This glossary defines general machine learning and 🤗 Transformers terms to help you better understand the documentation.
The attention mask is an optional argument used when batching sequences together.
This argument indicates to the model which tokens should be attended to, and which should not.
For example, consider these two sequences:
from transformers import BertTokenizer tokenizer = BertTokenizer.from_pretrained("bert-base-cased") sequence_a = "This is a short sequence." sequence_b = "This is a rather long sequence. It is at least longer than the sequence A." encoded_sequence_a = tokenizer(sequence_a)["input_ids"] encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
The encoded versions have different lengths:
len(encoded_sequence_a), len(encoded_sequence_b) (8, 19)
Therefore, we can’t put them together in the same tensor as-is. The first sequence needs to be padded up to the length of the second one, or the second one needs to be truncated down to the length of the first one.
In the first case, the list of IDs will be extended by the padding indices. We can pass a list to the tokenizer and ask it to pad like this:
True)padded_sequences = tokenizer([sequence_a, sequence_b], padding=
We can see that 0s have been added on the right of the first sentence to make it the same length as the second one:
"input_ids"] [[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]padded_sequences[
This can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating the
position of the padded indices so that the model does not attend to them. For the BertTokenizer,
1 indicates a
value that should be attended to, while
0 indicates a padded value. This attention mask is in the dictionary returned
by the tokenizer under the key “attention_mask”:
"attention_mask"] [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 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]]padded_sequences[
The backbone is the network (embeddings and layers) that outputs the raw hidden states or features. It is usually connected to a head which accepts the features as its input to make a prediction. For example, ViTModel is a backbone without a specific head on top. Other models can also use
VitModel as a backbone such as DPT.
Color images are made up of some combination of values in three channels - red, green, and blue (RGB) - and grayscale images only have one channel. In 🤗 Transformers, the channel can be the first or last dimension of an image’s tensor: [
width] or [
causal language modeling
A pretraining task where the model reads the texts in order and has to predict the next word. It’s usually done by reading the whole sentence but using a mask inside the model to hide the future tokens at a certain timestep.
connectionist temporal classification (CTC)
An algorithm which allows a model to learn without knowing exactly how the input and output are aligned; CTC calculates the distribution of all possible outputs for a given input and chooses the most likely output from it. CTC is commonly used in speech recognition tasks because speech doesn’t always cleanly align with the transcript for a variety of reasons such as a speaker’s different speech rates.
A type of layer in a neural network where the input matrix is multiplied element-wise by a smaller matrix (kernel or filter) and the values are summed up in a new matrix. This is known as a convolutional operation which is repeated over the entire input matrix. Each operation is applied to a different segment of the input matrix. Convolutional neural networks (CNNs) are commonly used in computer vision.
decoder input IDs
This input is specific to encoder-decoder models, and contains the input IDs that will be fed to the decoder. These inputs should be used for sequence to sequence tasks, such as translation or summarization, and are usually built in a way specific to each model.
Most encoder-decoder models (BART, T5) create their
decoder_input_ids on their own from the
labels. In such models,
labels is the preferred way to handle training.
Please check each model’s docs to see how they handle these input IDs for sequence to sequence training.
Machine learning algorithms which uses neural networks with several layers.
feed forward chunking
In each residual attention block in transformers the self-attention layer is usually followed by 2 feed forward layers.
The intermediate embedding size of the feed forward layers is often bigger than the hidden size of the model (e.g., for
For an input of size
[batch_size, sequence_length], the memory required to store the intermediate feed forward
[batch_size, sequence_length, config.intermediate_size] can account for a large fraction of the memory
use. The authors of Reformer: The Efficient Transformer noticed that since the
computation is independent of the
sequence_length dimension, it is mathematically equivalent to compute the output
embeddings of both feed forward layers
[batch_size, config.hidden_size]_0, ..., [batch_size, config.hidden_size]_n
individually and concat them afterward to
[batch_size, sequence_length, config.hidden_size] with
n = sequence_length, which trades increased computation time against reduced memory use, but yields a mathematically
For models employing the function apply_chunking_to_forward(), the
chunk_size defines the number of output
embeddings that are computed in parallel and thus defines the trade-off between memory and time complexity. If
chunk_size is set to 0, no feed forward chunking is done.
The model head refers to the last layer of a neural network that accepts the raw hidden states and projects them onto a different dimension. There is a different model head for each task. For example:
- GPT2ForSequenceClassification is a sequence classification head - a linear layer - on top of the base GPT2Model.
- ViTForImageClassification is an image classification head - a linear layer on top of the final hidden state of the
CLStoken - on top of the base ViTModel.
- Wav2Vec2ForCTC ia a language modeling head with CTC on top of the base Wav2Vec2Model.
Vision-based Transformers models split an image into smaller patches which are linearly embedded, and then passed as a sequence to the model. You can find the
patch_size - or resolution - of the model in it’s configuration.
The input ids are often the only required parameters to be passed to the model as input. They are token indices, numerical representations of tokens building the sequences that will be used as input by the model.
Each tokenizer works differently but the underlying mechanism remains the same. Here’s an example using the BERT tokenizer, which is a WordPiece tokenizer:
from transformers import BertTokenizer tokenizer = BertTokenizer.from_pretrained("bert-base-cased") sequence = "A Titan RTX has 24GB of VRAM"
The tokenizer takes care of splitting the sequence into tokens available in the tokenizer vocabulary.
tokenized_sequence = tokenizer.tokenize(sequence)
The tokens are either words or subwords. Here for instance, “VRAM” wasn’t in the model vocabulary, so it’s been split in “V”, “RA” and “M”. To indicate those tokens are not separate words but parts of the same word, a double-hash prefix is added for “RA” and “M”:
print(tokenized_sequence) ['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
These tokens can then be converted into IDs which are understandable by the model. This can be done by directly feeding the sentence to the tokenizer, which leverages the Rust implementation of 🤗 Tokenizers for peak performance.
inputs = tokenizer(sequence)
The tokenizer returns a dictionary with all the arguments necessary for its corresponding model to work properly. The
token indices are under the key
"input_ids"] print(encoded_sequence) [101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]encoded_sequence = inputs[
Note that the tokenizer automatically adds “special tokens” (if the associated model relies on them) which are special IDs the model sometimes uses.
If we decode the previous sequence of ids,
decoded_sequence = tokenizer.decode(encoded_sequence)
we will see
print(decoded_sequence) [CLS] A Titan RTX has 24GB of VRAM [SEP]
because this is the way a BertModel is going to expect its inputs.
The labels are an optional argument which can be passed in order for the model to compute the loss itself. These labels should be the expected prediction of the model: it will use the standard loss in order to compute the loss between its predictions and the expected value (the label).
These labels are different according to the model head, for example:
- For sequence classification models, (BertForSequenceClassification), the model expects a tensor of dimension
(batch_size)with each value of the batch corresponding to the expected label of the entire sequence.
- For token classification models, (BertForTokenClassification), the model expects a tensor of dimension
(batch_size, seq_length)with each value corresponding to the expected label of each individual token.
- For masked language modeling, (BertForMaskedLM), the model expects a tensor of dimension
(batch_size, seq_length)with each value corresponding to the expected label of each individual token: the labels being the token ID for the masked token, and values to be ignored for the rest (usually -100).
- For sequence to sequence tasks, (BartForConditionalGeneration, MBartForConditionalGeneration), the model
expects a tensor of dimension
(batch_size, tgt_seq_length)with each value corresponding to the target sequences associated with each input sequence. During training, both BART and T5 will make the appropriate
decoder_input_idsand decoder attention masks internally. They usually do not need to be supplied. This does not apply to models leveraging the Encoder-Decoder framework.
- For image classification models, (ViTForImageClassification), the model expects a tensor of dimension
(batch_size)with each value of the batch corresponding to the expected label of each individual image.
- For semantic segmentation models, (SegformerForSemanticSegmentation), the model expects a tensor of dimension
(batch_size, height, width)with each value of the batch corresponding to the expected label of each individual pixel.
- For object detection models, (DetrForObjectDetection), the model expects a list of dictionaries with a
boxeskey where each value of the batch corresponds to the expected label and number of bounding boxes of each individual image.
- For automatic speech recognition models, (Wav2Vec2ForCTC), the model expects a tensor of dimension
(batch_size, target_length)with each value corresponding to the expected label of each individual token.
Each model’s labels may be different, so be sure to always check the documentation of each model for more information about their specific labels!
The base models (BertModel) do not accept labels, as these are the base transformer models, simply outputting features.
masked language modeling
A pretraining task where the model sees a corrupted version of the texts, usually done by masking some tokens randomly, and has to predict the original text.
A task that combines texts with another kind of inputs (for instance images).
Natural language generation
All tasks related to generating text (for instance talk with transformers, translation).
Natural language processing
A generic way to say “deal with texts”.
Natural language understanding
All tasks related to understanding what is in a text (for instance classifying the whole text, individual words).
A tensor of the numerical representations of an image that is passed to a model. The pixel values have a shape of [
width], and are generated from an image processor.
An operation that reduces a matrix into a smaller matrix, either by taking the maximum or average of the pooled dimension(s). Pooling layers are commonly found between convolutional layers to downsample the feature representation.
Contrary to RNNs that have the position of each token embedded within them, transformers are unaware of the position of
each token. Therefore, the position IDs (
position_ids) are used by the model to identify each token’s position in the
list of tokens.
They are an optional parameter. If no
position_ids are passed to the model, the IDs are automatically created as
absolute positional embeddings.
Absolute positional embeddings are selected in the range
[0, config.max_position_embeddings - 1]. Some models use
other types of positional embeddings, such as sinusoidal position embeddings or relative position embeddings.
A model that has been pretrained on some data (for instance all of Wikipedia). Pretraining methods involve a self-supervised objective, which can be reading the text and trying to predict the next word (see causal language modeling) or masking some words and trying to predict them (see masked language modeling).
Speech and vision models have their own pretraining objectives. For example, Wav2Vec2 is a speech model pretrained on a contrastive task which requires the model to identify the “true” speech representation from a set of “false” speech representations. On the other hand, BEiT is a vision model pretrained on a masked image modeling task which masks some of the image patches and requires the model to predict the masked patches (similar to the masked language modeling objective).
recurrent neural network
A type of model that uses a loop over a layer to process texts.
A measurement in hertz of the number of samples (the audio signal) taken per second. The sampling rate is a result of discretizing a continuous signal such as speech.
Each element of the input finds out which other elements of the input they should attend to.
Models that generate a new sequence from an input, like translation models, or summarization models (such as Bart or T5).
In convolution or pooling, the stride refers to the distance the kernel is moved over a matrix. A stride of 1 means the kernel is moved one pixel over at a time, and a stride of 2 means the kernel is moved two pixels over at a time.
A part of a sentence, usually a word, but can also be a subword (non-common words are often split in subwords) or a punctuation symbol.
token Type IDs
Some models’ purpose is to do classification on pairs of sentences or question answering.
These require two different sequences to be joined in a single “input_ids” entry, which usually is performed with the
help of special tokens, such as the classifier (
[CLS]) and separator (
[SEP]) tokens. For example, the BERT model
builds its two sequence input as such:
# [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
We can use our tokenizer to automatically generate such a sentence by passing the two sequences to
tokenizer as two
arguments (and not a list, like before) like this:
from transformers import BertTokenizer tokenizer = BertTokenizer.from_pretrained("bert-base-cased") sequence_a = "HuggingFace is based in NYC" sequence_b = "Where is HuggingFace based?" encoded_dict = tokenizer(sequence_a, sequence_b) decoded = tokenizer.decode(encoded_dict["input_ids"])
which will return:
print(decoded) [CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]
This is enough for some models to understand where one sequence ends and where another begins. However, other models, such as BERT, also deploy token type IDs (also called segment IDs). They are represented as a binary mask identifying the two types of sequence in the model.
The tokenizer returns this mask as the “token_type_ids” entry:
"token_type_ids"] [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]encoded_dict[
The first sequence, the “context” used for the question, has all its tokens represented by a
0, whereas the second
sequence, corresponding to the “question”, has all its tokens represented by a
Some models, like XLNetModel use an additional token represented by a
Self-attention based deep learning model architecture.