preprocessing.mdx

<!--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.
-->

# Preprocess

[[open-in-colab]]

Before you can use your data in a model, the data needs to be processed into an acceptable format for the model. A model does not understand raw text, images or audio. These inputs need to be converted into numbers and assembled into tensors. In this tutorial, you will:

* Preprocess textual data with a tokenizer.
* Preprocess image or audio data with a feature extractor.
* Preprocess data for a multimodal task with a processor.

## NLP

<Youtube id="Yffk5aydLzg"/>

The main tool for processing textual data is a [tokenizer](main_classes/tokenizer). A tokenizer starts by splitting text into *tokens* according to a set of rules. The tokens are converted into numbers, which are used to build tensors as input to a model. Any additional inputs required by a model are also added by the tokenizer.

<Tip>

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 referrred to as the *vocab*) during pretraining.

</Tip>

Get started quickly by loading a pretrained tokenizer with the [`AutoTokenizer`] class. This downloads the *vocab* used when a model is pretrained.

### Tokenize

Load a pretrained tokenizer with [`AutoTokenizer.from_pretrained`]:

```py
>>> from transformers import AutoTokenizer

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

Then pass your sentence to the tokenizer:

```py
>>> 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 itmes:

* [input_ids](glossary#input-ids) are the indices corresponding to each token in the sentence.
* [attention_mask](glossary#attention-mask) indicates whether a token should be attended to or not.
* [token_type_ids](glossary#token-type-ids) identifies which sequence a token belongs to when there is more than one sequence.

You can decode the `input_ids` to return the original input:

```py
>>> 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 will automatically add them for you.

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

```py
>>> 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

This brings us to an important topic. When you process a batch of sentences, they aren't always the same length. This is a problem because tensors, the input to the model, need to have a uniform shape. Padding is a strategy for ensuring tensors are rectangular by adding a special *padding token* to sentences with fewer tokens.

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

```py
>>> 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]]}
```

Notice the tokenizer padded the first and third sentences with a `0` 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 will 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:

```py
>>> 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 are fed to the model.

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

<frameworkcontent>
<pt>

```py
>>> batch_sentences = [
...     "But what about second breakfast?",
...     "Don't think he knows about second breakfast, Pip.",
...     "What about elevensies?",
... ]
>>> encoded_input = tokenizer(batch, padding=True, truncation=True, return_tensors="pt")
>>> print(encoded_input)
{'input_ids': tensor([[  101,   153,  7719, 21490,  1122,  1114,  9582,  1623,   102],
                      [  101,  5226,  1122,  9649,  1199,  2610,  1236,   102,     0]]), 
 'token_type_ids': tensor([[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, 1],
                           [1, 1, 1, 1, 1, 1, 1, 1, 0]])}
```
</pt>
<tf>
```py
>>> 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,   153,  7719, 21490,  1122,  1114,  9582,  1623,   102],
       [  101,  5226,  1122,  9649,  1199,  2610,  1236,   102,     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]], dtype=int32)>, 
 'attention_mask': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
array([[1, 1, 1, 1, 1, 1, 1, 1, 1],
       [1, 1, 1, 1, 1, 1, 1, 1, 0]], dtype=int32)>}
```
</tf>
</frameworkcontent>

## Audio

Audio inputs are preprocessed differently than textual inputs, but the end goal remains the same: create numerical sequences the model can understand. A [feature extractor](main_classes/feature_extractor) is designed for the express purpose of extracting features from raw image or audio data and converting them into tensors. Before you begin, install 🤗 Datasets to load an audio dataset to experiment with:

```bash
pip install datasets
```

Load the keyword spotting task from the [SUPERB](https://huggingface.co/datasets/superb) benchmark (see the 🤗 [Datasets tutorial](https://huggingface.co/docs/datasets/load_hub.html) for more details on how to load a dataset):

```py
>>> from datasets import load_dataset, Audio

>>> dataset = load_dataset("superb", "ks")
```

Access the first element of the `audio` column to take a look at the input. Calling the `audio` column will automatically load and resample the audio file:

```py
>>> dataset["train"][0]["audio"]
{'array': array([ 0.        ,  0.        ,  0.        , ..., -0.00592041,
        -0.00405884, -0.00253296], dtype=float32),
 'path': '/root/.cache/huggingface/datasets/downloads/extracted/05734a36d88019a09725c20cc024e1c4e7982e37d7d55c0c1ca1742ea1cdd47f/_background_noise_/doing_the_dishes.wav',
 'sampling_rate': 16000}
```

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.

### Resample

For this tutorial, you will use the [Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base) model. As you can see from the model card, the Wav2Vec2 model 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 audio data. 

For example, load the [LJ Speech](https://huggingface.co/datasets/lj_speech) dataset which has a sampling rate of 22050kHz. In order to use the Wav2Vec2 model with this dataset, downsample the sampling rate to 16kHz:

```py
>>> lj_speech = load_dataset("lj_speech", split="train")
>>> 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}
```

1. Use 🤗 Datasets' [`cast_column`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.cast_column) method to downsample the sampling rate to 16kHz:

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

2. Load the audio file:

```py
>>> lj_speech[0]["audio"]
{'array': array([-0.00064146, -0.00074657, -0.00068768, ...,  0.00068341,
         0.00014045,  0.        ], dtype=float32),
 'path': '/root/.cache/huggingface/datasets/downloads/extracted/917ece08c95cf0c4115e45294e3cd0dee724a1165b7fc11798369308a465bd26/LJSpeech-1.1/wavs/LJ001-0001.wav',
 'sampling_rate': 16000}
```

As you can see, the `sampling_rate` was downsampled to 16kHz. Now that you know how resampling works, let's return to our previous example with the SUPERB dataset!

### Feature extractor

The next step is to 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, and the audio feature extractor will add a `0` - interpreted as silence - to `array`.

Load the feature extractor with [`AutoFeatureExtractor.from_pretrained`]:

```py
>>> 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.

```py
>>> audio_input = [dataset["train"][0]["audio"]["array"]]
>>> feature_extractor(audio_input, sampling_rate=16000)
{'input_values': [array([ 0.00045439,  0.00045439,  0.00045439, ..., -0.1578519 , -0.10807519, -0.06727459], dtype=float32)]}
```

### Pad and truncate

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:

```py
>>> dataset["train"][0]["audio"]["array"].shape
(1522930,)

>>> dataset["train"][1]["audio"]["array"].shape
(988891,)
```

As you can see, the first sample has a longer sequence than the second sample. Let's create a function that will preprocess the dataset. Specify a maximum sample length, and the feature extractor will either pad or truncate the sequences to match it:

```py
>>> 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=1000000,
...         truncation=True,
...     )
...     return inputs
```

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

```py
>>> processed_dataset = preprocess_function(dataset["train"][:5])
```

Now take another look at the processed sample lengths:

```py
>>> processed_dataset["input_values"][0].shape
(1000000,)

>>> processed_dataset["input_values"][1].shape
(1000000,)
```

The lengths of the first two samples now match the maximum length you specified.

## Vision

A feature extractor is also used to process images for vision tasks. Once again, the goal is to convert the raw image into a batch of tensors as input.

Let's load the [food101](https://huggingface.co/datasets/food101) dataset for this tutorial. Use 🤗 Datasets `split` parameter to only load a small sample from the training split since the dataset is quite large:

```py
>>> from datasets import load_dataset

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

Next, take a look at the image with 🤗 Datasets [`Image`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=image#datasets.Image) feature:

```py
>>> dataset[0]["image"]
```

![vision-preprocess-tutorial.png](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/vision-preprocess-tutorial.png)

### Feature extractor

Load the feature extractor with [`AutoFeatureExtractor.from_pretrained`]:

```py
>>> from transformers import AutoFeatureExtractor

>>> feature_extractor = AutoFeatureExtractor.from_pretrained("google/vit-base-patch16-224")
```

### Data augmentation

For vision tasks, it is common to add some type of data augmentation to the images as a part of preprocessing. You can add augmentations with any library you'd like, but in this tutorial, you will use torchvision's [`transforms`](https://pytorch.org/vision/stable/transforms.html) module.

1. Normalize the image and use [`Compose`](https://pytorch.org/vision/master/generated/torchvision.transforms.Compose.html) to chain some transforms - [`RandomResizedCrop`](https://pytorch.org/vision/main/generated/torchvision.transforms.RandomResizedCrop.html) and [`ColorJitter`](https://pytorch.org/vision/main/generated/torchvision.transforms.ColorJitter.html) - together:

```py
>>> from torchvision.transforms import Compose, Normalize, RandomResizedCrop, ColorJitter, ToTensor

>>> normalize = Normalize(mean=feature_extractor.image_mean, std=feature_extractor.image_std)
>>> _transforms = Compose(
...     [RandomResizedCrop(feature_extractor.size), ColorJitter(brightness=0.5, hue=0.5), ToTensor(), normalize]
... )
```

2. The model accepts [`pixel_values`](model_doc/visionencoderdecoder#transformers.VisionEncoderDecoderModel.forward.pixel_values) as it's input. This value is generated by the feature extractor. Create a function that generates `pixel_values` from the transforms:

```py
>>> def transforms(examples):
...     examples["pixel_values"] = [_transforms(image.convert("RGB")) for image in examples["image"]]
...     return examples
```

3. Then use 🤗 Datasets [`set_transform`](https://huggingface.co/docs/datasets/process.html#format-transform) to apply the transforms on-the-fly:

```py
>>> dataset.set_transform(transforms)
```

4. Now when you access the image, you will notice the feature extractor has added the model input `pixel_values`:

```py
>>> dataset[0]["image"]
{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=384x512 at 0x7F1A7B0630D0>,
 'label': 6,
 'pixel_values': tensor([[[ 0.0353,  0.0745,  0.1216,  ..., -0.9922, -0.9922, -0.9922],
          [-0.0196,  0.0667,  0.1294,  ..., -0.9765, -0.9843, -0.9922],
          [ 0.0196,  0.0824,  0.1137,  ..., -0.9765, -0.9686, -0.8667],
          ...,
          [ 0.0275,  0.0745,  0.0510,  ..., -0.1137, -0.1216, -0.0824],
          [ 0.0667,  0.0824,  0.0667,  ..., -0.0588, -0.0745, -0.0980],
          [ 0.0353,  0.0353,  0.0431,  ..., -0.0039, -0.0039, -0.0588]],
 
         [[ 0.2078,  0.2471,  0.2863,  ..., -0.9451, -0.9373, -0.9451],
          [ 0.1608,  0.2471,  0.3098,  ..., -0.9373, -0.9451, -0.9373],
          [ 0.2078,  0.2706,  0.3020,  ..., -0.9608, -0.9373, -0.8275],
          ...,
          [-0.0353,  0.0118, -0.0039,  ..., -0.2392, -0.2471, -0.2078],
          [ 0.0196,  0.0353,  0.0196,  ..., -0.1843, -0.2000, -0.2235],
          [-0.0118, -0.0039, -0.0039,  ..., -0.0980, -0.0980, -0.1529]],
 
         [[ 0.3961,  0.4431,  0.4980,  ..., -0.9216, -0.9137, -0.9216],
          [ 0.3569,  0.4510,  0.5216,  ..., -0.9059, -0.9137, -0.9137],
          [ 0.4118,  0.4745,  0.5216,  ..., -0.9137, -0.8902, -0.7804],
          ...,
          [-0.2314, -0.1922, -0.2078,  ..., -0.4196, -0.4275, -0.3882],
          [-0.1843, -0.1686, -0.2000,  ..., -0.3647, -0.3804, -0.4039],
          [-0.1922, -0.1922, -0.1922,  ..., -0.2941, -0.2863, -0.3412]]])}
```

Here is what the image looks like after you preprocess it. Just as you'd expect from the applied transforms, the image has been randomly cropped and it's color properties are different.

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

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

![preprocessed_image](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/preprocessed_image.png)

## Multimodal

For multimodal tasks. you will use a combination of everything you've learned so far and apply your skills to a automatic speech recognition (ASR) task. This means you will need a:

* Feature extractor to preprocess the audio data.
* Tokenizer to process the text.

Let's return to the [LJ Speech](https://huggingface.co/datasets/lj_speech) dataset:

```py
>>> from datasets import load_dataset

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

Since you are mainly interested in the `audio` and `text` column, remove the other columns:

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

Now take a look at the `audio` and `text` columns:

```py
>>> 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 from the earlier section on processing audio data, you should always [resample](preprocessing#audio) your audio data's sampling rate to match the sampling rate of the dataset used to pretrain a model:

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

### Processor

A processor combines a feature extractor and tokenizer. Load a processor with [`AutoProcessor.from_pretrained]:

```py
>>> from transformers import AutoProcessor

>>> processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base-960h")
```

1. Create a function to process the audio data to `input_values`, and tokenizes the text to `labels`. These are your inputs to the model:

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

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

...     with processor.as_target_processor():
...         example["labels"] = processor(example["text"]).input_ids
...     return example
```

2. Apply the `prepare_dataset` function to a sample:

```py
>>> prepare_dataset(lj_speech[0])
```

Notice the processor has added `input_values` and `labels`. The sampling rate has also been correctly downsampled to 16kHz.

Awesome, you should now be able to preprocess data for any modality and even combine different modalities! In the next tutorial, learn how to fine-tune a model on your newly preprocessed data.