README.md

---
language: en
tags:
- bert
- long context
pipeline_tag: fill-mask
---

# LSG model 
**Transformers >= 4.36.1**\
**This model relies on a custom modeling file, you need to add trust_remote_code=True**\
**See [\#13467](https://github.com/huggingface/transformers/pull/13467)**

LSG ArXiv [paper](https://arxiv.org/abs/2210.15497). \
Github/conversion script is available at this [link](https://github.com/ccdv-ai/convert_checkpoint_to_lsg).

* [Usage](#usage)
* [Parameters](#parameters)
* [Sparse selection type](#sparse-selection-type)
* [Tasks](#tasks)
* [Training global tokens](#training-global-tokens)

This model is adapted from [BERT-base-uncased](https://huggingface.co/bert-base-uncased) without additional pretraining yet. It uses the same number of parameters/layers and the same tokenizer.

This model can handle long sequences but faster and more efficiently than Longformer or BigBird (from Transformers) and relies on Local + Sparse + Global attention (LSG).

The model requires sequences whose length is a multiple of the block size. The model is "adaptive" and automatically pads the sequences if needed (adaptive=True in config). It is however recommended, thanks to the tokenizer, to truncate the inputs (truncation=True) and optionally to pad with a multiple of the block size (pad_to_multiple_of=...).

Support encoder-decoder but I didnt test it extensively.\
Implemented in PyTorch.

![attn](attn.png)

## Usage
The model relies on a custom modeling file, you need to add trust_remote_code=True to use it.

```python: 
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("ccdv/lsg-bert-base-uncased-4096", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("ccdv/lsg-bert-base-uncased-4096")
``` 

## Parameters
You can change various parameters like : 
* the number of global tokens (num_global_tokens=1)
* local block size (block_size=128)
* sparse block size (sparse_block_size=128)
* sparsity factor (sparsity_factor=2)
* mask_first_token (mask first token since it is redundant with the first global token)
* see config.json file

Default parameters work well in practice. If you are short on memory, reduce block sizes, increase sparsity factor and remove dropout in the attention score matrix.

```python:
from transformers import AutoModel

model = AutoModel.from_pretrained("ccdv/lsg-bert-base-uncased-4096", 
    trust_remote_code=True, 
    num_global_tokens=16,
    block_size=64,
    sparse_block_size=64,
    attention_probs_dropout_prob=0.0
    sparsity_factor=4,
    sparsity_type="none",
    mask_first_token=True
)
``` 

## Sparse selection type

There are 6 different sparse selection patterns. The best type is task dependent. \
If `sparse_block_size=0` or `sparsity_type="none"`, only local attention is considered. \
Note that for sequences with length < 2*block_size, the type has no effect.
* `sparsity_type="bos_pooling"` (new)
    * weighted average pooling using the BOS token 
    * Works best in general, especially with a rather large sparsity_factor (8, 16, 32)
    * Additional parameters:
        * None
* `sparsity_type="norm"`, select highest norm tokens
    * Works best for a small sparsity_factor (2 to 4)
    * Additional parameters:
        * None
* `sparsity_type="pooling"`, use average pooling to merge tokens
    * Works best for a small sparsity_factor (2 to 4)
    * Additional parameters:
        * None
* `sparsity_type="lsh"`, use the LSH algorithm to cluster similar tokens
    * Works best for a large sparsity_factor (4+)
    * LSH relies on random projections, thus inference may differ slightly with different seeds
    * Additional parameters:
        * lsg_num_pre_rounds=1, pre merge tokens n times before computing centroids
* `sparsity_type="stride"`, use a striding mecanism per head
    * Each head will use different tokens strided by sparsify_factor
    * Not recommended if sparsify_factor > num_heads
* `sparsity_type="block_stride"`, use a striding mecanism per head
    * Each head will use block of tokens strided by sparsify_factor
    * Not recommended if sparsify_factor > num_heads

## Tasks
Fill mask example:
```python:
from transformers import FillMaskPipeline, AutoModelForMaskedLM, AutoTokenizer

model = AutoModelForMaskedLM.from_pretrained("ccdv/lsg-bert-base-uncased-4096", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("ccdv/lsg-bert-base-uncased-4096")

SENTENCES = "Paris is the [MASK] of France."
pipeline = FillMaskPipeline(model, tokenizer)
output = pipeline(SENTENCES)

> 'Paris is the capital of France.'
```


Classification example:
```python:
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained("ccdv/lsg-bert-base-uncased-4096", 
    trust_remote_code=True, 
    pool_with_global=True, # pool with a global token instead of first token
)
tokenizer = AutoTokenizer.from_pretrained("ccdv/lsg-bert-base-uncased-4096")

SENTENCE = "This is a test for sequence classification. " * 300
token_ids = tokenizer(
    SENTENCE, 
    return_tensors="pt", 
    #pad_to_multiple_of=... # Optional
    truncation=True
    )
output = model(**token_ids)

> SequenceClassifierOutput(loss=None, logits=tensor([[-0.3051, -0.1762]], grad_fn=<AddmmBackward>), hidden_states=None, attentions=None)
```

## Training global tokens
To train global tokens and the classification head only:
```python:
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained("ccdv/lsg-bert-base-uncased-4096", 
    trust_remote_code=True, 
    pool_with_global=True, # pool with a global token instead of first token
    num_global_tokens=16
)
tokenizer = AutoTokenizer.from_pretrained("ccdv/lsg-bert-base-uncased-4096")

for name, param in model.named_parameters():
    if "global_embeddings" not in name:
        param.requires_grad = False
    else:
        param.required_grad = True
```

**BERT**
```
@article{DBLP:journals/corr/abs-1810-04805,
  author    = {Jacob Devlin and
               Ming{-}Wei Chang and
               Kenton Lee and
               Kristina Toutanova},
  title     = {{BERT:} Pre-training of Deep Bidirectional Transformers for Language
               Understanding},
  journal   = {CoRR},
  volume    = {abs/1810.04805},
  year      = {2018},
  url       = {http://arxiv.org/abs/1810.04805},
  archivePrefix = {arXiv},
  eprint    = {1810.04805},
  timestamp = {Tue, 30 Oct 2018 20:39:56 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1810-04805.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}
```