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 Tokenizer 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 Feature extractor to extract sequential features from audio waveforms and convert them into tensors.
• Image inputs use a ImageProcessor to convert images into tensors.
• Multimodal inputs, use a Processor to combine a tokenizer and a feature extractor or image processor.

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

pip install datasets

Natural Language Processing

The main tool for preprocessing textual data is a tokenizer. 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.

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

>>> from transformers import AutoTokenizer

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

Then pass your text to the tokenizer:

>>> 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:

• input_ids are the indices corresponding to each token in the sentence.
• attention_mask indicates whether a token should be attended to or not.
• token_type_ids identifies which sequence a token belongs to when there is more than one sequence.

Return your input by decoding the input_ids:

>>> 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:

>>> 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:

>>> 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:

>>> 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:

>>> 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]])}

>>> 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

For audio tasks, you’ll need a feature extractor 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 MInDS-14 dataset (see the 🤗 Datasets tutorial for more details on how to load a dataset) to see how you can use a feature extractor with audio datasets:

>>> 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:

>>> dataset[0]["audio"]
{'array': array([ 0.        ,  0.00024414, -0.00024414, ..., -0.00024414,
         0.        ,  0.        ], dtype=float32),
 'path': '/root/.cache/huggingface/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.

For this tutorial, you’ll use the Wav2Vec2 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.

1. Use 🤗 Datasets’ cast_column method to upsample the sampling rate to 16kHz:

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

2. Call the audio column again to resample the audio file:

>>> 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/huggingface/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.

Load the feature extractor with AutoFeatureExtractor.from_pretrained():

>>> 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.

>>> 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:

>>> 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:

>>> 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 first few examples in the dataset:

>>> 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!

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

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

Computer vision

For computer vision tasks, you’ll need an image processor 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 food101 dataset (see the 🤗 Datasets tutorial for more details on how to load a dataset) to see how you can use an image processor with computer vision datasets:

>>> from datasets import load_dataset

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

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

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

Load the image processor with AutoImageProcessor.from_pretrained():

>>> from transformers import AutoImageProcessor

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

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

1. Here we use Compose to chain together a couple of transforms - RandomResizedCrop and ColorJitter. 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.

>>> 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)])

2. The model accepts pixel_values 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:

>>> 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

3. Then use 🤗 Datasetsset_transform to apply the transforms on the fly:

>>> dataset.set_transform(transforms)

4. 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!

>>> 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.

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

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

Pad

In some cases, for instance, when fine-tuning DETR, 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 DetrImageProcessor and define a custom collate_fn to batch images together.

>>> 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

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

Load the LJ Speech dataset (see the 🤗 Datasets tutorial for more details on how to load a dataset) to see how you can use a processor for automatic speech recognition (ASR):

>>> 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:

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

Now take a look at the audio and text columns:

>>> 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/huggingface/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'

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

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

Load a processor with AutoProcessor.from_pretrained():

>>> 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:

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

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

...     return example

2. Apply the prepare_dataset function to a sample:

>>> 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!