Merge branch 'huggingface:master' into master

.github/workflows/self-scheduled.yml

@@ -37,6 +37,7 @@ jobs:
           pip install --upgrade pip
           pip install .[integrations,sklearn,testing,onnxruntime,sentencepiece,torch-speech,vision,timm]
           pip install https://github.com/kpu/kenlm/archive/master.zip
+          python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
 
       - name: Are GPUs recognized by our DL frameworks
         run: |
@@ -241,6 +242,7 @@ jobs:
           pip install --upgrade pip
           pip install .[integrations,sklearn,testing,onnxruntime,sentencepiece,torch-speech,vision,timm]
           pip install https://github.com/kpu/kenlm/archive/master.zip
+          python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'
 
       - name: Are GPUs recognized by our DL frameworks
         run: |

examples/flax/image-captioning/README.md

+# Image Captioning (vision-encoder-text-decoder model) training example
+
+The following example showcases how to finetune a vision-encoder-text-decoder model for image captioning
+using the JAX/Flax backend, leveraging 🤗 Transformers library's [FlaxVisionEncoderDecoderModel](https://huggingface.co/docs/transformers/model_doc/visionencoderdecoder#transformers.FlaxVisionEncoderDecoderModel).
+
+JAX/Flax allows you to trace pure functions and compile them into efficient, fused accelerator code on both GPU and TPU.
+Models written in JAX/Flax are **immutable** and updated in a purely functional
+way which enables simple and efficient model parallelism.
+
+`run_image_captioning_flax.py` is a lightweight example of how to download and preprocess a dataset from the 🤗 Datasets
+library or use your own files (jsonlines or csv), then fine-tune one of the architectures above on it.
+
+For custom datasets in `jsonlines` format please see: https://huggingface.co/docs/datasets/loading_datasets.html#json-files and you also will find examples of these below.
+
+### Download COCO dataset (2017)
+This example uses COCO dataset (2017) through a custom dataset script, which requires users to manually download the
+COCO dataset before training.
+
+```bash
+mkdir data
+cd data
+wget http://images.cocodataset.org/zips/train2017.zip
+wget http://images.cocodataset.org/zips/val2017.zip
+wget http://images.cocodataset.org/zips/test2017.zip
+wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
+wget http://images.cocodataset.org/annotations/image_info_test2017.zip
+cd ..
+```
+
+### Create a model from a vision encoder model and a text decoder model
+Next, we create a [FlaxVisionEncoderDecoderModel](https://huggingface.co/docs/transformers/model_doc/visionencoderdecoder#transformers.FlaxVisionEncoderDecoderModel) instance from a pre-trained vision encoder ([ViT](https://huggingface.co/docs/transformers/model_doc/vit#transformers.FlaxViTModel)) and a pre-trained text decoder ([GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2#transformers.FlaxGPT2Model)):
+
+```bash
+python3 create_model_from_encoder_decoder_models.py \
+    --output_dir model \
+    --encoder_model_name_or_path google/vit-base-patch16-224-in21k \
+    --decoder_model_name_or_path gpt2
+```
+
+### Train the model
+Finally, we can run the example script to train the model:
+
+```bash
+python3 run_image_captioning_flax.py \
+	--output_dir ./image-captioning-training-results \
+	--model_name_or_path model \
+	--dataset_name ydshieh/coco_dataset_script \
+	--dataset_config_name=2017 \
+	--data_dir $PWD/data \
+	--image_column image_path \
+	--caption_column caption \
+	--do_train --do_eval --predict_with_generate \
+	--num_train_epochs 1 \
+	--eval_steps 500 \
+	--learning_rate 3e-5 --warmup_steps 0 \
+	--per_device_train_batch_size 32 \
+	--per_device_eval_batch_size 32 \
+	--overwrite_output_dir \
+	--max_target_length 32 \
+	--num_beams 8 \
+	--preprocessing_num_workers 16 \
+	--logging_steps 10 \
+	--block_size 16384 \
+	--push_to_hub
+```
+
+This should finish in about 1h30 on Cloud TPU, with validation loss and ROUGE2 score of 2.0153 and 14.64 respectively
+after 1 epoch. Training statistics can be accessed on [Models](https://huggingface.co/ydshieh/image-captioning-training-results/tensorboard).

examples/flax/image-captioning/create_model_from_encoder_decoder_models.py

+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2022 The HuggingFace Team All rights reserved.
+#
+# 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.
+"""
+Create a VisionEncoderDecoderModel instance from pretrained encoder/decoder models.
+
+The cross-attention will be randomly initialized.
+"""
+
+from dataclasses import dataclass, field
+from typing import Optional
+
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoTokenizer,
+    FlaxVisionEncoderDecoderModel,
+    HfArgumentParser,
+)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    output_dir: str = field(
+        metadata={"help": "The output directory where the model will be written."},
+    )
+    encoder_model_name_or_path: str = field(
+        metadata={
+            "help": "The encoder model checkpoint for weights initialization."
+            "Don't set if you want to train an encoder model from scratch."
+        },
+    )
+    decoder_model_name_or_path: str = field(
+        metadata={
+            "help": "The decoder model checkpoint for weights initialization."
+            "Don't set if you want to train a decoder model from scratch."
+        },
+    )
+    encoder_config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained encoder config name or path if not the same as encoder_model_name"}
+    )
+    decoder_config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained decoder config name or path if not the same as decoder_model_name"}
+    )
+
+
+def main():
+    parser = HfArgumentParser((ModelArguments,))
+    (model_args,) = parser.parse_args_into_dataclasses()
+
+    # Load pretrained model and tokenizer
+
+    # Use explicit specified encoder config
+    if model_args.encoder_config_name:
+        encoder_config = AutoConfig.from_pretrained(model_args.encoder_config_name)
+    # Use pretrained encoder model's config
+    else:
+        encoder_config = AutoConfig.from_pretrained(model_args.encoder_model_name_or_path)
+
+    # Use explicit specified decoder config
+    if model_args.decoder_config_name:
+        decoder_config = AutoConfig.from_pretrained(model_args.decoder_config_name)
+    # Use pretrained decoder model's config
+    else:
+        decoder_config = AutoConfig.from_pretrained(model_args.decoder_model_name_or_path)
+
+    # necessary for `from_encoder_decoder_pretrained` when `decoder_config` is passed
+    decoder_config.is_decoder = True
+    decoder_config.add_cross_attention = True
+
+    model = FlaxVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
+        encoder_pretrained_model_name_or_path=model_args.encoder_model_name_or_path,
+        decoder_pretrained_model_name_or_path=model_args.decoder_model_name_or_path,
+        encoder_config=encoder_config,
+        decoder_config=decoder_config,
+    )
+
+    # GPT2 only has bos/eos tokens but not decoder_start/pad tokens
+    decoder_start_token_id = decoder_config.decoder_start_token_id
+    pad_token_id = decoder_config.pad_token_id
+    if decoder_start_token_id is None:
+        decoder_start_token_id = decoder_config.bos_token_id
+    if pad_token_id is None:
+        pad_token_id = decoder_config.eos_token_id
+
+    # This is necessary to make Flax's generate() work
+    model.config.eos_token_id = decoder_config.eos_token_id
+    model.config.decoder_start_token_id = decoder_start_token_id
+    model.config.pad_token_id = pad_token_id
+
+    feature_extractor = AutoFeatureExtractor.from_pretrained(model_args.encoder_model_name_or_path)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_args.decoder_model_name_or_path)
+    tokenizer.pad_token = tokenizer.convert_ids_to_tokens(model.config.pad_token_id)
+
+    model.save_pretrained(model_args.output_dir)
+    feature_extractor.save_pretrained(model_args.output_dir)
+    tokenizer.save_pretrained(model_args.output_dir)
+
+
+if __name__ == "__main__":
+    main()

examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py

@@ -359,7 +359,7 @@ def main():
         raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
 
     if data_args.max_eval_samples is not None:
-        raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_train_samples))
+        raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
 
     def prepare_dataset(batch):
         # process audio

src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py

@@ -16,6 +16,22 @@
 
 ####################################################################################################
 
+#
+# Note: If when running this conversion script you're getting an exception:
+#     ModuleNotFoundError: No module named 'megatron.model.enums'
+# you need to tell python where to find the clone of Megatron-LM, e.g.:
+#
+# cd /tmp
+# git clone https://github.com/NVIDIA/Megatron-LM
+# PYTHONPATH=/tmp/Megatron-LM python src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py ...
+#
+# if you already have it cloned elsewhere, simply adjust the path to the existing path
+#
+# If the training was done using a Megatron-LM fork, e.g.,
+# https://github.com/microsoft/Megatron-DeepSpeed/ then chances are that you need to have that one
+# in your path, i.e., /path/to/Megatron-DeepSpeed/
+#
+
 import argparse
 import json
 import os

src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py

@@ -16,6 +16,22 @@
 
 ####################################################################################################
 
+#
+# Note: If when running this conversion script you're getting an exception:
+#     ModuleNotFoundError: No module named 'megatron.model.enums'
+# you need to tell python where to find the clone of Megatron-LM, e.g.:
+#
+# cd /tmp
+# git clone https://github.com/NVIDIA/Megatron-LM
+# PYTHONPATH=/tmp/Megatron-LM python src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py ...
+#
+# if you already have it cloned elsewhere, simply adjust the path to the existing path
+#
+# If the training was done using a Megatron-LM fork, e.g.,
+# https://github.com/microsoft/Megatron-DeepSpeed/ then chances are that you need to have that one
+# in your path, i.e., /path/to/Megatron-DeepSpeed/
+#
+
 import argparse
 import os
 import re

src/transformers/models/speech_encoder_decoder/modeling_speech_encoder_decoder.py

@@ -380,7 +380,7 @@ class SpeechEncoderDecoderModel(PreTrainedModel):
                 )
 
             if "config" not in kwargs_encoder:
-                encoder_config = AutoConfig.from_pretrained(encoder_pretrained_model_name_or_path, **kwargs_encoder)
+                encoder_config = AutoConfig.from_pretrained(encoder_pretrained_model_name_or_path)
                 if encoder_config.is_decoder is True or encoder_config.add_cross_attention is True:
                     logger.info(
                         f"Initializing {encoder_pretrained_model_name_or_path} as a encoder model "
@@ -391,7 +391,7 @@ class SpeechEncoderDecoderModel(PreTrainedModel):
 
                 kwargs_encoder["config"] = encoder_config
 
-            encoder = AutoModel.from_pretrained(encoder_pretrained_model_name_or_path, *model_args)
+            encoder = AutoModel.from_pretrained(encoder_pretrained_model_name_or_path, *model_args, **kwargs_encoder)
 
         decoder = kwargs_decoder.pop("model", None)
         if decoder is None:
@@ -402,7 +402,7 @@ class SpeechEncoderDecoderModel(PreTrainedModel):
                 )
 
             if "config" not in kwargs_decoder:
-                decoder_config = AutoConfig.from_pretrained(decoder_pretrained_model_name_or_path, **kwargs_decoder)
+                decoder_config = AutoConfig.from_pretrained(decoder_pretrained_model_name_or_path)
                 if decoder_config.is_decoder is False or decoder_config.add_cross_attention is False:
                     logger.info(
                         f"Initializing {decoder_pretrained_model_name_or_path} as a decoder model. "
@@ -424,7 +424,7 @@ class SpeechEncoderDecoderModel(PreTrainedModel):
                     "`decoder_config` to `.from_encoder_decoder_pretrained(...)`"
                 )
 
-            decoder = AutoModelForCausalLM.from_pretrained(decoder_pretrained_model_name_or_path)
+            decoder = AutoModelForCausalLM.from_pretrained(decoder_pretrained_model_name_or_path, **kwargs_decoder)
 
         # instantiate config with corresponding kwargs
         config = SpeechEncoderDecoderConfig.from_encoder_decoder_configs(encoder.config, decoder.config, **kwargs)

src/transformers/models/unispeech/modeling_unispeech.py

@@ -1290,6 +1290,7 @@ class UniSpeechForPreTraining(UniSpeechPreTrainedModel):
         >>> batch_size, raw_sequence_length = input_values.shape
         >>> sequence_length = model._get_feat_extract_output_lengths(raw_sequence_length)
         >>> mask_time_indices = _compute_mask_indices((batch_size, sequence_length), mask_prob=0.2, mask_length=2)
+        >>> mask_time_indices = torch.tensor(mask_time_indices, device=input_values.device, dtype=torch.long)
 
         >>> with torch.no_grad():
         ...     outputs = model(input_values, mask_time_indices=mask_time_indices)

src/transformers/models/unispeech_sat/modeling_unispeech_sat.py

@@ -1322,6 +1322,7 @@ class UniSpeechSatForPreTraining(UniSpeechSatPreTrainedModel):
         >>> batch_size, raw_sequence_length = input_values.shape
         >>> sequence_length = model._get_feat_extract_output_lengths(raw_sequence_length)
         >>> mask_time_indices = _compute_mask_indices((batch_size, sequence_length), mask_prob=0.2, mask_length=2)
+        >>> mask_time_indices = torch.tensor(mask_time_indices, device=input_values.device, dtype=torch.long)
 
         >>> with torch.no_grad():
         ...     outputs = model(input_values, mask_time_indices=mask_time_indices)

src/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -1460,6 +1460,7 @@ class Wav2Vec2ForPreTraining(Wav2Vec2PreTrainedModel):
         >>> batch_size, raw_sequence_length = input_values.shape
         >>> sequence_length = model._get_feat_extract_output_lengths(raw_sequence_length)
         >>> mask_time_indices = _compute_mask_indices((batch_size, sequence_length), mask_prob=0.2, mask_length=2)
+        >>> mask_time_indices = torch.tensor(mask_time_indices, device=input_values.device, dtype=torch.long)
 
         >>> with torch.no_grad():
         ...     outputs = model(input_values, mask_time_indices=mask_time_indices)

src/transformers/pipelines/base.py

@@ -742,6 +742,8 @@ class Pipeline(_ScikitCompat):
             self.model.config.update(task_specific_params.get(task))
 
         self.call_count = 0
+        self._batch_size = kwargs.pop("batch_size", None)
+        self._num_workers = kwargs.pop("num_workers", None)
         self._preprocess_params, self._forward_params, self._postprocess_params = self._sanitize_parameters(**kwargs)
 
     def save_pretrained(self, save_directory: str):
@@ -947,9 +949,21 @@ class Pipeline(_ScikitCompat):
         final_iterator = PipelineIterator(model_iterator, self.postprocess, postprocess_params)
         return final_iterator
 
-    def __call__(self, inputs, *args, num_workers=0, batch_size=1, **kwargs):
+    def __call__(self, inputs, *args, num_workers=None, batch_size=None, **kwargs):
         if args:
             logger.warning(f"Ignoring args : {args}")
+
+        if num_workers is None:
+            if self._num_workers is None:
+                num_workers = 0
+            else:
+                num_workers = self._num_workers
+        if batch_size is None:
+            if self._batch_size is None:
+                batch_size = 1
+            else:
+                batch_size = self._batch_size
+
         preprocess_params, forward_params, postprocess_params = self._sanitize_parameters(**kwargs)
 
         # Fuse __init__ params and __call__ params without modifying the __init__ ones.

src/transformers/pipelines/image_classification.py

 from typing import List, Union
 
-from ..file_utils import add_end_docstrings, is_torch_available, is_vision_available, requires_backends
+from ..file_utils import (
+    add_end_docstrings,
+    is_tf_available,
+    is_torch_available,
+    is_vision_available,
+    requires_backends,
+)
 from ..utils import logging
 from .base import PIPELINE_INIT_ARGS, Pipeline
 
@@ -10,6 +16,11 @@ if is_vision_available():
 
     from ..image_utils import load_image
 
+if is_tf_available():
+    import tensorflow as tf
+
+    from ..models.auto.modeling_tf_auto import TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
+
 if is_torch_available():
     from ..models.auto.modeling_auto import MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
 
@@ -31,12 +42,12 @@ class ImageClassificationPipeline(Pipeline):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-
-        if self.framework == "tf":
-            raise ValueError(f"The {self.__class__} is only available in PyTorch.")
-
         requires_backends(self, "vision")
-        self.check_model_type(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING)
+        self.check_model_type(
+            TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
+            if self.framework == "tf"
+            else MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
+        )
 
     def _sanitize_parameters(self, top_k=None):
         postprocess_params = {}
@@ -77,7 +88,7 @@ class ImageClassificationPipeline(Pipeline):
 
     def preprocess(self, image):
         image = load_image(image)
-        model_inputs = self.feature_extractor(images=image, return_tensors="pt")
+        model_inputs = self.feature_extractor(images=image, return_tensors=self.framework)
         return model_inputs
 
     def _forward(self, model_inputs):
@@ -87,8 +98,16 @@ class ImageClassificationPipeline(Pipeline):
     def postprocess(self, model_outputs, top_k=5):
         if top_k > self.model.config.num_labels:
             top_k = self.model.config.num_labels
-        probs = model_outputs.logits.softmax(-1)[0]
-        scores, ids = probs.topk(top_k)
+
+        if self.framework == "pt":
+            probs = model_outputs.logits.softmax(-1)[0]
+            scores, ids = probs.topk(top_k)
+        elif self.framework == "tf":
+            probs = tf.nn.softmax(model_outputs.logits, axis=-1)[0]
+            topk = tf.math.top_k(probs, k=top_k)
+            scores, ids = topk.values.numpy(), topk.indices.numpy()
+        else:
+            raise ValueError(f"Unsupported framework: {self.framework}")
 
         scores = scores.tolist()
         ids = ids.tolist()

tests/test_modeling_bert_generation.py

@@ -307,7 +307,8 @@ class BertGenerationEncoderIntegrationTest(unittest.TestCase):
     def test_inference_no_head_absolute_embedding(self):
         model = BertGenerationEncoder.from_pretrained("google/bert_for_seq_generation_L-24_bbc_encoder")
         input_ids = torch.tensor([[101, 7592, 1010, 2026, 3899, 2003, 10140, 102]])
-        output = model(input_ids)[0]
+        with torch.no_grad():
+            output = model(input_ids)[0]
         expected_shape = torch.Size([1, 8, 1024])
         self.assertEqual(output.shape, expected_shape)
         expected_slice = torch.tensor(
@@ -322,7 +323,8 @@ class BertGenerationDecoderIntegrationTest(unittest.TestCase):
     def test_inference_no_head_absolute_embedding(self):
         model = BertGenerationDecoder.from_pretrained("google/bert_for_seq_generation_L-24_bbc_encoder")
         input_ids = torch.tensor([[101, 7592, 1010, 2026, 3899, 2003, 10140, 102]])
-        output = model(input_ids)[0]
+        with torch.no_grad():
+            output = model(input_ids)[0]
         expected_shape = torch.Size([1, 8, 50358])
         self.assertEqual(output.shape, expected_shape)
         expected_slice = torch.tensor(

tests/test_modeling_clip.py

@@ -858,6 +858,7 @@ def prepare_img():
 
 
 @require_vision
+@require_torch
 class CLIPModelIntegrationTest(unittest.TestCase):
     @slow
     def test_inference(self):

tests/test_modeling_layoutlmv2.py

@@ -478,11 +478,11 @@ class LayoutLMv2ModelTest(ModelTesterMixin, unittest.TestCase):
 def prepare_layoutlmv2_batch_inputs():
     # Here we prepare a batch of 2 sequences to test a LayoutLMv2 forward pass on:
     # fmt: off
-    input_ids = torch.tensor([[101,1019,1014,1016,1037,12849,4747,1004,14246,2278,5439,4524,5002,2930,2193,2930,4341,3208,1005,1055,2171,2848,11300,3531,102],[101,4070,4034,7020,1024,3058,1015,1013,2861,1013,6070,19274,2772,6205,27814,16147,16147,4343,2047,10283,10969,14389,1012,2338,102]],device=torch_device)  # noqa: E231
-    bbox = torch.tensor([[[0,0,0,0],[423,237,440,251],[427,272,441,287],[419,115,437,129],[961,885,992,912],[256,38,330,58],[256,38,330,58],[336,42,353,57],[360,39,401,56],[360,39,401,56],[411,39,471,59],[479,41,528,59],[533,39,630,60],[67,113,134,131],[141,115,209,132],[68,149,133,166],[141,149,187,164],[195,148,287,165],[195,148,287,165],[195,148,287,165],[295,148,349,165],[441,149,492,166],[497,149,546,164],[64,201,125,218],[1000,1000,1000,1000]],[[0,0,0,0],[662,150,754,166],[665,199,742,211],[519,213,554,228],[519,213,554,228],[134,433,187,454],[130,467,204,480],[130,467,204,480],[130,467,204,480],[130,467,204,480],[130,467,204,480],[314,469,376,482],[504,684,582,706],[941,825,973,900],[941,825,973,900],[941,825,973,900],[941,825,973,900],[610,749,652,765],[130,659,168,672],[176,657,237,672],[238,657,312,672],[443,653,628,672],[443,653,628,672],[716,301,825,317],[1000,1000,1000,1000]]],device=torch_device)  # noqa: E231
+    input_ids = torch.tensor([[101,1019,1014,1016,1037,12849,4747,1004,14246,2278,5439,4524,5002,2930,2193,2930,4341,3208,1005,1055,2171,2848,11300,3531,102],[101,4070,4034,7020,1024,3058,1015,1013,2861,1013,6070,19274,2772,6205,27814,16147,16147,4343,2047,10283,10969,14389,1012,2338,102]])  # noqa: E231
+    bbox = torch.tensor([[[0,0,0,0],[423,237,440,251],[427,272,441,287],[419,115,437,129],[961,885,992,912],[256,38,330,58],[256,38,330,58],[336,42,353,57],[360,39,401,56],[360,39,401,56],[411,39,471,59],[479,41,528,59],[533,39,630,60],[67,113,134,131],[141,115,209,132],[68,149,133,166],[141,149,187,164],[195,148,287,165],[195,148,287,165],[195,148,287,165],[295,148,349,165],[441,149,492,166],[497,149,546,164],[64,201,125,218],[1000,1000,1000,1000]],[[0,0,0,0],[662,150,754,166],[665,199,742,211],[519,213,554,228],[519,213,554,228],[134,433,187,454],[130,467,204,480],[130,467,204,480],[130,467,204,480],[130,467,204,480],[130,467,204,480],[314,469,376,482],[504,684,582,706],[941,825,973,900],[941,825,973,900],[941,825,973,900],[941,825,973,900],[610,749,652,765],[130,659,168,672],[176,657,237,672],[238,657,312,672],[443,653,628,672],[443,653,628,672],[716,301,825,317],[1000,1000,1000,1000]]])  # noqa: E231
     image = ImageList(torch.randn((2,3,224,224)), image_sizes=[(224,224), (224,224)])  # noqa: E231
-    attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],],device=torch_device)  # noqa: E231
-    token_type_ids = torch.tensor([[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]],device=torch_device)  # noqa: E231
+    attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],])  # noqa: E231
+    token_type_ids = torch.tensor([[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]])  # noqa: E231
     # fmt: on
 
     return input_ids, bbox, image, attention_mask, token_type_ids
@@ -505,11 +505,11 @@ class LayoutLMv2ModelIntegrationTest(unittest.TestCase):
 
         # forward pass
         outputs = model(
-            input_ids=input_ids,
-            bbox=bbox,
-            image=image,
-            attention_mask=attention_mask,
-            token_type_ids=token_type_ids,
+            input_ids=input_ids.to(torch_device),
+            bbox=bbox.to(torch_device),
+            image=image.to(torch_device),
+            attention_mask=attention_mask.to(torch_device),
+            token_type_ids=token_type_ids.to(torch_device),
         )
 
         # verify the sequence output

tests/test_modeling_roberta.py

@@ -485,7 +485,8 @@ class RobertaModelIntegrationTest(TestCasePlus):
         model = RobertaForMaskedLM.from_pretrained("roberta-base")
 
         input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
-        output = model(input_ids)[0]
+        with torch.no_grad():
+            output = model(input_ids)[0]
         expected_shape = torch.Size((1, 11, 50265))
         self.assertEqual(output.shape, expected_shape)
         # compare the actual values for a slice.
@@ -504,7 +505,8 @@ class RobertaModelIntegrationTest(TestCasePlus):
         model = RobertaModel.from_pretrained("roberta-base")
 
         input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
-        output = model(input_ids)[0]
+        with torch.no_grad():
+            output = model(input_ids)[0]
         # compare the actual values for a slice.
         expected_slice = torch.tensor(
             [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0540, -0.0175], [0.0548, 0.0799, 0.1687]]]
@@ -521,7 +523,8 @@ class RobertaModelIntegrationTest(TestCasePlus):
         model = RobertaForSequenceClassification.from_pretrained("roberta-large-mnli")
 
         input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
-        output = model(input_ids)[0]
+        with torch.no_grad():
+            output = model(input_ids)[0]
         expected_shape = torch.Size((1, 3))
         self.assertEqual(output.shape, expected_shape)
         expected_tensor = torch.tensor([[-0.9469, 0.3913, 0.5118]])

tests/test_modeling_tf_clip.py

@@ -254,7 +254,7 @@ class TFCLIPVisionModelTest(TFModelTesterMixin, unittest.TestCase):
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
-            model = TFCLIPVisionModel.from_pretrained(model_name, from_pt=True)
+            model = TFCLIPVisionModel.from_pretrained(model_name)
             self.assertIsNotNone(model)
 
 
@@ -359,7 +359,7 @@ class TFCLIPTextModelTest(TFModelTesterMixin, unittest.TestCase):
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
-            model = TFCLIPTextModel.from_pretrained(model_name, from_pt=True)
+            model = TFCLIPTextModel.from_pretrained(model_name)
             self.assertIsNotNone(model)
 
 
@@ -618,7 +618,7 @@ class TFCLIPModelTest(TFModelTesterMixin, unittest.TestCase):
     @slow
     def test_model_from_pretrained(self):
         for model_name in TF_CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
-            model = TFCLIPModel.from_pretrained(model_name, from_pt=True)
+            model = TFCLIPModel.from_pretrained(model_name)
             self.assertIsNotNone(model)
 
 
@@ -630,11 +630,12 @@ def prepare_img():
 
 
 @require_vision
+@require_tf
 class TFCLIPModelIntegrationTest(unittest.TestCase):
     @slow
     def test_inference(self):
         model_name = "openai/clip-vit-base-patch32"
-        model = TFCLIPModel.from_pretrained(model_name, from_pt=True)
+        model = TFCLIPModel.from_pretrained(model_name)
         processor = CLIPProcessor.from_pretrained(model_name)
 
         image = prepare_img()

tests/test_pipelines_common.py

@@ -299,6 +299,16 @@ class CommonPipelineTest(unittest.TestCase):
 
         self.assertIsInstance(pipe, TextClassificationPipeline)
 
+    @require_torch
+    def test_pipeline_batch_size_global(self):
+        pipe = pipeline(model="hf-internal-testing/tiny-random-distilbert")
+        self.assertEqual(pipe._batch_size, None)
+        self.assertEqual(pipe._num_workers, None)
+
+        pipe = pipeline(model="hf-internal-testing/tiny-random-distilbert", batch_size=2, num_workers=1)
+        self.assertEqual(pipe._batch_size, 2)
+        self.assertEqual(pipe._num_workers, 1)
+
     @require_torch
     def test_pipeline_override(self):
         class MyPipeline(TextClassificationPipeline):

tests/test_pipelines_image_classification.py

@@ -14,7 +14,12 @@
 
 import unittest
 
-from transformers import MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING, PreTrainedTokenizer, is_vision_available
+from transformers import (
+    MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
+    TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
+    PreTrainedTokenizer,
+    is_vision_available,
+)
 from transformers.pipelines import ImageClassificationPipeline, pipeline
 from transformers.testing_utils import (
     is_pipeline_test,
@@ -40,9 +45,9 @@ else:
 
 @is_pipeline_test
 @require_vision
-@require_torch
 class ImageClassificationPipelineTests(unittest.TestCase, metaclass=PipelineTestCaseMeta):
     model_mapping = MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
+    tf_model_mapping = TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING
 
     def get_test_pipeline(self, model, tokenizer, feature_extractor):
         image_classifier = ImageClassificationPipeline(model=model, feature_extractor=feature_extractor, top_k=2)
@@ -145,9 +150,42 @@ class ImageClassificationPipelineTests(unittest.TestCase, metaclass=PipelineTest
         )
 
     @require_tf
-    @unittest.skip("Image classification is not implemented for TF")
     def test_small_model_tf(self):
-        pass
+        small_model = "lysandre/tiny-vit-random"
+        image_classifier = pipeline("image-classification", model=small_model)
+
+        outputs = image_classifier("http://images.cocodataset.org/val2017/000000039769.jpg")
+        self.assertEqual(
+            nested_simplify(outputs, decimals=4),
+            [
+                {"score": 0.0015, "label": "chambered nautilus, pearly nautilus, nautilus"},
+                {"score": 0.0015, "label": "pajama, pyjama, pj's, jammies"},
+                {"score": 0.0014, "label": "trench coat"},
+                {"score": 0.0014, "label": "handkerchief, hankie, hanky, hankey"},
+                {"score": 0.0014, "label": "baboon"},
+            ],
+        )
+
+        outputs = image_classifier(
+            [
+                "http://images.cocodataset.org/val2017/000000039769.jpg",
+                "http://images.cocodataset.org/val2017/000000039769.jpg",
+            ],
+            top_k=2,
+        )
+        self.assertEqual(
+            nested_simplify(outputs, decimals=4),
+            [
+                [
+                    {"score": 0.0015, "label": "chambered nautilus, pearly nautilus, nautilus"},
+                    {"score": 0.0015, "label": "pajama, pyjama, pj's, jammies"},
+                ],
+                [
+                    {"score": 0.0015, "label": "chambered nautilus, pearly nautilus, nautilus"},
+                    {"score": 0.0015, "label": "pajama, pyjama, pj's, jammies"},
+                ],
+            ],
+        )
 
     def test_custom_tokenizer(self):
         tokenizer = PreTrainedTokenizer()