6.2 KiB
CANINE
Overview
The CANINE model was proposed in CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting. It's among the first papers that trains a Transformer without using an explicit tokenization step (such as Byte Pair Encoding (BPE), WordPiece or SentencePiece). Instead, the model is trained directly at a Unicode character-level. Training at a character-level inevitably comes with a longer sequence length, which CANINE solves with an efficient downsampling strategy, before applying a deep Transformer encoder.
The abstract from the paper is the following:
Pipelined NLP systems have largely been superseded by end-to-end neural modeling, yet nearly all commonly-used models still require an explicit tokenization step. While recent tokenization approaches based on data-derived subword lexicons are less brittle than manually engineered tokenizers, these techniques are not equally suited to all languages, and the use of any fixed vocabulary may limit a model's ability to adapt. In this paper, we present CANINE, a neural encoder that operates directly on character sequences, without explicit tokenization or vocabulary, and a pre-training strategy that operates either directly on characters or optionally uses subwords as a soft inductive bias. To use its finer-grained input effectively and efficiently, CANINE combines downsampling, which reduces the input sequence length, with a deep transformer stack, which encodes context. CANINE outperforms a comparable mBERT model by 2.8 F1 on TyDi QA, a challenging multilingual benchmark, despite having 28% fewer model parameters.
This model was contributed by nielsr. The original code can be found here.
Usage tips
- CANINE uses no less than 3 Transformer encoders internally: 2 "shallow" encoders (which only consist of a single layer) and 1 "deep" encoder (which is a regular BERT encoder). First, a "shallow" encoder is used to contextualize the character embeddings, using local attention. Next, after downsampling, a "deep" encoder is applied. Finally, after upsampling, a "shallow" encoder is used to create the final character embeddings. Details regarding up- and downsampling can be found in the paper.
- CANINE uses a max sequence length of 2048 characters by default. One can use [
CanineTokenizer] to prepare text for the model. - Classification can be done by placing a linear layer on top of the final hidden state of the special [CLS] token (which has a predefined Unicode code point). For token classification tasks however, the downsampled sequence of tokens needs to be upsampled again to match the length of the original character sequence (which is 2048). The details for this can be found in the paper.
Model checkpoints:
- google/canine-c: Pre-trained with autoregressive character loss, 12-layer, 768-hidden, 12-heads, 121M parameters (size ~500 MB).
- google/canine-s: Pre-trained with subword loss, 12-layer, 768-hidden, 12-heads, 121M parameters (size ~500 MB).
Usage example
CANINE works on raw characters, so it can be used without a tokenizer:
>>> from transformers import CanineModel
>>> import torch
>>> model = CanineModel.from_pretrained("google/canine-c") # model pre-trained with autoregressive character loss
>>> text = "hello world"
>>> # use Python's built-in ord() function to turn each character into its unicode code point id
>>> input_ids = torch.tensor([[ord(char) for char in text]])
>>> outputs = model(input_ids) # forward pass
>>> pooled_output = outputs.pooler_output
>>> sequence_output = outputs.last_hidden_state
For batched inference and training, it is however recommended to make use of the tokenizer (to pad/truncate all sequences to the same length):
>>> from transformers import CanineTokenizer, CanineModel
>>> model = CanineModel.from_pretrained("google/canine-c")
>>> tokenizer = CanineTokenizer.from_pretrained("google/canine-c")
>>> inputs = ["Life is like a box of chocolates.", "You never know what you gonna get."]
>>> encoding = tokenizer(inputs, padding="longest", truncation=True, return_tensors="pt")
>>> outputs = model(**encoding) # forward pass
>>> pooled_output = outputs.pooler_output
>>> sequence_output = outputs.last_hidden_state
Resources
- Text classification task guide
- Token classification task guide
- Question answering task guide
- Multiple choice task guide
CanineConfig
autodoc CanineConfig
CanineTokenizer
autodoc CanineTokenizer - build_inputs_with_special_tokens - get_special_tokens_mask - create_token_type_ids_from_sequences
CANINE specific outputs
autodoc models.canine.modeling_canine.CanineModelOutputWithPooling
CanineModel
autodoc CanineModel - forward
CanineForSequenceClassification
autodoc CanineForSequenceClassification - forward
CanineForMultipleChoice
autodoc CanineForMultipleChoice - forward
CanineForTokenClassification
autodoc CanineForTokenClassification - forward
CanineForQuestionAnswering
autodoc CanineForQuestionAnswering - forward