|
--- |
|
datasets: |
|
- multi_nli |
|
- snli |
|
- scitail |
|
metrics: |
|
- accuracy |
|
- f1 |
|
pipeline_tag: zero-shot-classification |
|
language: |
|
- en |
|
--- |
|
# T5ForSequenceClassification |
|
**T5ForSequenceClassification** adapts the original [T5](https://github.com/google-research/text-to-text-transfer-transformer) architecture for sequence classification tasks. |
|
|
|
T5 was originally built for text-to-text tasks and excels in it. |
|
It can handle any NLP task if it has been converted to a text-to-text format, including sequence classification task! |
|
You can find [here](https://huggingface.co/google/flan-t5-base?text=Premise%3A++At+my+age+you+will+probably+have+learnt+one+lesson.+Hypothesis%3A++It%27s+not+certain+how+many+lessons+you%27ll+learn+by+your+thirties.+Does+the+premise+entail+the+hypothesis%3F) how the original T5 is used for sequence classification task. |
|
|
|
Our motivations for building **T5ForSequenceClassification** is that the full original T5 architecture is not needed for most NLU tasks. Indeed, NLU tasks generally do not require to generate text and thus a large decoder is unnecessary. |
|
By removing the decoder we can *half the original number of parameters* (thus half the computation cost) and *efficiently optimize* the network for the given task. |
|
|
|
## Table of Contents |
|
|
|
1. [Why use T5ForSequenceClassification?](##why-use-t5forsequenceclassification?) |
|
2. [T5ForClassification vs T5](##t5forclassification-vs-t5) |
|
|
|
## Why use T5ForSequenceClassification? |
|
Models based on the [BERT](https://huggingface.co/bert-large-uncased) architecture like [RoBERTa](https://huggingface.co/roberta-large) and [DeBERTa](https://huggingface.co/microsoft/deberta-v2-xxlarge) have shown very strong performance on sequence classification task and are still widely used today. |
|
However, those models only scale up to ~1.5B parameters (DeBERTa xxlarge) resulting in a limited knowledge compare to bigger models. |
|
On the other hand, models based on the T5 architecture scale up to ~11B parameters (t5-xxl) and innovations with this architecture are very recent and keeps improving ([mT5](https://huggingface.co/google/mt5-xxl), [Flan-T5](https://huggingface.co/google/flan-t5-xxl), [UL2](https://huggingface.co/google/ul2), [Flan-UL2](https://huggingface.co/google/flan-ul2), and probably more...) |
|
|
|
## T5ForClassification vs T5 |
|
**T5ForClassification** Architecture: |
|
- Encoder: same as original T5 |
|
- Decoder: only the first layer (for pooling purpose) |
|
- Classification head: simple Linear layer on top of the decoder |
|
|
|
Benefits and Drawbacks: |
|
- (**+**) Keeps T5 encoding strength |
|
- (**+**) Parameters size is half |
|
- (**+**) Interpretable outputs (class logits) |
|
- (**+**) No generation mistakes and faster prediction (no generation latency) |
|
- (**-**) Looses text-to-text ability |
|
|
|
|
|
Special thanks to [philschmid](https://huggingface.co/philschmid) for making a Flan-T5-xxl [checkpoint](https://huggingface.co/philschmid/flan-t5-xxl-sharded-fp16) in fp16. |
|
|
