Transformers
  • 🌍GET STARTED
    • Transformers
    • Quick tour
    • Installation
  • 🌍TUTORIALS
    • Run inference with pipelines
    • Write portable code with AutoClass
    • Preprocess data
    • Fine-tune a pretrained model
    • Train with a script
    • Set up distributed training with BOINC AI Accelerate
    • Load and train adapters with BOINC AI PEFT
    • Share your model
    • Agents
    • Generation with LLMs
  • 🌍TASK GUIDES
    • 🌍NATURAL LANGUAGE PROCESSING
      • Text classification
      • Token classification
      • Question answering
      • Causal language modeling
      • Masked language modeling
      • Translation
      • Summarization
      • Multiple choice
    • 🌍AUDIO
      • Audio classification
      • Automatic speech recognition
    • 🌍COMPUTER VISION
      • Image classification
      • Semantic segmentation
      • Video classification
      • Object detection
      • Zero-shot object detection
      • Zero-shot image classification
      • Depth estimation
    • 🌍MULTIMODAL
      • Image captioning
      • Document Question Answering
      • Visual Question Answering
      • Text to speech
    • 🌍GENERATION
      • Customize the generation strategy
    • 🌍PROMPTING
      • Image tasks with IDEFICS
  • 🌍DEVELOPER GUIDES
    • Use fast tokenizers from BOINC AI Tokenizers
    • Run inference with multilingual models
    • Use model-specific APIs
    • Share a custom model
    • Templates for chat models
    • Run training on Amazon SageMaker
    • Export to ONNX
    • Export to TFLite
    • Export to TorchScript
    • Benchmarks
    • Notebooks with examples
    • Community resources
    • Custom Tools and Prompts
    • Troubleshoot
  • 🌍PERFORMANCE AND SCALABILITY
    • Overview
    • 🌍EFFICIENT TRAINING TECHNIQUES
      • Methods and tools for efficient training on a single GPU
      • Multiple GPUs and parallelism
      • Efficient training on CPU
      • Distributed CPU training
      • Training on TPUs
      • Training on TPU with TensorFlow
      • Training on Specialized Hardware
      • Custom hardware for training
      • Hyperparameter Search using Trainer API
    • 🌍OPTIMIZING INFERENCE
      • Inference on CPU
      • Inference on one GPU
      • Inference on many GPUs
      • Inference on Specialized Hardware
    • Instantiating a big model
    • Troubleshooting
    • XLA Integration for TensorFlow Models
    • Optimize inference using `torch.compile()`
  • 🌍CONTRIBUTE
    • How to contribute to transformers?
    • How to add a model to BOINC AI Transformers?
    • How to convert a BOINC AI Transformers model to TensorFlow?
    • How to add a pipeline to BOINC AI Transformers?
    • Testing
    • Checks on a Pull Request
  • 🌍CONCEPTUAL GUIDES
    • Philosophy
    • Glossary
    • What BOINC AI Transformers can do
    • How BOINC AI Transformers solve tasks
    • The Transformer model family
    • Summary of the tokenizers
    • Attention mechanisms
    • Padding and truncation
    • BERTology
    • Perplexity of fixed-length models
    • Pipelines for webserver inference
    • Model training anatomy
  • 🌍API
    • 🌍MAIN CLASSES
      • Agents and Tools
      • 🌍Auto Classes
        • Extending the Auto Classes
        • AutoConfig
        • AutoTokenizer
        • AutoFeatureExtractor
        • AutoImageProcessor
        • AutoProcessor
        • Generic model classes
          • AutoModel
          • TFAutoModel
          • FlaxAutoModel
        • Generic pretraining classes
          • AutoModelForPreTraining
          • TFAutoModelForPreTraining
          • FlaxAutoModelForPreTraining
        • Natural Language Processing
          • AutoModelForCausalLM
          • TFAutoModelForCausalLM
          • FlaxAutoModelForCausalLM
          • AutoModelForMaskedLM
          • TFAutoModelForMaskedLM
          • FlaxAutoModelForMaskedLM
          • AutoModelForMaskGenerationge
          • TFAutoModelForMaskGeneration
          • AutoModelForSeq2SeqLM
          • TFAutoModelForSeq2SeqLM
          • FlaxAutoModelForSeq2SeqLM
          • AutoModelForSequenceClassification
          • TFAutoModelForSequenceClassification
          • FlaxAutoModelForSequenceClassification
          • AutoModelForMultipleChoice
          • TFAutoModelForMultipleChoice
          • FlaxAutoModelForMultipleChoice
          • AutoModelForNextSentencePrediction
          • TFAutoModelForNextSentencePrediction
          • FlaxAutoModelForNextSentencePrediction
          • AutoModelForTokenClassification
          • TFAutoModelForTokenClassification
          • FlaxAutoModelForTokenClassification
          • AutoModelForQuestionAnswering
          • TFAutoModelForQuestionAnswering
          • FlaxAutoModelForQuestionAnswering
          • AutoModelForTextEncoding
          • TFAutoModelForTextEncoding
        • Computer vision
          • AutoModelForDepthEstimation
          • AutoModelForImageClassification
          • TFAutoModelForImageClassification
          • FlaxAutoModelForImageClassification
          • AutoModelForVideoClassification
          • AutoModelForMaskedImageModeling
          • TFAutoModelForMaskedImageModeling
          • AutoModelForObjectDetection
          • AutoModelForImageSegmentation
          • AutoModelForImageToImage
          • AutoModelForSemanticSegmentation
          • TFAutoModelForSemanticSegmentation
          • AutoModelForInstanceSegmentation
          • AutoModelForUniversalSegmentation
          • AutoModelForZeroShotImageClassification
          • TFAutoModelForZeroShotImageClassification
          • AutoModelForZeroShotObjectDetection
        • Audio
          • AutoModelForAudioClassification
          • AutoModelForAudioFrameClassification
          • TFAutoModelForAudioFrameClassification
          • AutoModelForCTC
          • AutoModelForSpeechSeq2Seq
          • TFAutoModelForSpeechSeq2Seq
          • FlaxAutoModelForSpeechSeq2Seq
          • AutoModelForAudioXVector
          • AutoModelForTextToSpectrogram
          • AutoModelForTextToWaveform
        • Multimodal
          • AutoModelForTableQuestionAnswering
          • TFAutoModelForTableQuestionAnswering
          • AutoModelForDocumentQuestionAnswering
          • TFAutoModelForDocumentQuestionAnswering
          • AutoModelForVisualQuestionAnswering
          • AutoModelForVision2Seq
          • TFAutoModelForVision2Seq
          • FlaxAutoModelForVision2Seq
      • Callbacks
      • Configuration
      • Data Collator
      • Keras callbacks
      • Logging
      • Models
      • Text Generation
      • ONNX
      • Optimization
      • Model outputs
      • Pipelines
      • Processors
      • Quantization
      • Tokenizer
      • Trainer
      • DeepSpeed Integration
      • Feature Extractor
      • Image Processor
    • 🌍MODELS
      • 🌍TEXT MODELS
        • ALBERT
        • BART
        • BARThez
        • BARTpho
        • BERT
        • BertGeneration
        • BertJapanese
        • Bertweet
        • BigBird
        • BigBirdPegasus
        • BioGpt
        • Blenderbot
        • Blenderbot Small
        • BLOOM
        • BORT
        • ByT5
        • CamemBERT
        • CANINE
        • CodeGen
        • CodeLlama
        • ConvBERT
        • CPM
        • CPMANT
        • CTRL
        • DeBERTa
        • DeBERTa-v2
        • DialoGPT
        • DistilBERT
        • DPR
        • ELECTRA
        • Encoder Decoder Models
        • ERNIE
        • ErnieM
        • ESM
        • Falcon
        • FLAN-T5
        • FLAN-UL2
        • FlauBERT
        • FNet
        • FSMT
        • Funnel Transformer
        • GPT
        • GPT Neo
        • GPT NeoX
        • GPT NeoX Japanese
        • GPT-J
        • GPT2
        • GPTBigCode
        • GPTSAN Japanese
        • GPTSw3
        • HerBERT
        • I-BERT
        • Jukebox
        • LED
        • LLaMA
        • LLama2
        • Longformer
        • LongT5
        • LUKE
        • M2M100
        • MarianMT
        • MarkupLM
        • MBart and MBart-50
        • MEGA
        • MegatronBERT
        • MegatronGPT2
        • Mistral
        • mLUKE
        • MobileBERT
        • MPNet
        • MPT
        • MRA
        • MT5
        • MVP
        • NEZHA
        • NLLB
        • NLLB-MoE
        • Nyströmformer
        • Open-Llama
        • OPT
        • Pegasus
        • PEGASUS-X
        • Persimmon
        • PhoBERT
        • PLBart
        • ProphetNet
        • QDQBert
        • RAG
        • REALM
        • Reformer
        • RemBERT
        • RetriBERT
        • RoBERTa
        • RoBERTa-PreLayerNorm
        • RoCBert
        • RoFormer
        • RWKV
        • Splinter
        • SqueezeBERT
        • SwitchTransformers
        • T5
        • T5v1.1
        • TAPEX
        • Transformer XL
        • UL2
        • UMT5
        • X-MOD
        • XGLM
        • XLM
        • XLM-ProphetNet
        • XLM-RoBERTa
        • XLM-RoBERTa-XL
        • XLM-V
        • XLNet
        • YOSO
      • 🌍VISION MODELS
        • BEiT
        • BiT
        • Conditional DETR
        • ConvNeXT
        • ConvNeXTV2
        • CvT
        • Deformable DETR
        • DeiT
        • DETA
        • DETR
        • DiNAT
        • DINO V2
        • DiT
        • DPT
        • EfficientFormer
        • EfficientNet
        • FocalNet
        • GLPN
        • ImageGPT
        • LeViT
        • Mask2Former
        • MaskFormer
        • MobileNetV1
        • MobileNetV2
        • MobileViT
        • MobileViTV2
        • NAT
        • PoolFormer
        • Pyramid Vision Transformer (PVT)
        • RegNet
        • ResNet
        • SegFormer
        • SwiftFormer
        • Swin Transformer
        • Swin Transformer V2
        • Swin2SR
        • Table Transformer
        • TimeSformer
        • UperNet
        • VAN
        • VideoMAE
        • Vision Transformer (ViT)
        • ViT Hybrid
        • ViTDet
        • ViTMAE
        • ViTMatte
        • ViTMSN
        • ViViT
        • YOLOS
      • 🌍AUDIO MODELS
        • Audio Spectrogram Transformer
        • Bark
        • CLAP
        • EnCodec
        • Hubert
        • MCTCT
        • MMS
        • MusicGen
        • Pop2Piano
        • SEW
        • SEW-D
        • Speech2Text
        • Speech2Text2
        • SpeechT5
        • UniSpeech
        • UniSpeech-SAT
        • VITS
        • Wav2Vec2
        • Wav2Vec2-Conformer
        • Wav2Vec2Phoneme
        • WavLM
        • Whisper
        • XLS-R
        • XLSR-Wav2Vec2
      • 🌍MULTIMODAL MODELS
        • ALIGN
        • AltCLIP
        • BLIP
        • BLIP-2
        • BridgeTower
        • BROS
        • Chinese-CLIP
        • CLIP
        • CLIPSeg
        • Data2Vec
        • DePlot
        • Donut
        • FLAVA
        • GIT
        • GroupViT
        • IDEFICS
        • InstructBLIP
        • LayoutLM
        • LayoutLMV2
        • LayoutLMV3
        • LayoutXLM
        • LiLT
        • LXMERT
        • MatCha
        • MGP-STR
        • Nougat
        • OneFormer
        • OWL-ViT
        • Perceiver
        • Pix2Struct
        • Segment Anything
        • Speech Encoder Decoder Models
        • TAPAS
        • TrOCR
        • TVLT
        • ViLT
        • Vision Encoder Decoder Models
        • Vision Text Dual Encoder
        • VisualBERT
        • X-CLIP
      • 🌍REINFORCEMENT LEARNING MODELS
        • Decision Transformer
        • Trajectory Transformer
      • 🌍TIME SERIES MODELS
        • Autoformer
        • Informer
        • Time Series Transformer
      • 🌍GRAPH MODELS
        • Graphormer
  • 🌍INTERNAL HELPERS
    • Custom Layers and Utilities
    • Utilities for pipelines
    • Utilities for Tokenizers
    • Utilities for Trainer
    • Utilities for Generation
    • Utilities for Image Processors
    • Utilities for Audio processing
    • General Utilities
    • Utilities for Time Series
Powered by GitBook
On this page
  • Preprocess
  • Natural Language Processing
  • Audio
  • Computer vision
  • Multimodal
  1. TUTORIALS

Preprocess data

PreviousWrite portable code with AutoClassNextFine-tune a pretrained model

Last updated 1 year ago

Preprocess

Before you can train a model on a dataset, it needs to be preprocessed into the expected model input format. Whether your data is text, images, or audio, they need to be converted and assembled into batches of tensors. 🌎 Transformers provides a set of preprocessing classes to help prepare your data for the model. In this tutorial, you’ll learn that for:

  • Text, use a to convert text into a sequence of tokens, create a numerical representation of the tokens, and assemble them into tensors.

  • Speech and audio, use a to extract sequential features from audio waveforms and convert them into tensors.

  • Image inputs use a to convert images into tensors.

  • Multimodal inputs, use a to combine a tokenizer and a feature extractor or image processor.

AutoProcessor always works and automatically chooses the correct class for the model you’re using, whether you’re using a tokenizer, image processor, feature extractor or processor.

Before you begin, install 🌎 Datasets so you can load some datasets to experiment with:

Copied

pip install datasets

Natural Language Processing

The main tool for preprocessing textual data is a . A tokenizer splits text into tokens according to a set of rules. The tokens are converted into numbers and then tensors, which become the model inputs. Any additional inputs required by the model are added by the tokenizer.

If you plan on using a pretrained model, it’s important to use the associated pretrained tokenizer. This ensures the text is split the same way as the pretraining corpus, and uses the same corresponding tokens-to-index (usually referred to as the vocab) during pretraining.

Get started by loading a pretrained tokenizer with the method. This downloads the vocab a model was pretrained with:

Copied

>>> from transformers import AutoTokenizer

>>> tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")

Then pass your text to the tokenizer:

Copied

>>> encoded_input = tokenizer("Do not meddle in the affairs of wizards, for they are subtle and quick to anger.")
>>> print(encoded_input)
{'input_ids': [101, 2079, 2025, 19960, 10362, 1999, 1996, 3821, 1997, 16657, 1010, 2005, 2027, 2024, 11259, 1998, 4248, 2000, 4963, 1012, 102],
 'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}

The tokenizer returns a dictionary with three important items:

Return your input by decoding the input_ids:

Copied

>>> tokenizer.decode(encoded_input["input_ids"])
'[CLS] Do not meddle in the affairs of wizards, for they are subtle and quick to anger. [SEP]'

As you can see, the tokenizer added two special tokens - CLS and SEP (classifier and separator) - to the sentence. Not all models need special tokens, but if they do, the tokenizer automatically adds them for you.

If there are several sentences you want to preprocess, pass them as a list to the tokenizer:

Copied

>>> batch_sentences = [
...     "But what about second breakfast?",
...     "Don't think he knows about second breakfast, Pip.",
...     "What about elevensies?",
... ]
>>> encoded_inputs = tokenizer(batch_sentences)
>>> print(encoded_inputs)
{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102],
               [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
               [101, 1327, 1164, 5450, 23434, 136, 102]],
 'token_type_ids': [[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]],
 'attention_mask': [[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]]}

Pad

Sentences aren’t always the same length which can be an issue because tensors, the model inputs, need to have a uniform shape. Padding is a strategy for ensuring tensors are rectangular by adding a special padding token to shorter sentences.

Set the padding parameter to True to pad the shorter sequences in the batch to match the longest sequence:

Copied

>>> batch_sentences = [
...     "But what about second breakfast?",
...     "Don't think he knows about second breakfast, Pip.",
...     "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch_sentences, padding=True)
>>> print(encoded_input)
{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
               [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
               [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],
 'token_type_ids': [[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]],
 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
                    [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}

The first and third sentences are now padded with 0’s because they are shorter.

Truncation

On the other end of the spectrum, sometimes a sequence may be too long for a model to handle. In this case, you’ll need to truncate the sequence to a shorter length.

Set the truncation parameter to True to truncate a sequence to the maximum length accepted by the model:

Copied

>>> batch_sentences = [
...     "But what about second breakfast?",
...     "Don't think he knows about second breakfast, Pip.",
...     "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch_sentences, padding=True, truncation=True)
>>> print(encoded_input)
{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
               [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
               [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],
 'token_type_ids': [[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]],
 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
                    [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}

Build tensors

Finally, you want the tokenizer to return the actual tensors that get fed to the model.

Set the return_tensors parameter to either pt for PyTorch, or tf for TensorFlow:

PytorchHide Pytorch contentCopied

>>> batch_sentences = [
...     "But what about second breakfast?",
...     "Don't think he knows about second breakfast, Pip.",
...     "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="pt")
>>> print(encoded_input)
{'input_ids': tensor([[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
                      [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
                      [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]]),
 'token_type_ids': 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]]),
 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
                           [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
                           [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]])}

TensorFlowHide TensorFlow contentCopied

>>> batch_sentences = [
...     "But what about second breakfast?",
...     "Don't think he knows about second breakfast, Pip.",
...     "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="tf")
>>> print(encoded_input)
{'input_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
array([[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
       [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
       [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],
      dtype=int32)>,
 'token_type_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
array([[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]], dtype=int32)>,
 'attention_mask': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
array([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
       [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
       [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>}

Audio

Copied

>>> from datasets import load_dataset, Audio

>>> dataset = load_dataset("PolyAI/minds14", name="en-US", split="train")

Access the first element of the audio column to take a look at the input. Calling the audio column automatically loads and resamples the audio file:

Copied

>>> dataset[0]["audio"]
{'array': array([ 0.        ,  0.00024414, -0.00024414, ..., -0.00024414,
         0.        ,  0.        ], dtype=float32),
 'path': '/root/.cache/boincai/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',
 'sampling_rate': 8000}

This returns three items:

  • array is the speech signal loaded - and potentially resampled - as a 1D array.

  • path points to the location of the audio file.

  • sampling_rate refers to how many data points in the speech signal are measured per second.

Copied

>>> dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
  1. Call the audio column again to resample the audio file:

Copied

>>> dataset[0]["audio"]
{'array': array([ 2.3443763e-05,  2.1729663e-04,  2.2145823e-04, ...,
         3.8356509e-05, -7.3497440e-06, -2.1754686e-05], dtype=float32),
 'path': '/root/.cache/boincai/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',
 'sampling_rate': 16000}

Next, load a feature extractor to normalize and pad the input. When padding textual data, a 0 is added for shorter sequences. The same idea applies to audio data. The feature extractor adds a 0 - interpreted as silence - to array.

Copied

>>> from transformers import AutoFeatureExtractor

>>> feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base")

Pass the audio array to the feature extractor. We also recommend adding the sampling_rate argument in the feature extractor in order to better debug any silent errors that may occur.

Copied

>>> audio_input = [dataset[0]["audio"]["array"]]
>>> feature_extractor(audio_input, sampling_rate=16000)
{'input_values': [array([ 3.8106556e-04,  2.7506407e-03,  2.8015103e-03, ...,
        5.6335266e-04,  4.6588284e-06, -1.7142107e-04], dtype=float32)]}

Just like the tokenizer, you can apply padding or truncation to handle variable sequences in a batch. Take a look at the sequence length of these two audio samples:

Copied

>>> dataset[0]["audio"]["array"].shape
(173398,)

>>> dataset[1]["audio"]["array"].shape
(106496,)

Create a function to preprocess the dataset so the audio samples are the same lengths. Specify a maximum sample length, and the feature extractor will either pad or truncate the sequences to match it:

Copied

>>> def preprocess_function(examples):
...     audio_arrays = [x["array"] for x in examples["audio"]]
...     inputs = feature_extractor(
...         audio_arrays,
...         sampling_rate=16000,
...         padding=True,
...         max_length=100000,
...         truncation=True,
...     )
...     return inputs

Apply the preprocess_function to the the first few examples in the dataset:

Copied

>>> processed_dataset = preprocess_function(dataset[:5])

The sample lengths are now the same and match the specified maximum length. You can pass your processed dataset to the model now!

Copied

>>> processed_dataset["input_values"][0].shape
(100000,)

>>> processed_dataset["input_values"][1].shape
(100000,)

Computer vision

Image preprocessing often follows some form of image augmentation. Both image preprocessing and image augmentation transform image data, but they serve different purposes:

  • Image augmentation alters images in a way that can help prevent overfitting and increase the robustness of the model. You can get creative in how you augment your data - adjust brightness and colors, crop, rotate, resize, zoom, etc. However, be mindful not to change the meaning of the images with your augmentations.

  • Image preprocessing guarantees that the images match the model’s expected input format. When fine-tuning a computer vision model, images must be preprocessed exactly as when the model was initially trained.

You can use any library you like for image augmentation. For image preprocessing, use the ImageProcessor associated with the model.

Use 🌎 Datasets split parameter to only load a small sample from the training split since the dataset is quite large!

Copied

>>> from datasets import load_dataset

>>> dataset = load_dataset("food101", split="train[:100]")

Copied

>>> dataset[0]["image"]

Copied

>>> from transformers import AutoImageProcessor

>>> image_processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")

Copied

>>> from torchvision.transforms import RandomResizedCrop, ColorJitter, Compose

>>> size = (
...     image_processor.size["shortest_edge"]
...     if "shortest_edge" in image_processor.size
...     else (image_processor.size["height"], image_processor.size["width"])
... )

>>> _transforms = Compose([RandomResizedCrop(size), ColorJitter(brightness=0.5, hue=0.5)])

Copied

>>> def transforms(examples):
...     images = [_transforms(img.convert("RGB")) for img in examples["image"]]
...     examples["pixel_values"] = image_processor(images, do_resize=False, return_tensors="pt")["pixel_values"]
...     return examples

In the example above we set do_resize=False because we have already resized the images in the image augmentation transformation, and leveraged the size attribute from the appropriate image_processor. If you do not resize images during image augmentation, leave this parameter out. By default, ImageProcessor will handle the resizing.

If you wish to normalize images as a part of the augmentation transformation, use the image_processor.image_mean, and image_processor.image_std values.

Copied

>>> dataset.set_transform(transforms)
  1. Now when you access the image, you’ll notice the image processor has added pixel_values. You can pass your processed dataset to the model now!

Copied

>>> dataset[0].keys()

Here is what the image looks like after the transforms are applied. The image has been randomly cropped and it’s color properties are different.

Copied

>>> import numpy as np
>>> import matplotlib.pyplot as plt

>>> img = dataset[0]["pixel_values"]
>>> plt.imshow(img.permute(1, 2, 0))

For tasks like object detection, semantic segmentation, instance segmentation, and panoptic segmentation, ImageProcessor offers post processing methods. These methods convert model’s raw outputs into meaningful predictions such as bounding boxes, or segmentation maps.

Pad

Copied

>>> def collate_fn(batch):
...     pixel_values = [item["pixel_values"] for item in batch]
...     encoding = image_processor.pad(pixel_values, return_tensors="pt")
...     labels = [item["labels"] for item in batch]
...     batch = {}
...     batch["pixel_values"] = encoding["pixel_values"]
...     batch["pixel_mask"] = encoding["pixel_mask"]
...     batch["labels"] = labels
...     return batch

Multimodal

Copied

>>> from datasets import load_dataset

>>> lj_speech = load_dataset("lj_speech", split="train")

For ASR, you’re mainly focused on audio and text so you can remove the other columns:

Copied

>>> lj_speech = lj_speech.map(remove_columns=["file", "id", "normalized_text"])

Now take a look at the audio and text columns:

Copied

>>> lj_speech[0]["audio"]
{'array': array([-7.3242188e-04, -7.6293945e-04, -6.4086914e-04, ...,
         7.3242188e-04,  2.1362305e-04,  6.1035156e-05], dtype=float32),
 'path': '/root/.cache/boincai/datasets/downloads/extracted/917ece08c95cf0c4115e45294e3cd0dee724a1165b7fc11798369308a465bd26/LJSpeech-1.1/wavs/LJ001-0001.wav',
 'sampling_rate': 22050}

>>> lj_speech[0]["text"]
'Printing, in the only sense with which we are at present concerned, differs from most if not from all the arts and crafts represented in the Exhibition'

Copied

>>> lj_speech = lj_speech.cast_column("audio", Audio(sampling_rate=16_000))

Copied

>>> from transformers import AutoProcessor

>>> processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base-960h")
  1. Create a function to process the audio data contained in array to input_values, and tokenize text to labels. These are the inputs to the model:

Copied

>>> def prepare_dataset(example):
...     audio = example["audio"]

...     example.update(processor(audio=audio["array"], text=example["text"], sampling_rate=16000))

...     return example
  1. Apply the prepare_dataset function to a sample:

Copied

>>> prepare_dataset(lj_speech[0])

The processor has now added input_values and labels, and the sampling rate has also been correctly downsampled to 16kHz. You can pass your processed dataset to the model now!

are the indices corresponding to each token in the sentence.

indicates whether a token should be attended to or not.

identifies which sequence a token belongs to when there is more than one sequence.

Check out the concept guide to learn more different padding and truncation arguments.

For audio tasks, you’ll need a to prepare your dataset for the model. The feature extractor is designed to extract features from raw audio data, and convert them into tensors.

Load the dataset (see the 🌎 for more details on how to load a dataset) to see how you can use a feature extractor with audio datasets:

For this tutorial, you’ll use the model. Take a look at the model card, and you’ll learn Wav2Vec2 is pretrained on 16kHz sampled speech audio. It is important your audio data’s sampling rate matches the sampling rate of the dataset used to pretrain the model. If your data’s sampling rate isn’t the same, then you need to resample your data.

Use 🌎 Datasets’ method to upsample the sampling rate to 16kHz:

Load the feature extractor with :

For computer vision tasks, you’ll need an to prepare your dataset for the model. Image preprocessing consists of several steps that convert images into the input expected by the model. These steps include but are not limited to resizing, normalizing, color channel correction, and converting images to tensors.

Load the dataset (see the 🌎 for more details on how to load a dataset) to see how you can use an image processor with computer vision datasets:

Next, take a look at the image with 🌎 Datasets feature:

Load the image processor with :

First, let’s add some image augmentation. You can use any library you prefer, but in this tutorial, we’ll use torchvision’s module. If you’re interested in using another data augmentation library, learn how in the or .

Here we use to chain together a couple of transforms - and . Note that for resizing, we can get the image size requirements from the image_processor. For some models, an exact height and width are expected, for others only the shortest_edge is defined.

The model accepts as its input. ImageProcessor can take care of normalizing the images, and generating appropriate tensors. Create a function that combines image augmentation and image preprocessing for a batch of images and generates pixel_values:

Then use 🌎 Datasets to apply the transforms on the fly:

In some cases, for instance, when fine-tuning , the model applies scale augmentation at training time. This may cause images to be different sizes in a batch. You can use DetrImageProcessor.pad() from and define a custom collate_fn to batch images together.

For tasks involving multimodal inputs, you’ll need a to prepare your dataset for the model. A processor couples together two processing objects such as as tokenizer and feature extractor.

Load the dataset (see the 🌎 for more details on how to load a dataset) to see how you can use a processor for automatic speech recognition (ASR):

Remember you should always your audio dataset’s sampling rate to match the sampling rate of the dataset used to pretrain a model!

Load a processor with :

🌍
input_ids
attention_mask
token_type_ids
Padding and truncation
feature extractor
MInDS-14
Datasets tutorial
Wav2Vec2
cast_column
AutoFeatureExtractor.from_pretrained()
image processor
food101
Datasets tutorial
Image
AutoImageProcessor.from_pretrained()
transforms
Albumentations
Kornia notebooks
Compose
RandomResizedCrop
ColorJitter
pixel_values
set_transform
DETR
DetrImageProcessor
processor
LJ Speech
Datasets tutorial
resample
AutoProcessor.from_pretrained()
Tokenizer
Feature extractor
ImageProcessor
Processor
tokenizer
AutoTokenizer.from_pretrained()