Skip to content

GitLab

Unverified Commit aa39967b authored by Steven Liu's avatar Steven Liu Committed by GitHub
Browse files

reorganize glossary (#20010)

parent 305e8718
Hide whitespace changes
Inline Side-by-side
Showing with 198 additions and 146 deletions
docs/source/en/glossary.mdx
View file @ aa39967b
......@@ -12,39 +12,122 @@ specific language governing permissions and limitations under the License.
# Glossary
## General terms
- autoencoding models: see MLM
- autoregressive models: see CLM
- CLM: 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.
- deep learning: machine learning algorithms which uses neural networks with several layers.
- MLM: 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.
- multimodal: a task that combines texts with another kind of inputs (for instance images).
- NLG: natural language generation, all tasks related to generating text (for instance talk with transformers,
translation).
- NLP: natural language processing, a generic way to say "deal with texts".
- NLU: natural language understanding, all tasks related to understanding what is in a text (for instance classifying
the whole text, individual words).
- pretrained model: 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 CLM) or
masking some words and trying to predict them (see MLM).
- RNN: recurrent neural network, a type of model that uses a loop over a layer to process texts.
- self-attention: each element of the input finds out which other elements of the input they should attend to.
- seq2seq or sequence-to-sequence: models that generate a new sequence from an input, like translation models, or
summarization models (such as [Bart](model_doc/bart) or [T5](model_doc/t5)).
- token: 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.
- transformer: self-attention based deep learning model architecture.
## Model inputs
Every model is different yet bears similarities with the others. Therefore most models use the same inputs, which are
detailed here alongside usage examples.
This glossary defines general machine learning and 🤗 Transformers terms to help you better understand the
documentation.
## A
### Attention mask
The attention mask is an optional argument used when batching sequences together.
<Youtube id="M6adb1j2jPI"/>
This argument indicates to the model which tokens should be attended to, and which should not.
For example, consider these two sequences:
```python
>>> 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:
```python
>>> 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:
```python
>>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
```
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:
```python
>>> padded_sequences["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]]
```
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":
```python
>>> padded_sequences["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]]
```
### autoencoding models
see [MLM](#mlm)
### autoregressive models
see [CLM](#clm)
## C
### CLM
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.
## D
### 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,
passing the `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.
### deep learning
Machine learning algorithms which uses neural networks with several layers.
## F
### 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
`bert-base-uncased`).
For an input of size `[batch_size, sequence_length]`, the memory required to store the intermediate feed forward
embeddings `[batch_size, sequence_length, config.intermediate_size]` can account for a large fraction of the memory
use. The authors of [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) 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
**equivalent** result.
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.
## I
### Input IDs
......@@ -80,7 +163,8 @@ is added for "RA" and "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](https://github.com/huggingface/tokenizers) for peak performance.
the sentence to the tokenizer, which leverages the Rust implementation of [🤗
Tokenizers](https://github.com/huggingface/tokenizers) for peak performance.
```python
>>> inputs = tokenizer(sequence)
......@@ -113,65 +197,103 @@ we will see
because this is the way a [`BertModel`] is going to expect its inputs.
## L
### Labels
### Attention mask
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).
The attention mask is an optional argument used when batching sequences together.
These labels are different according to the model head, for example:
<Youtube id="M6adb1j2jPI"/>
- For sequence classification models (e.g., [`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 (e.g., [`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 (e.g., [`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,(e.g., [`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_ids* and decoder attention masks internally. They usually do not need to be supplied. This does not
apply to models leveraging the Encoder-Decoder framework. See the documentation of each model for more information on
each specific model's labels.
This argument indicates to the model which tokens should be attended to, and which should not.
The base models (e.g., [`BertModel`]) do not accept labels, as these are the base transformer models, simply outputting
features.
For example, consider these two sequences:
## M
```python
>>> from transformers import BertTokenizer
### MLM
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
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.
>>> sequence_a = "This is a short sequence."
>>> sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
### multimodal
>>> encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
>>> encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
```
A task that combines texts with another kind of inputs (for instance images).
The encoded versions have different lengths:
## N
```python
>>> len(encoded_sequence_a), len(encoded_sequence_b)
(8, 19)
```
### NLG
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.
Natural language generation, all tasks related to generating text (for instance talk with transformers, translation).
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:
### NLP
```python
>>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
```
Natural language processing, a generic way to say "deal with texts".
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:
### NLU
```python
>>> padded_sequences["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]]
```
Natural language understanding, all tasks related to understanding what is in a text (for instance classifying the
whole text, individual words).
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":
## P
```python
>>> padded_sequences["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]]
```
### Position IDs
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.
### pretrained model
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 CLM) or masking some
words and trying to predict them (see MLM).
## R
### RNN
Recurrent neural network, a type of model that uses a loop over a layer to process texts.
## S
### self-attention
Each element of the input finds out which other elements of the input they should attend to.
### seq2seq or sequence-to-sequence
Models that generate a new sequence from an input, like translation models, or summarization models (such as
[Bart](model_doc/bart) or [T5](model_doc/t5)).
## T
### token
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
......@@ -180,8 +302,8 @@ Some models' purpose is to do classification on pairs of sentences or question a
<Youtube id="0u3ioSwev3s"/>
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:
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:
```python
>>> # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
......@@ -219,81 +341,11 @@ The tokenizer returns this mask as the "token_type_ids" entry:
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
```
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 `1`.
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 `1`.
Some models, like [`XLNetModel`] use an additional token represented by a `2`.
### transformer
### Position IDs
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.
### Labels
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 (e.g., [`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 (e.g., [`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 (e.g., [`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,(e.g., [`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_ids* and decoder attention masks internally.
They usually do not need to be supplied. This does not apply to models leveraging the Encoder-Decoder framework. See
the documentation of each model for more information on each specific model's labels.
The base models (e.g., [`BertModel`]) do not accept labels, as these are the base transformer
models, simply outputting features.
### 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, passing the `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.
### 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
`bert-base-uncased`).
For an input of size `[batch_size, sequence_length]`, the memory required to store the intermediate feed forward
embeddings `[batch_size, sequence_length, config.intermediate_size]` can account for a large fraction of the memory
use. The authors of [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) 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
**equivalent** result.
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.
Self-attention based deep learning model architecture.
\ No newline at end of file
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment