T5 | Transformers

T5

Overview

The T5 model was presented in Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.

The abstract from the paper is the following:

Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pretraining objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.

Tips:

T5 is an encoder-decoder model pre-trained on a multi-task mixture of unsupervised and supervised tasks and for which each task is converted into a text-to-text format. T5 works well on a variety of tasks out-of-the-box by prepending a different prefix to the input corresponding to each task, e.g., for translation: translate English to German: …, for summarization: summarize: ….
The pretraining includes both supervised and self-supervised training. Supervised training is conducted on downstream tasks provided by the GLUE and SuperGLUE benchmarks (converting them into text-to-text tasks as explained above).
Self-supervised training uses corrupted tokens, by randomly removing 15% of the tokens and replacing them with individual sentinel tokens (if several consecutive tokens are marked for removal, the whole group is replaced with a single sentinel token). The input of the encoder is the corrupted sentence, the input of the decoder is the original sentence and the target is then the dropped out tokens delimited by their sentinel tokens.
T5 uses relative scalar embeddings. Encoder input padding can be done on the left and on the right.
See the training, inference and scripts sections below for all details regarding usage.

T5 comes in different sizes:

Based on the original T5 model, Google has released some follow-up works:

T5v1.1: T5v1.1 is an improved version of T5 with some architectural tweaks, and is pre-trained on C4 only without mixing in the supervised tasks. Refer to the documentation of T5v1.1 which can be found here.
mT5: mT5 is a multilingual T5 model. It is pre-trained on the mC4 corpus, which includes 101 languages. Refer to the documentation of mT5 which can be found here.
byT5: byT5 is a T5 model pre-trained on byte sequences rather than SentencePiece subword token sequences. Refer to the documentation of byT5 which can be found here.
UL2: UL2 is a T5 like model pretrained on various denoising objectives
Flan-T5: Flan is a pretraining methods that is based on prompting. The Flan-T5 are T5 models trained on the Flan collection of datasets which include: taskmaster2, djaym7/wiki_dialog, deepmind/code_contests, lambada, gsm8k, aqua_rat, esnli, quasc and qed.
FLan-UL2 : the UL2 model finetuned using the “Flan” prompt tuning and dataset collection.
UMT5: UmT5 is a multilingual T5 model trained on an improved and refreshed mC4 multilingual corpus, 29 trillion characters across 107 language, using a new sampling method, UniMax. Refer to the documentation of mT5 which can be found here.

All checkpoints can be found on the hub.

This model was contributed by thomwolf. The original code can be found here.

Training

T5 is an encoder-decoder model and converts all NLP problems into a text-to-text format. It is trained using teacher forcing. This means that for training, we always need an input sequence and a corresponding target sequence. The input sequence is fed to the model using input_ids. The target sequence is shifted to the right, i.e., prepended by a start-sequence token and fed to the decoder using the decoder_input_ids. In teacher-forcing style, the target sequence is then appended by the EOS token and corresponds to the labels. The PAD token is hereby used as the start-sequence token. T5 can be trained / fine-tuned both in a supervised and unsupervised fashion.

One can use T5ForConditionalGeneration (or the Tensorflow/Flax variant), which includes the language modeling head on top of the decoder.

Unsupervised denoising training

In this setup, spans of the input sequence are masked by so-called sentinel tokens (a.k.a unique mask tokens) and the output sequence is formed as a concatenation of the same sentinel tokens and the real masked tokens. Each sentinel token represents a unique mask token for this sentence and should start with <extra_id_0>, <extra_id_1>, … up to <extra_id_99>. As a default, 100 sentinel tokens are available in T5Tokenizer.

For instance, the sentence “The cute dog walks in the park” with the masks put on “cute dog” and “the” should be processed as follows:

Copied

>>> from transformers import T5Tokenizer, T5ForConditionalGeneration

>>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
>>> model = T5ForConditionalGeneration.from_pretrained("t5-small")

>>> input_ids = tokenizer("The <extra_id_0> walks in <extra_id_1> park", return_tensors="pt").input_ids
>>> labels = tokenizer("<extra_id_0> cute dog <extra_id_1> the <extra_id_2>", return_tensors="pt").input_ids

>>> # the forward function automatically creates the correct decoder_input_ids
>>> loss = model(input_ids=input_ids, labels=labels).loss
>>> loss.item()
3.7837

If you’re interested in pre-training T5 on a new corpus, check out the run_t5_mlm_flax.py script in the Examples directory.

Supervised training

In this setup, the input sequence and output sequence are a standard sequence-to-sequence input-output mapping. Suppose that we want to fine-tune the model for translation for example, and we have a training example: the input sequence “The house is wonderful.” and output sequence “Das Haus ist wunderbar.”, then they should be prepared for the model as follows:

Copied

>>> from transformers import T5Tokenizer, T5ForConditionalGeneration

>>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
>>> model = T5ForConditionalGeneration.from_pretrained("t5-small")

>>> input_ids = tokenizer("translate English to German: The house is wonderful.", return_tensors="pt").input_ids
>>> labels = tokenizer("Das Haus ist wunderbar.", return_tensors="pt").input_ids

>>> # the forward function automatically creates the correct decoder_input_ids
>>> loss = model(input_ids=input_ids, labels=labels).loss
>>> loss.item()
0.2542

As you can see, only 2 inputs are required for the model in order to compute a loss: input_ids (which are the input_ids of the encoded input sequence) and labels (which are the input_ids of the encoded target sequence). The model will automatically create the decoder_input_ids based on the labels, by shifting them one position to the right and prepending the config.decoder_start_token_id, which for T5 is equal to 0 (i.e. the id of the pad token). Also note the task prefix: we prepend the input sequence with ‘translate English to German: ’ before encoding it. This will help in improving the performance, as this task prefix was used during T5’s pre-training.

However, the example above only shows a single training example. In practice, one trains deep learning models in batches. This entails that we must pad/truncate examples to the same length. For encoder-decoder models, one typically defines a max_source_length and max_target_length, which determine the maximum length of the input and output sequences respectively (otherwise they are truncated). These should be carefully set depending on the task.

In addition, we must make sure that padding token id’s of the labels are not taken into account by the loss function. In PyTorch and Tensorflow, this can be done by replacing them with -100, which is the ignore_index of the CrossEntropyLoss. In Flax, one can use the decoder_attention_mask to ignore padded tokens from the loss (see the Flax summarization script for details). We also pass attention_mask as additional input to the model, which makes sure that padding tokens of the inputs are ignored. The code example below illustrates all of this.

Copied

>>> from transformers import T5Tokenizer, T5ForConditionalGeneration
>>> import torch

>>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
>>> model = T5ForConditionalGeneration.from_pretrained("t5-small")

>>> # the following 2 hyperparameters are task-specific
>>> max_source_length = 512
>>> max_target_length = 128

>>> # Suppose we have the following 2 training examples:
>>> input_sequence_1 = "Welcome to NYC"
>>> output_sequence_1 = "Bienvenue à NYC"

>>> input_sequence_2 = "BOINCAI is a company"
>>> output_sequence_2 = "BOINCAI est une entreprise"

>>> # encode the inputs
>>> task_prefix = "translate English to French: "
>>> input_sequences = [input_sequence_1, input_sequence_2]

>>> encoding = tokenizer(
...     [task_prefix + sequence for sequence in input_sequences],
...     padding="longest",
...     max_length=max_source_length,
...     truncation=True,
...     return_tensors="pt",
... )

>>> input_ids, attention_mask = encoding.input_ids, encoding.attention_mask

>>> # encode the targets
>>> target_encoding = tokenizer(
...     [output_sequence_1, output_sequence_2],
...     padding="longest",
...     max_length=max_target_length,
...     truncation=True,
...     return_tensors="pt",
... )
>>> labels = target_encoding.input_ids

>>> # replace padding token id's of the labels by -100 so it's ignored by the loss
>>> labels[labels == tokenizer.pad_token_id] = -100

>>> # forward pass
>>> loss = model(input_ids=input_ids, attention_mask=attention_mask, labels=labels).loss
>>> loss.item()
0.188

Additional training tips:

T5 models need a slightly higher learning rate than the default one set in the Trainer when using the AdamW optimizer. Typically, 1e-4 and 3e-4 work well for most problems (classification, summarization, translation, question answering, question generation). Note that T5 was pre-trained using the AdaFactor optimizer.

According to this forum post, task prefixes matter when (1) doing multi-task training (2) your task is similar or related to one of the supervised tasks used in T5’s pre-training mixture (see Appendix D of the paper for the task prefixes used).

If training on TPU, it is recommended to pad all examples of the dataset to the same length or make use of pad_to_multiple_of to have a small number of predefined bucket sizes to fit all examples in. Dynamically padding batches to the longest example is not recommended on TPU as it triggers a recompilation for every batch shape that is encountered during training thus significantly slowing down the training. only padding up to the longest example in a batch) leads to very slow training on TPU.

Inference

At inference time, it is recommended to use generate(). This method takes care of encoding the input and feeding the encoded hidden states via cross-attention layers to the decoder and auto-regressively generates the decoder output. Check out this blog post to know all the details about generating text with Transformers. There’s also this blog post which explains how generation works in general in encoder-decoder models.

Copied

>>> from transformers import T5Tokenizer, T5ForConditionalGeneration

>>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
>>> model = T5ForConditionalGeneration.from_pretrained("t5-small")

>>> input_ids = tokenizer("translate English to German: The house is wonderful.", return_tensors="pt").input_ids
>>> outputs = model.generate(input_ids)
>>> print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Das Haus ist wunderbar.

Note that T5 uses the pad_token_id as the decoder_start_token_id, so when doing generation without using generate(), make sure you start it with the pad_token_id.

The example above only shows a single example. You can also do batched inference, like so:

Copied

>>> from transformers import T5Tokenizer, T5ForConditionalGeneration

>>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
>>> model = T5ForConditionalGeneration.from_pretrained("t5-small")

>>> task_prefix = "translate English to German: "
>>> # use different length sentences to test batching
>>> sentences = ["The house is wonderful.", "I like to work in NYC."]

>>> inputs = tokenizer([task_prefix + sentence for sentence in sentences], return_tensors="pt", padding=True)

>>> output_sequences = model.generate(
...     input_ids=inputs["input_ids"],
...     attention_mask=inputs["attention_mask"],
...     do_sample=False,  # disable sampling to test if batching affects output
... )

>>> print(tokenizer.batch_decode(output_sequences, skip_special_tokens=True))
['Das Haus ist wunderbar.', 'Ich arbeite gerne in NYC.']

Because T5 has been trained with the span-mask denoising objective, it can be used to predict the sentinel (masked-out) tokens during inference. The predicted tokens will then be placed between the sentinel tokens.

Copied

>>> from transformers import T5Tokenizer, T5ForConditionalGeneration

>>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
>>> model = T5ForConditionalGeneration.from_pretrained("t5-small")

>>> input_ids = tokenizer("The <extra_id_0> walks in <extra_id_1> park", return_tensors="pt").input_ids

>>> sequence_ids = model.generate(input_ids)
>>> sequences = tokenizer.batch_decode(sequence_ids)
>>> sequences
['<pad><extra_id_0> park offers<extra_id_1> the<extra_id_2> park.</s>']

Performance

If you’d like a faster training and inference performance, install apex and then the model will automatically use apex.normalization.FusedRMSNorm instead of T5LayerNorm. The former uses an optimized fused kernel which is several times faster than the latter.

Resources

A list of official BOINC AI and community (indicated by 🌎) resources to help you get started with T5. If you’re interested in submitting a resource to be included here, please feel free to open a Pull Request and we’ll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.

Text Classification

A notebook for how to finetune T5 for classification and multiple choice.
A notebook for how to finetune T5 for sentiment span extraction. 🌎

Token Classification

A notebook for how to finetune T5 for named entity recognition. 🌎

Text Generation

A notebook for Finetuning CodeT5 for generating docstrings from Ruby code.

Summarization

A notebook to Finetune T5-base-dutch to perform Dutch abstractive summarization on a TPU.
A notebook for how to finetune T5 for summarization in PyTorch and track experiments with WandB. 🌎
A blog post on Distributed Training: Train BART/T5 for Summarization using 🌎 Transformers and Amazon SageMaker.
T5ForConditionalGeneration is supported by this example script and notebook.
TFT5ForConditionalGeneration is supported by this example script and notebook.
FlaxT5ForConditionalGeneration is supported by this example script.
Summarization chapter of the 🌎 BOINC AI course.
Summarization task guide

Fill-Mask

FlaxT5ForConditionalGeneration is supported by this example script for training T5 with a span-masked language model objective. The script also shows how to train a T5 tokenizer. FlaxT5ForConditionalGeneration is also supported by this notebook.

Translation

T5ForConditionalGeneration is supported by this example script and notebook.
TFT5ForConditionalGeneration is supported by this example script and notebook.
Translation task guide

Question Answering

A notebook on how to finetune T5 for question answering with TensorFlow 2. 🌎
A notebook on how to finetune T5 for question answering on a TPU.

🚀 Deploy

A blog post on how to deploy T5 11B for inference for less than $500.

T5

Overview

Training

Inference

Performance

Resources

T5Config

class transformers.T5Config

T5Tokenizer

class transformers.T5Tokenizer

T5TokenizerFast

class transformers.T5TokenizerFast

T5Model

class transformers.T5Model

T5ForConditionalGeneration

class transformers.T5ForConditionalGeneration

T5EncoderModel

class transformers.T5EncoderModel

T5ForSequenceClassification

class transformers.T5ForSequenceClassification

T5ForQuestionAnswering

class transformers.T5ForQuestionAnswering

TFT5Model

class transformers.TFT5Model

TFT5ForConditionalGeneration

class transformers.TFT5ForConditionalGeneration

TFT5EncoderModel

class transformers.TFT5EncoderModel

FlaxT5Model

class transformers.FlaxT5Model

FlaxT5ForConditionalGeneration

class transformers.FlaxT5ForConditionalGeneration

FlaxT5EncoderModel

class transformers.FlaxT5EncoderModel