Implementation of Token Shift GPT - An autoregressive model that solely relies on shifting the sequence space for mixing

Last update: Oct 14, 2022

Overview

Token Shift GPT

Implementation of Token Shift GPT - An autoregressive model that relies solely on shifting along the sequence dimension and feedforwards.

Update: Inexplicably, it actually works quite well. The feedforward module follows the same design as gMLP, except the feature dimension of the gate tensor is divided up into log2(seq_len) chunks, and the mean pool of the past consecutive segments (length 1, 2, 4, 8, etc. into the past) are shifted into each chunk before a projection along the feature dimension.

Install

$ pip install token-shift-gpt

Usage

import torch
from token_shift_gpt import TokenShiftGPT

model = TokenShiftGPT(
    num_tokens = 256,
    dim = 512,
    max_seq_len = 1024,
    depth = 12,
    ff_mult = 8   # when working with small model dimensions, you may want to increase the intermediate feedforward dimension (here, 8x instead of the usual 4x), so the learning is not bottlenecked by the dimensions of the shifted chunk
)

x = torch.randint(0, 256, (1, 1024))
logits = model(x) # (1, 1024, 256)

To use the discounted cumulative sum approach (which only uses one chunk and seems to be just as effective as the above), just set use_discounted_cumsum = True

First install an additional library

$ pip install torch-discounted-cumsum

Then

import torch
from token_shift_gpt import TokenShiftGPT

model = TokenShiftGPT(
    num_tokens = 256,
    dim = 512,
    max_seq_len = 1024,
    depth = 12,
    ff_mult = 8,
    use_discounted_cumsum = True,
    discounted_gamma = 0.9              # gamma factor for discount
)

x = torch.randint(0, 256, (1, 1024))
logits = model(x) # (1, 1024, 256)

Citations

@misc{yu2021s2mlp,
    title   = {S$^2$-MLP: Spatial-Shift MLP Architecture for Vision}, 
    author  = {Tan Yu and Xu Li and Yunfeng Cai and Mingming Sun and Ping Li},
    year    = {2021},
    eprint  = {2106.07477},
    archivePrefix = {arXiv},
    primaryClass = {cs.CV}
}

@misc{liu2021pay,
    title   = {Pay Attention to MLPs}, 
    author  = {Hanxiao Liu and Zihang Dai and David R. So and Quoc V. Le},
    year    = {2021},
    eprint  = {2105.08050},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG}
}

@software{peng_bo_2021_5196578,
    author       = {PENG Bo},
    title        = {BlinkDL/RWKV-LM: 0.01},
    month        = {aug},
    year         = {2021},
    publisher    = {Zenodo},
    version      = {0.01},
    doi          = {10.5281/zenodo.5196578},
    url          = {https://doi.org/10.5281/zenodo.5196578}
}

You might also like...

Sequence-to-Sequence Framework in PyTorch

nmtpytorch allows training of various end-to-end neural architectures including but not limited to neural machine translation, image captioning and au

395 Nov 21, 2022

Pervasive Attention: 2D Convolutional Networks for Sequence-to-Sequence Prediction

This is a fork of Fairseq(-py) with implementations of the following models: Pervasive Attention - 2D Convolutional Neural Networks for Sequence-to-Se

490 Dec 15, 2022

MASS: Masked Sequence to Sequence Pre-training for Language Generation

1.1k Dec 17, 2022

Sequence-to-Sequence learning using PyTorch

Seq2Seq in PyTorch This is a complete suite for training sequence-to-sequence models in PyTorch. It consists of several models and code to both train

514 Nov 17, 2022

Ptorch NLU, a Chinese text classification and sequence annotation toolkit, supports multi class and multi label classification tasks of Chinese long text and short text, and supports sequence annotation tasks such as Chinese named entity recognition, part of speech tagging and word segmentation.

Pytorch-NLU，一个中文文本分类、序列标注工具包，支持中文长文本、短文本的多类、多标签分类任务，支持中文命名实体识别、词性标注、分词等序列标注任务。 Ptorch NLU, a Chinese text classification and sequence annotation toolkit, supports multi class and multi label classification tasks of Chinese long text and short text, and supports sequence annotation tasks such as Chinese named entity recognition, part of speech tagging and word segmentation.

186 Dec 24, 2022

Code for the paper: Sequence-to-Sequence Learning with Latent Neural Grammars

43 Dec 23, 2022

The official implementation of VAENAR-TTS, a VAE based non-autoregressive TTS model.

VAENAR-TTS This repo contains code accompanying the paper "VAENAR-TTS: Variational Auto-Encoder based Non-AutoRegressive Text-to-Speech Synthesis". Sa

138 Oct 28, 2022

An implementation of model parallel GPT-3-like models on GPUs, based on the DeepSpeed library. Designed to be able to train models in the hundreds of billions of parameters or larger.

GPT-NeoX An implementation of model parallel GPT-3-like models on GPUs, based on the DeepSpeed library. Designed to be able to train models in the hun

3.1k Jan 8, 2023

Python package to easily retrain OpenAI's GPT-2 text-generating model on new texts

gpt-2-simple A simple Python package that wraps existing model fine-tuning and generation scripts for OpenAI's GPT-2 text generation model (specifical

3.1k Jan 7, 2023

Comments

self.gate is never used?

it seems like self.gate is never actually used, or am I missing something?

https://github.com/lucidrains/token-shift-gpt/blob/1449b263f1fb222279d00f9128c29f25dbef976b/token_shift_gpt/token_shift_gpt.py#L79

opened by inspirit 1

Implementation of Token Shift GPT - An autoregressive model that solely relies on shifting the sequence space for mixing

Related tags

Overview

Token Shift GPT

Install

Usage

Citations

You might also like...

Sequence-to-Sequence Framework in PyTorch

Pervasive Attention: 2D Convolutional Networks for Sequence-to-Sequence Prediction

MASS: Masked Sequence to Sequence Pre-training for Language Generation

Sequence-to-Sequence learning using PyTorch

Ptorch NLU, a Chinese text classification and sequence annotation toolkit, supports multi class and multi label classification tasks of Chinese long text and short text, and supports sequence annotation tasks such as Chinese named entity recognition, part of speech tagging and word segmentation.

Code for the paper: Sequence-to-Sequence Learning with Latent Neural Grammars

The official implementation of VAENAR-TTS, a VAE based non-autoregressive TTS model.

An implementation of model parallel GPT-3-like models on GPUs, based on the DeepSpeed library. Designed to be able to train models in the hundreds of billions of parameters or larger.

Python package to easily retrain OpenAI's GPT-2 text-generating model on new texts

Comments

self.gate is never used?

Releases(0.0.2)

0.0.2(Aug 18, 2021)

0.0.1(Aug 18, 2021)

Owner

Phil Wang

KoBERT - Korean BERT pre-trained cased (KoBERT)

HAIS_2GNN: 3D Visual Grounding with Graph and Attention

DiY Oxygen Concentrator based on the OxiKit

Library for Russian imprecise rhymes generation

NAACL 2022: MCSE: Multimodal Contrastive Learning of Sentence Embeddings

Create a semantic search engine with a neural network (i.e. BERT) whose knowledge base can be updated

This code is the implementation of Text Emotion Recognition (TER) with linguistic features

硕士期间自学的NLP子任务，供学习参考

Extract rooms type, door, neibour rooms, rooms corners nad bounding boxes, and generate graph from rplan dataset

Beyond Paragraphs: NLP for Long Sequences

Tensorflow implementation of paper: Learning to Diagnose with LSTM Recurrent Neural Networks.

Fake news detector filters - Smart filter project allow to classify the quality of information and web pages

Code for "Generating Disentangled Arguments with Prompts: a Simple Event Extraction Framework that Works"

Sequence-to-Sequence learning using PyTorch

Translates basic English sentences into the Huna language (hoo-NAH)

Tools to download and cleanup Common Crawl data

edge-SR: Super-Resolution For The Masses

Code for the ACL 2021 paper "Structural Guidance for Transformer Language Models"

PyTorch implementation and pretrained models for XCiT models. See XCiT: Cross-Covariance Image Transformer

A method for cleaning and classifying text using transformers.