File size: 3,008 Bytes
09bb874 b8757a8 d98c6b5 e25c9fb d98c6b5 8fa24cb e25c9fb 8fa24cb 312454f e25c9fb d98c6b5 e25c9fb d98c6b5 d6ae002 d73f03f d6ae002 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 |
---
license: apache-2.0
language:
- en
---
## Model weights for Parallel Roberta-Large model ##
We provide the [weights](https://huggingface.co/luffycodes/Parallel-Roberta-Large) for the Parallel Attention and Feedforward design (PAF) for RoBERTa-Large.
To use this model, use the following [paf_modeling_roberta.py](https://github.com/luffycodes/Parallel-Transformers-Pytorch/blob/main/paf_modeling_roberta.py) file.
## Here is how to use this model to get the features of a given text in PyTorch
```python
from transformers import RobertaTokenizer
from paf_modeling_roberta import RobertaModel
tokenizer = RobertaTokenizer.from_pretrained('roberta-large')
model = RobertaModel.from_pretrained('luffycodes/parallel-roberta-large')
text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
```
## Efficient GPU implementation
[gpu_paf_modeling_roberta.py](https://github.com/luffycodes/Parallel-Transformers-Pytorch/blob/main/gpu_paf_modeling_roberta.py) provides an efficient gpu implementation of PAF design for pytorch.
It clubs the computation of key, query, value, and first feedforward network sub-layer(intermediate) computation into one.
```
self.kqv_ffn1.weight.data = torch.cat((attention.self.key.weight.data, attention.self.query.weight.data,
attention.self.value.weight.data,
intermediate.dense.weight.data))
```
However, I could not efficiently optimize the second feedforward network sub-layer computation to run in parallel.
## What is Parallel Attention and Feed-Forward Design?
*On the left is the standard Series Attention and Feed-Forward Net Design (SAF) for transformers models. On the right is the Parallel Attention and Feed-Forward Net Design (PAF) used in transformer models like PaLM (Chowdhery et al., 2022) and Mesh-Transformers (Wang, 2021)*
## Evaluation results of [PAF-RoBERTa-Large](https://huggingface.co/luffycodes/parallel-roberta-large).
When fine-tuned on downstream tasks, this model achieves the following results:
Glue test results:
| Task | MNLI | QQP | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE |
|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|
| | 89.3 | 91.7 | 94.3 | 96.2 | 64.0 | 91.0 | 90.4 | 80.1 |
If you use this work, please cite:
Investigating the Role of Feed-Forward Networks in Transformers Using Parallel Attention and Feed-Forward Net Design:
https://arxiv.org/abs/2305.13297
```
@misc{sonkar2023investigating,
title={Investigating the Role of Feed-Forward Networks in Transformers Using Parallel Attention and Feed-Forward Net Design},
author={Shashank Sonkar and Richard G. Baraniuk},
year={2023},
eprint={2305.13297},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
```
|