @@ class SpecialTokensMixin:
         """
         return self.convert_tokens_to_ids(self.additional_special_tokens)
 
+    @bos_token_id.setter
+    def bos_token_id(self, value):
+        self._bos_token = self.convert_tokens_to_ids(value)
+
+    @eos_token_id.setter
+    def eos_token_id(self, value):
+        self._eos_token = self.convert_tokens_to_ids(value)
+
+    @unk_token_id.setter
+    def unk_token_id(self, value):
+        self._unk_token = self.convert_tokens_to_ids(value)
+
+    @sep_token_id.setter
+    def sep_token_id(self, value):
+        self._sep_token = self.convert_tokens_to_ids(value)
+
+    @pad_token_id.setter
+    def pad_token_id(self, value):
+        self._pad_token = self.convert_tokens_to_ids(value)
+
+    @cls_token_id.setter
+    def cls_token_id(self, value):
+        self._cls_token = self.convert_tokens_to_ids(value)
+
+    @mask_token_id.setter
+    def mask_token_id(self, value):
+        self._mask_token = self.convert_tokens_to_ids(value)
+
+    @additional_special_tokens_ids.setter
+    def additional_special_tokens_ids(self, values):
+        self._additional_special_tokens = [self.convert_tokens_to_ids(value) for value in values]
+
     @property
     def special_tokens_map(self) -> Dict[str, Union[str, List[str]]]:
         """
@@ -1424,6 +1473,18 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
             f"padding_side='{self.padding_side}', special_tokens={self.special_tokens_map_extended})"
         )
 
+    def get_vocab(self) -> Dict[str, int]:
+        """
+        Returns the vocabulary as a dictionary of token to index.
+
+        :obj:`tokenizer.get_vocab()[token]` is equivalent to :obj:`tokenizer.convert_tokens_to_ids(token)` when
+        :obj:`token` is in the vocab.
+
+        Returns:
+            :obj:`Dict[str, int]`: The vocabulary.
+        """
+        raise NotImplementedError()
+
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *init_inputs, **kwargs):
         r"""
@@ -1852,6 +1913,32 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
         """
         raise NotImplementedError
 
+    def tokenize(self, text: str, pair: Optional[str] = None, add_special_tokens: bool = False, **kwargs) -> List[str]:
+        """
+        Converts a string in a sequence of tokens, using the backend Rust tokenizer.
+
+        Note that this method behave differently between fast and slow tokenizers:
+            - in fast tokenizers (instances of :class:`~transformers.PreTrainedTokenizerFast`), this method
+                will replace the unknown tokens with the :obj:`unk_token`,
+            - in slow tokenizers (instances of :class:`~transformers.PreTrainedTokenizer`), this method
+                keep unknown tokens unchanged.
+
+        Args:
+            text (:obj:`str`):
+                The sequence to be encoded.
+            pair (:obj:`str`, `optional`):
+                A second sequence to be encoded with the first.
+            add_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to add the special tokens associated with the corresponding model.
+            kwargs (additional keyword arguments, `optional`):
+                Will be passed to the underlying model specific encode method.
+                See details in :meth:`~transformers.PreTrainedTokenizer.__call__`
+
+        Returns:
+            :obj:`List[str]`: The list of tokens.
+        """
+        raise NotImplementedError
+
     @add_end_docstrings(
         ENCODE_KWARGS_DOCSTRING,
         """
@@ -2456,18 +2543,6 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
                     f"Should be one of a python, numpy, pytorch or tensorflow object."
                 )
 
-            def to_py_obj(obj):
-                if isinstance(obj, (list, tuple)):
-                    return [to_py_obj(o) for o in obj]
-                elif is_tf_available() and isinstance(obj, tf.Tensor):
-                    return obj.numpy().tolist()
-                elif is_torch_available() and isinstance(obj, torch.Tensor):
-                    return obj.cpu().tolist()
-                elif isinstance(obj, np.ndarray):
-                    return obj.tolist()
-                else:
-                    return obj
-
             for key, value in encoded_inputs.items():
                 encoded_inputs[key] = to_py_obj(value)
 
@@ -2862,33 +2937,53 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
 
         return encoded_inputs
 
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """
+        Converts a sequence of token ids in a single string.
+        The most simple way to do it is ``" ".join(tokens)`` but we often want to remove
+        sub-word tokenization artifacts at the same time.
+        Args:
+            tokens (:obj:`List[str]`): The token to join in a string.
+        Return: The joined tokens.
+        """
+        raise NotImplementedError
+
     def batch_decode(
-        self, sequences: List[List[int]], skip_special_tokens: bool = False, clean_up_tokenization_spaces: bool = True
+        self,
+        sequences: Union[List[int], List[List[int]], "np.ndarray", "torch.Tensor", "tf.Tensor"],
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: bool = True,
+        **kwargs
     ) -> List[str]:
         """
         Convert a list of lists of token ids into a list of strings by calling decode.
 
         Args:
-            sequences (:obj:`List[List[int]]`):
+            sequences (:obj:`Union[List[int], List[List[int]], np.ndarray, torch.Tensor, tf.Tensor]`):
                 List of tokenized input ids. Can be obtained using the ``__call__`` method.
             skip_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
                 Whether or not to remove special tokens in the decoding.
             clean_up_tokenization_spaces (:obj:`bool`, `optional`, defaults to :obj:`True`):
                 Whether or not to clean up the tokenization spaces.
+            kwargs (additional keyword arguments, `optional`):
+                Will be passed to the underlying model specific decode method.
 
         Returns:
             :obj:`List[str]`: The list of decoded sentences.
         """
         return [
             self.decode(
-                seq, skip_special_tokens=skip_special_tokens, clean_up_tokenization_spaces=clean_up_tokenization_spaces
+                seq,
+                skip_special_tokens=skip_special_tokens,
+                clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+                **kwargs,
             )
             for seq in sequences
         ]
 
     def decode(
         self,
-        token_ids: List[int],
+        token_ids: Union[int, List[int], "np.ndarray", "torch.Tensor", "tf.Tensor"],
         skip_special_tokens: bool = False,
         clean_up_tokenization_spaces: bool = True,
         **kwargs
@@ -2900,16 +2995,35 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
         Similar to doing ``self.convert_tokens_to_string(self.convert_ids_to_tokens(token_ids))``.
 
         Args:
-            token_ids (:obj:`List[int]`):
+            token_ids (:obj:`Union[int, List[int], np.ndarray, torch.Tensor, tf.Tensor]`):
                 List of tokenized input ids. Can be obtained using the ``__call__`` method.
             skip_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
                 Whether or not to remove special tokens in the decoding.
             clean_up_tokenization_spaces (:obj:`bool`, `optional`, defaults to :obj:`True`):
                 Whether or not to clean up the tokenization spaces.
+            kwargs (additional keyword arguments, `optional`):
+                Will be passed to the underlying model specific decode method.
 
         Returns:
             :obj:`str`: The decoded sentence.
         """
+        # Convert inputs to python lists
+        token_ids = to_py_obj(token_ids)
+
+        return self._decode(
+            token_ids=token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    def _decode(
+        self,
+        token_ids: Union[int, List[int]],
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: bool = True,
+        **kwargs
+    ) -> str:
         raise NotImplementedError
 
     def get_special_tokens_mask(

src/transformers/tokenization_utils_fast.py

@@ -122,17 +122,12 @@ class PreTrainedTokenizerFast(PreTrainedTokenizerBase):
         return self._tokenizer.get_vocab_size(with_added_tokens=False)
 
     def get_vocab(self) -> Dict[str, int]:
-        """
-        Returns the vocabulary as a dictionary of token to index.
-
-        :obj:`tokenizer.get_vocab()[token]` is equivalent to :obj:`tokenizer.convert_tokens_to_ids(token)` when
-        :obj:`token` is in the vocab.
-
-        Returns:
-            :obj:`Dict[str, int]`: The vocabulary.
-        """
         return self._tokenizer.get_vocab(with_added_tokens=True)
 
+    @property
+    def vocab(self) -> Dict[str, int]:
+        return self.get_vocab()
+
     def get_added_vocab(self) -> Dict[str, int]:
         """
         Returns the added tokens in the vocabulary as a dictionary of token to index.
@@ -291,25 +286,8 @@ class PreTrainedTokenizerFast(PreTrainedTokenizerBase):
             tokens.append(self._tokenizer.id_to_token(index))
         return tokens
 
-    def tokenize(self, text: str, pair: Optional[str] = None, add_special_tokens: bool = False) -> List[str]:
-        """
-        Converts a string in a sequence of tokens, using the backend Rust tokenizer.
-
-        Note that, unlike slow tokenizers (instances of :class:`~transformers.PreTrainedTokenizer`), this method
-        will replace the unknown tokens with the :obj:`unk_token`.
-
-        Args:
-            text (:obj:`str`):
-                The sequence to be encoded.
-            pair (:obj:`str`, `optional`):
-                A second sequence to be encoded with the first.
-            add_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
-                Whether or not to add the special tokens associated with the corresponding model.
-
-        Returns:
-            :obj:`List[str]`: The list of tokens.
-        """
-        return self._tokenizer.encode(text, pair, add_special_tokens=add_special_tokens).tokens
+    def tokenize(self, text: str, pair: Optional[str] = None, add_special_tokens: bool = False, **kwargs) -> List[str]:
+        return self.encode_plus(text=text, text_pair=pair, add_special_tokens=add_special_tokens, **kwargs).tokens()
 
     def set_truncation_and_padding(
         self,
@@ -405,29 +383,11 @@ class PreTrainedTokenizerFast(PreTrainedTokenizerBase):
             pad_to_multiple_of=pad_to_multiple_of,
         )
 
-        # Avoid thread overhead if only one example.
-        if len(batch_text_or_text_pairs) == 1:
-            if isinstance(batch_text_or_text_pairs[0], tuple):
-                # We got a Tuple with a pair of sequences
-                encodings = self._tokenizer.encode(
-                    *batch_text_or_text_pairs[0],
-                    add_special_tokens=add_special_tokens,
-                    is_pretokenized=is_split_into_words,
-                )
-            else:
-                # We got a single sequence
-                encodings = self._tokenizer.encode(
-                    batch_text_or_text_pairs[0],
-                    add_special_tokens=add_special_tokens,
-                    is_pretokenized=is_split_into_words,
-                )
-            encodings = [encodings]
-        else:
-            encodings = self._tokenizer.encode_batch(
-                batch_text_or_text_pairs,
-                add_special_tokens=add_special_tokens,
-                is_pretokenized=is_split_into_words,
-            )
+        encodings = self._tokenizer.encode_batch(
+            batch_text_or_text_pairs,
+            add_special_tokens=add_special_tokens,
+            is_pretokenized=is_split_into_words,
+        )
 
         # Convert encoding to dict
         # `Tokens` has type: List[Dict[str, List[List[int]]]] or List[Dict[str, 2D-Tensor]]
@@ -525,30 +485,16 @@ class PreTrainedTokenizerFast(PreTrainedTokenizerBase):
 
         return batched_output
 
-    def decode(
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        return self.backend_tokenizer.decoder.decode(tokens)
+
+    def _decode(
         self,
         token_ids: Union[int, List[int]],
         skip_special_tokens: bool = False,
         clean_up_tokenization_spaces: bool = True,
         **kwargs
     ) -> str:
-        """
-        Converts a sequence of ids in a string, using the tokenizer and vocabulary
-        with options to remove special tokens and clean up tokenization spaces.
-
-        Similar to doing ``self.convert_tokens_to_string(self.convert_ids_to_tokens(token_ids))``.
-
-        Args:
-            token_ids (:obj:`Union[int, List[int]]`):
-                List of tokenized input ids. Can be obtained using the ``__call__`` method.
-            skip_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
-                Whether or not to remove special tokens in the decoding.
-            clean_up_tokenization_spaces (:obj:`bool`, `optional`, defaults to :obj:`True`):
-                Whether or not to clean up the tokenization spaces.
-
-        Returns:
-            :obj:`str`: The decoded sentence.
-        """
         if isinstance(token_ids, int):
             token_ids = [token_ids]
         text = self._tokenizer.decode(token_ids, skip_special_tokens=skip_special_tokens)

tests/conftest.py

@@ -15,3 +15,10 @@ sys.path.insert(1, git_repo_path)
 # silence FutureWarning warnings in tests since often we can't act on them until
 # they become normal warnings - i.e. the tests still need to test the current functionality
 warnings.simplefilter(action="ignore", category=FutureWarning)
+
+
+def pytest_configure(config):
+    config.addinivalue_line("markers", "is_pipeline_test: mark test to run only when pipeline are tested")
+    config.addinivalue_line(
+        "markers", "is_pt_tf_cross_test: mark test to run only when PT and TF interactions are tested"
+    )

tests/test_modeling_tf_bart.py

@@ -19,7 +19,7 @@ import unittest
 
 from transformers import is_tf_available
 from transformers.file_utils import cached_property
-from transformers.testing_utils import require_tf, require_torch, slow
+from transformers.testing_utils import is_pt_tf_cross_test, require_tf, slow
 
 from .test_configuration_common import ConfigTester
 from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
@@ -231,8 +231,7 @@ def _long_tensor(tok_lst):
 TOLERANCE = 1e-4
 
 
-@require_tf
-@require_torch
+@is_pt_tf_cross_test
 @slow
 class TFBartModelIntegrationTest(unittest.TestCase):
     def test_inference_no_head(self):

tests/test_modeling_tf_common.py

@@ -23,8 +23,8 @@ import unittest
 from importlib import import_module
 from typing import List, Tuple
 
-from transformers import is_tf_available, is_torch_available
-from transformers.testing_utils import _tf_gpu_memory_limit, require_tf, slow
+from transformers import is_tf_available
+from transformers.testing_utils import _tf_gpu_memory_limit, is_pt_tf_cross_test, require_tf, slow
 
 
 if is_tf_available():
@@ -291,9 +291,8 @@ class TFModelTesterMixin:
         max_diff = np.amax(np.abs(out_1 - out_2))
         self.assertLessEqual(max_diff, 1e-5)
 
+    @is_pt_tf_cross_test
     def test_pt_tf_model_equivalence(self):
-        if not is_torch_available():
-            return
 
         import torch
 

tests/test_modeling_tf_pytorch.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+from transformers import is_tf_available, is_torch_available
+from transformers.testing_utils import DUMMY_UNKWOWN_IDENTIFIER, SMALL_MODEL_IDENTIFIER, is_pt_tf_cross_test, slow
+
+
+if is_tf_available():
+    from transformers import (
+        AutoConfig,
+        BertConfig,
+        GPT2Config,
+        T5Config,
+        TFAutoModel,
+        TFAutoModelForCausalLM,
+        TFAutoModelForMaskedLM,
+        TFAutoModelForPreTraining,
+        TFAutoModelForQuestionAnswering,
+        TFAutoModelForSeq2SeqLM,
+        TFAutoModelForSequenceClassification,
+        TFAutoModelWithLMHead,
+        TFBertForMaskedLM,
+        TFBertForPreTraining,
+        TFBertForQuestionAnswering,
+        TFBertForSequenceClassification,
+        TFBertModel,
+        TFGPT2LMHeadModel,
+        TFRobertaForMaskedLM,
+        TFT5ForConditionalGeneration,
+    )
+    from transformers.modeling_tf_bert import TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST
+    from transformers.modeling_tf_gpt2 import TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST
+    from transformers.modeling_tf_t5 import TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST
+
+if is_torch_available():
+    from transformers import (
+        AutoModel,
+        AutoModelForCausalLM,
+        AutoModelForMaskedLM,
+        AutoModelForPreTraining,
+        AutoModelForQuestionAnswering,
+        AutoModelForSeq2SeqLM,
+        AutoModelForSequenceClassification,
+        AutoModelWithLMHead,
+        BertForMaskedLM,
+        BertForPreTraining,
+        BertForQuestionAnswering,
+        BertForSequenceClassification,
+        BertModel,
+        GPT2LMHeadModel,
+        RobertaForMaskedLM,
+        T5ForConditionalGeneration,
+    )
+
+
+@is_pt_tf_cross_test
+class TFPTAutoModelTest(unittest.TestCase):
+    @slow
+    def test_model_from_pretrained(self):
+        import h5py
+
+        self.assertTrue(h5py.version.hdf5_version.startswith("1.10"))
+
+        # for model_name in TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+        for model_name in ["bert-base-uncased"]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, BertConfig)
+
+            model = TFAutoModel.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFBertModel)
+
+            model = AutoModel.from_pretrained(model_name, from_tf=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, BertModel)
+
+    @slow
+    def test_model_for_pretraining_from_pretrained(self):
+        import h5py
+
+        self.assertTrue(h5py.version.hdf5_version.startswith("1.10"))
+
+        # for model_name in TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+        for model_name in ["bert-base-uncased"]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, BertConfig)
+
+            model = TFAutoModelForPreTraining.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFBertForPreTraining)
+
+            model = AutoModelForPreTraining.from_pretrained(model_name, from_tf=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, BertForPreTraining)
+
+    @slow
+    def test_model_for_causal_lm(self):
+        for model_name in TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, GPT2Config)
+
+            model = TFAutoModelForCausalLM.from_pretrained(model_name, from_pt=True)
+            model, loading_info = TFAutoModelForCausalLM.from_pretrained(
+                model_name, output_loading_info=True, from_pt=True
+            )
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFGPT2LMHeadModel)
+
+            model = AutoModelForCausalLM.from_pretrained(model_name, from_tf=True)
+            model, loading_info = AutoModelForCausalLM.from_pretrained(
+                model_name, output_loading_info=True, from_tf=True
+            )
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, GPT2LMHeadModel)
+
+    @slow
+    def test_lmhead_model_from_pretrained(self):
+        for model_name in TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, BertConfig)
+
+            model = TFAutoModelWithLMHead.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFBertForMaskedLM)
+
+            model = AutoModelWithLMHead.from_pretrained(model_name, from_tf=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, BertForMaskedLM)
+
+    @slow
+    def test_model_for_masked_lm(self):
+        for model_name in TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, BertConfig)
+
+            model = TFAutoModelForMaskedLM.from_pretrained(model_name, from_pt=True)
+            model, loading_info = TFAutoModelForMaskedLM.from_pretrained(
+                model_name, output_loading_info=True, from_pt=True
+            )
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFBertForMaskedLM)
+
+            model = AutoModelForMaskedLM.from_pretrained(model_name, from_tf=True)
+            model, loading_info = AutoModelForMaskedLM.from_pretrained(
+                model_name, output_loading_info=True, from_tf=True
+            )
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, BertForMaskedLM)
+
+    @slow
+    def test_model_for_encoder_decoder_lm(self):
+        for model_name in TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, T5Config)
+
+            model = TFAutoModelForSeq2SeqLM.from_pretrained(model_name, from_pt=True)
+            model, loading_info = TFAutoModelForSeq2SeqLM.from_pretrained(
+                model_name, output_loading_info=True, from_pt=True
+            )
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFT5ForConditionalGeneration)
+
+            model = AutoModelForSeq2SeqLM.from_pretrained(model_name, from_tf=True)
+            model, loading_info = AutoModelForSeq2SeqLM.from_pretrained(
+                model_name, output_loading_info=True, from_tf=True
+            )
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, T5ForConditionalGeneration)
+
+    @slow
+    def test_sequence_classification_model_from_pretrained(self):
+        # for model_name in TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+        for model_name in ["bert-base-uncased"]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, BertConfig)
+
+            model = TFAutoModelForSequenceClassification.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFBertForSequenceClassification)
+
+            model = AutoModelForSequenceClassification.from_pretrained(model_name, from_tf=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, BertForSequenceClassification)
+
+    @slow
+    def test_question_answering_model_from_pretrained(self):
+        # for model_name in TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+        for model_name in ["bert-base-uncased"]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, BertConfig)
+
+            model = TFAutoModelForQuestionAnswering.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, TFBertForQuestionAnswering)
+
+            model = AutoModelForQuestionAnswering.from_pretrained(model_name, from_tf=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, BertForQuestionAnswering)
+
+    def test_from_pretrained_identifier(self):
+        model = TFAutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER, from_pt=True)
+        self.assertIsInstance(model, TFBertForMaskedLM)
+        self.assertEqual(model.num_parameters(), 14830)
+        self.assertEqual(model.num_parameters(only_trainable=True), 14830)
+
+        model = AutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER, from_tf=True)
+        self.assertIsInstance(model, BertForMaskedLM)
+        self.assertEqual(model.num_parameters(), 14410)
+        self.assertEqual(model.num_parameters(only_trainable=True), 14410)
+
+    def test_from_identifier_from_model_type(self):
+        model = TFAutoModelWithLMHead.from_pretrained(DUMMY_UNKWOWN_IDENTIFIER, from_pt=True)
+        self.assertIsInstance(model, TFRobertaForMaskedLM)
+        self.assertEqual(model.num_parameters(), 14830)
+        self.assertEqual(model.num_parameters(only_trainable=True), 14830)
+
+        model = AutoModelWithLMHead.from_pretrained(DUMMY_UNKWOWN_IDENTIFIER, from_tf=True)
+        self.assertIsInstance(model, RobertaForMaskedLM)
+        self.assertEqual(model.num_parameters(), 14410)
+        self.assertEqual(model.num_parameters(only_trainable=True), 14410)

tests/test_pipelines_common.py

+import unittest
+from typing import List, Optional
+
+from transformers import is_tf_available, is_torch_available, pipeline
+from transformers.pipelines import DefaultArgumentHandler, Pipeline
+from transformers.testing_utils import _run_slow_tests, is_pipeline_test, require_tf, require_torch, slow
+
+
+VALID_INPUTS = ["A simple string", ["list of strings"]]
+
+
+@is_pipeline_test
+class CustomInputPipelineCommonMixin:
+    pipeline_task = None
+    pipeline_loading_kwargs = {}
+    small_models = None  # Models tested without the @slow decorator
+    large_models = None  # Models tested with the @slow decorator
+
+    def setUp(self) -> None:
+        if not is_tf_available() and not is_torch_available():
+            return  # Currently no JAX pipelines
+
+        # Download needed checkpoints
+        models = self.small_models
+        if _run_slow_tests:
+            models = models + self.large_models
+
+        for model_name in models:
+            if is_torch_available():
+                pipeline(
+                    self.pipeline_task,
+                    model=model_name,
+                    tokenizer=model_name,
+                    framework="pt",
+                    **self.pipeline_loading_kwargs,
+                )
+            if is_tf_available():
+                pipeline(
+                    self.pipeline_task,
+                    model=model_name,
+                    tokenizer=model_name,
+                    framework="tf",
+                    **self.pipeline_loading_kwargs,
+                )
+
+    @require_torch
+    @slow
+    def test_pt_defaults(self):
+        pipeline(self.pipeline_task, framework="pt")
+
+    @require_tf
+    @slow
+    def test_tf_defaults(self):
+        pipeline(self.pipeline_task, framework="tf")
+
+    @require_torch
+    def test_torch_small(self):
+        for model_name in self.small_models:
+            nlp = pipeline(task=self.pipeline_task, model=model_name, tokenizer=model_name, framework="pt")
+            self._test_pipeline(nlp)
+
+    @require_tf
+    def test_tf_small(self):
+        for model_name in self.small_models:
+            nlp = pipeline(task=self.pipeline_task, model=model_name, tokenizer=model_name, framework="tf")
+            self._test_pipeline(nlp)
+
+    @require_torch
+    @slow
+    def test_torch_large(self):
+        for model_name in self.large_models:
+            nlp = pipeline(task=self.pipeline_task, model=model_name, tokenizer=model_name, framework="pt")
+            self._test_pipeline(nlp)
+
+    @require_tf
+    @slow
+    def test_tf_large(self):
+        for model_name in self.large_models:
+            nlp = pipeline(task=self.pipeline_task, model=model_name, tokenizer=model_name, framework="tf")
+            self._test_pipeline(nlp)
+
+    def _test_pipeline(self, nlp: Pipeline):
+        raise NotImplementedError
+
+
+@is_pipeline_test
+class MonoInputPipelineCommonMixin:
+    pipeline_task = None
+    pipeline_loading_kwargs = {}  # Additional kwargs to load the pipeline with
+    pipeline_running_kwargs = {}  # Additional kwargs to run the pipeline with
+    small_models = []  # Models tested without the @slow decorator
+    large_models = []  # Models tested with the @slow decorator
+    mandatory_keys = {}  # Keys which should be in the output
+    valid_inputs = VALID_INPUTS  # inputs which are valid
+    invalid_inputs = [None]  # inputs which are not allowed
+    expected_multi_result: Optional[List] = None
+    expected_check_keys: Optional[List[str]] = None
+
+    def setUp(self) -> None:
+        if not is_tf_available() and not is_torch_available():
+            return  # Currently no JAX pipelines
+
+        for model_name in self.small_models:
+            pipeline(self.pipeline_task, model=model_name, tokenizer=model_name, **self.pipeline_loading_kwargs)
+        for model_name in self.large_models:
+            pipeline(self.pipeline_task, model=model_name, tokenizer=model_name, **self.pipeline_loading_kwargs)
+
+    @require_torch
+    @slow
+    def test_pt_defaults_loads(self):
+        pipeline(self.pipeline_task, framework="pt", **self.pipeline_loading_kwargs)
+
+    @require_tf
+    @slow
+    def test_tf_defaults_loads(self):
+        pipeline(self.pipeline_task, framework="tf", **self.pipeline_loading_kwargs)
+
+    @require_torch
+    def test_torch_small(self):
+        for model_name in self.small_models:
+            nlp = pipeline(
+                task=self.pipeline_task,
+                model=model_name,
+                tokenizer=model_name,
+                framework="pt",
+                **self.pipeline_loading_kwargs,
+            )
+            self._test_pipeline(nlp)
+
+    @require_tf
+    def test_tf_small(self):
+        for model_name in self.small_models:
+            nlp = pipeline(
+                task=self.pipeline_task,
+                model=model_name,
+                tokenizer=model_name,
+                framework="tf",
+                **self.pipeline_loading_kwargs,
+            )
+            self._test_pipeline(nlp)
+
+    @require_torch
+    @slow
+    def test_torch_large(self):
+        for model_name in self.large_models:
+            nlp = pipeline(
+                task=self.pipeline_task,
+                model=model_name,
+                tokenizer=model_name,
+                framework="pt",
+                **self.pipeline_loading_kwargs,
+            )
+            self._test_pipeline(nlp)
+
+    @require_tf
+    @slow
+    def test_tf_large(self):
+        for model_name in self.large_models:
+            nlp = pipeline(
+                task=self.pipeline_task,
+                model=model_name,
+                tokenizer=model_name,
+                framework="tf",
+                **self.pipeline_loading_kwargs,
+            )
+            self._test_pipeline(nlp)
+
+    def _test_pipeline(self, nlp: Pipeline):
+        self.assertIsNotNone(nlp)
+
+        mono_result = nlp(self.valid_inputs[0], **self.pipeline_running_kwargs)
+        self.assertIsInstance(mono_result, list)
+        self.assertIsInstance(mono_result[0], (dict, list))
+
+        if isinstance(mono_result[0], list):
+            mono_result = mono_result[0]
+
+        for key in self.mandatory_keys:
+            self.assertIn(key, mono_result[0])
+
+        multi_result = [nlp(input, **self.pipeline_running_kwargs) for input in self.valid_inputs]
+        self.assertIsInstance(multi_result, list)
+        self.assertIsInstance(multi_result[0], (dict, list))
+
+        if self.expected_multi_result is not None:
+            for result, expect in zip(multi_result, self.expected_multi_result):
+                for key in self.expected_check_keys or []:
+                    self.assertEqual(
+                        set([o[key] for o in result]),
+                        set([o[key] for o in expect]),
+                    )
+
+        if isinstance(multi_result[0], list):
+            multi_result = multi_result[0]
+
+        for result in multi_result:
+            for key in self.mandatory_keys:
+                self.assertIn(key, result)
+
+        self.assertRaises(Exception, nlp, self.invalid_inputs)
+
+
+@is_pipeline_test
+class DefaultArgumentHandlerTestCase(unittest.TestCase):
+    def setUp(self) -> None:
+        self.handler = DefaultArgumentHandler()
+
+    def test_kwargs_x(self):
+        mono_data = {"X": "This is a sample input"}
+        mono_args = self.handler(**mono_data)
+
+        self.assertTrue(isinstance(mono_args, list))
+        self.assertEqual(len(mono_args), 1)
+
+        multi_data = {"x": ["This is a sample input", "This is a second sample input"]}
+        multi_args = self.handler(**multi_data)
+
+        self.assertTrue(isinstance(multi_args, list))
+        self.assertEqual(len(multi_args), 2)
+
+    def test_kwargs_data(self):
+        mono_data = {"data": "This is a sample input"}
+        mono_args = self.handler(**mono_data)
+
+        self.assertTrue(isinstance(mono_args, list))
+        self.assertEqual(len(mono_args), 1)
+
+        multi_data = {"data": ["This is a sample input", "This is a second sample input"]}
+        multi_args = self.handler(**multi_data)
+
+        self.assertTrue(isinstance(multi_args, list))
+        self.assertEqual(len(multi_args), 2)
+
+    def test_multi_kwargs(self):
+        mono_data = {"data": "This is a sample input", "X": "This is a sample input 2"}
+        mono_args = self.handler(**mono_data)
+
+        self.assertTrue(isinstance(mono_args, list))
+        self.assertEqual(len(mono_args), 2)
+
+        multi_data = {
+            "data": ["This is a sample input", "This is a second sample input"],
+            "test": ["This is a sample input 2", "This is a second sample input 2"],
+        }
+        multi_args = self.handler(**multi_data)
+
+        self.assertTrue(isinstance(multi_args, list))
+        self.assertEqual(len(multi_args), 4)
+
+    def test_args(self):
+        mono_data = "This is a sample input"
+        mono_args = self.handler(mono_data)
+
+        self.assertTrue(isinstance(mono_args, list))
+        self.assertEqual(len(mono_args), 1)
+
+        mono_data = ["This is a sample input"]
+        mono_args = self.handler(mono_data)
+
+        self.assertTrue(isinstance(mono_args, list))
+        self.assertEqual(len(mono_args), 1)
+
+        multi_data = ["This is a sample input", "This is a second sample input"]
+        multi_args = self.handler(multi_data)
+
+        self.assertTrue(isinstance(multi_args, list))
+        self.assertEqual(len(multi_args), 2)
+
+        multi_data = ["This is a sample input", "This is a second sample input"]
+        multi_args = self.handler(*multi_data)
+
+        self.assertTrue(isinstance(multi_args, list))
+        self.assertEqual(len(multi_args), 2)

tests/test_pipelines_conversational.py

+import unittest
+
+from transformers import Conversation, pipeline
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from .test_pipelines_common import MonoInputPipelineCommonMixin
+
+
+DEFAULT_DEVICE_NUM = -1 if torch_device == "cpu" else 0
+
+
+class TextGenerationPipelineTests(MonoInputPipelineCommonMixin, unittest.TestCase):
+    pipeline_task = "conversational"
+    small_models = []  # Models tested without the @slow decorator
+    large_models = ["microsoft/DialoGPT-medium"]  # Models tested with the @slow decorator
+    valid_inputs = [Conversation("Hi there!"), [Conversation("Hi there!"), Conversation("How are you?")]]
+    invalid_inputs = ["Hi there!", Conversation()]
+
+    def _test_pipeline(
+        self, nlp
+    ):  # e overide the default test method to check that the output is a `Conversation` object
+        self.assertIsNotNone(nlp)
+
+        mono_result = nlp(self.valid_inputs[0])
+        self.assertIsInstance(mono_result, Conversation)
+
+        multi_result = nlp(self.valid_inputs[1])
+        self.assertIsInstance(multi_result, list)
+        self.assertIsInstance(multi_result[0], Conversation)
+        # Inactive conversations passed to the pipeline raise a ValueError
+        self.assertRaises(ValueError, nlp, self.valid_inputs[1])
+
+        for bad_input in self.invalid_inputs:
+            self.assertRaises(Exception, nlp, bad_input)
+        self.assertRaises(Exception, nlp, self.invalid_inputs)
+
+    @require_torch
+    @slow
+    def test_integration_torch_conversation(self):
+        # When
+        nlp = pipeline(task="conversational", device=DEFAULT_DEVICE_NUM)
+        conversation_1 = Conversation("Going to the movies tonight - any suggestions?")
+        conversation_2 = Conversation("What's the last book you have read?")
+        # Then
+        self.assertEqual(len(conversation_1.past_user_inputs), 0)
+        self.assertEqual(len(conversation_2.past_user_inputs), 0)
+        # When
+        result = nlp([conversation_1, conversation_2], do_sample=False, max_length=1000)
+        # Then
+        self.assertEqual(result, [conversation_1, conversation_2])
+        self.assertEqual(len(result[0].past_user_inputs), 1)
+        self.assertEqual(len(result[1].past_user_inputs), 1)
+        self.assertEqual(len(result[0].generated_responses), 1)
+        self.assertEqual(len(result[1].generated_responses), 1)
+        self.assertEqual(result[0].past_user_inputs[0], "Going to the movies tonight - any suggestions?")
+        self.assertEqual(result[0].generated_responses[0], "The Big Lebowski")
+        self.assertEqual(result[1].past_user_inputs[0], "What's the last book you have read?")
+        self.assertEqual(result[1].generated_responses[0], "The Last Question")
+        # When
+        conversation_2.add_user_input("Why do you recommend it?")
+        result = nlp(conversation_2, do_sample=False, max_length=1000)
+        # Then
+        self.assertEqual(result, conversation_2)
+        self.assertEqual(len(result.past_user_inputs), 2)
+        self.assertEqual(len(result.generated_responses), 2)
+        self.assertEqual(result.past_user_inputs[1], "Why do you recommend it?")
+        self.assertEqual(result.generated_responses[1], "It's a good book.")
+
+    @require_torch
+    @slow
+    def test_integration_torch_conversation_truncated_history(self):
+        # When
+        nlp = pipeline(task="conversational", min_length_for_response=24, device=DEFAULT_DEVICE_NUM)
+        conversation_1 = Conversation("Going to the movies tonight - any suggestions?")
+        # Then
+        self.assertEqual(len(conversation_1.past_user_inputs), 0)
+        # When
+        result = nlp(conversation_1, do_sample=False, max_length=36)
+        # Then
+        self.assertEqual(result, conversation_1)
+        self.assertEqual(len(result.past_user_inputs), 1)
+        self.assertEqual(len(result.generated_responses), 1)
+        self.assertEqual(result.past_user_inputs[0], "Going to the movies tonight - any suggestions?")
+        self.assertEqual(result.generated_responses[0], "The Big Lebowski")
+        # When
+        conversation_1.add_user_input("Is it an action movie?")
+        result = nlp(conversation_1, do_sample=False, max_length=36)
+        # Then
+        self.assertEqual(result, conversation_1)
+        self.assertEqual(len(result.past_user_inputs), 2)
+        self.assertEqual(len(result.generated_responses), 2)
+        self.assertEqual(result.past_user_inputs[1], "Is it an action movie?")
+        self.assertEqual(result.generated_responses[1], "It's a comedy.")