Implementation of Hierarchical Transformer Memory (HTM) for Pytorch

Last update: Dec 29, 2022

Overview

Hierarchical Transformer Memory (HTM) - Pytorch

Implementation of Hierarchical Transformer Memory (HTM) for Pytorch. This Deepmind paper proposes a simple method to allow transformers to attend to memories of the past efficiently. Original Jax repository

Install

$ pip install htm-pytorch

Usage

import torch
from htm_pytorch import HTMAttention

attn = HTMAttention(
    dim = 512,
    heads = 8,               # number of heads for within-memory attention
    dim_head = 64,           # dimension per head for within-memory attention
    topk_mems = 8,           # how many memory chunks to select for
    mem_chunk_size = 32,     # number of tokens in each memory chunk
    add_pos_enc = True       # whether to add positional encoding to the memories
)

queries = torch.randn(1, 128, 512)     # queries
memories = torch.randn(1, 20000, 512)  # memories, of any size
mask = torch.ones(1, 20000).bool()     # memory mask

attended = attn(queries, memories, mask = mask) # (1, 128, 512)

If you want the entire HTM Block (which contains the layernorm for the input followed by a skip connection), just import HTMBlock instead

import torch
from htm_pytorch import HTMBlock

block = HTMBlock(
    dim = 512,
    topk_mems = 8,
    mem_chunk_size = 32
)

queries = torch.randn(1, 128, 512)
memories = torch.randn(1, 20000, 512)
mask = torch.ones(1, 20000).bool()

out = block(queries, memories, mask = mask) # (1, 128, 512)

Citations

@misc{lampinen2021mental,
    title   = {Towards mental time travel: a hierarchical memory for reinforcement learning agents}, 
    author  = {Andrew Kyle Lampinen and Stephanie C. Y. Chan and Andrea Banino and Felix Hill},
    year    = {2021},
    eprint  = {2105.14039},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG}
}

Segcache: a memory-efficient and scalable in-memory key-value cache for small objects

Segcache: a memory-efficient and scalable in-memory key-value cache for small objects This repo contains the code of Segcache described in the followi

78 Jan 7, 2023

Episodic-memory - Ego4D Episodic Memory Benchmark

Ego4D Episodic Memory Benchmark EGO4D is the world's largest egocentric (first p

3 Feb 18, 2022

Implementation of H-Transformer-1D, Hierarchical Attention for Sequence Learning

H-Transformer-1D Implementation of H-Transformer-1D, Transformer using hierarchical Attention for sequence learning with subquadratic costs. For now,

123 Nov 17, 2022

Implementation of H-Transformer-1D, Hierarchical Attention for Sequence Learning using 🤗 transformers

hierarchical-transformer-1d Implementation of H-Transformer-1D, Hierarchical Attention for Sequence Learning using 🤗 transformers In Progress!! 2021.

7 Nov 6, 2022

VSR-Transformer - This paper proposes a new Transformer for video super-resolution (called VSR-Transformer).

VSR-Transformer By Jiezhang Cao, Yawei Li, Kai Zhang, Luc Van Gool This paper proposes a new Transformer for video super-resolution (called VSR-Transf

225 Nov 13, 2022

HiFT: Hierarchical Feature Transformer for Aerial Tracking (ICCV2021)

HiFT: Hierarchical Feature Transformer for Aerial Tracking Ziang Cao, Changhong Fu, Junjie Ye, Bowen Li, and Yiming Li Our paper is Accepted by ICCV 2

Intelligent Vision for Robotics in Complex Environment

55 Nov 23, 2022

The official code for paper "R2D2: Recursive Transformer based on Differentiable Tree for Interpretable Hierarchical Language Modeling".

R2D2 This is the official code for paper titled "R2D2: Recursive Transformer based on Differentiable Tree for Interpretable Hierarchical Language Mode

49 Dec 17, 2022

Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image classification, in Pytorch

Transformer in Transformer Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image c

272 Dec 23, 2022

Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

12.6k Jan 9, 2023

Comments

auto-regressive use case
Hi Phil! I was wondering if HTM part can be used in/for auto-regressive scenario? Full proposed arch in the paper has 3 blocks:

Self Att - this can be easily done with causal masking

next we have HTM block with memories - can it be used in autoregressive scenario i wonder?

Feed Forward block

please let me know your thoughts?
opened by inspirit 0

Implementation of Hierarchical Transformer Memory (HTM) for Pytorch

Related tags

Overview

Hierarchical Transformer Memory (HTM) - Pytorch

Install

Usage

Citations

You might also like...

Segcache: a memory-efficient and scalable in-memory key-value cache for small objects

Episodic-memory - Ego4D Episodic Memory Benchmark

Implementation of H-Transformer-1D, Hierarchical Attention for Sequence Learning

Implementation of H-Transformer-1D, Hierarchical Attention for Sequence Learning using 🤗 transformers

VSR-Transformer - This paper proposes a new Transformer for video super-resolution (called VSR-Transformer).

HiFT: Hierarchical Feature Transformer for Aerial Tracking (ICCV2021)

The official code for paper "R2D2: Recursive Transformer based on Differentiable Tree for Interpretable Hierarchical Language Modeling".

Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image classification, in Pytorch

Implementation of Vision Transformer, a simple way to achieve SOTA in vision classification with only a single transformer encoder, in Pytorch

Comments

auto-regressive use case

Releases(0.0.4)

0.0.4(Sep 15, 2021)

0.0.3(Sep 14, 2021)

0.0.2(Sep 14, 2021)

0.0.1(Sep 14, 2021)

Owner

Phil Wang

Details about the wide minima density hypothesis and metrics to compute width of a minima

MPRNet-Cloud-removal: Progressive cloud removal

PyTorch implementation of the TTC algorithm

Code release of paper Improving neural implicit surfaces geometry with patch warping

g9.py - Torch interactive graphics

UniLM AI - Large-scale Self-supervised Pre-training across Tasks, Languages, and Modalities

This is the official Pytorch implementation of the paper "Diverse Motion Stylization for Multiple Style Domains via Spatial-Temporal Graph-Based Generative Model"

This is a Keras implementation of a CNN for estimating age, gender and mask from a camera.

Multiview Neural Surface Reconstruction by Disentangling Geometry and Appearance

Self-Supervised Multi-Frame Monocular Scene Flow (CVPR 2021)

YKKDetector For Python

My solution for the 7th place / 245 in the Umoja Hack 2022 challenge

Official implementation for paper: A Latent Transformer for Disentangled Face Editing in Images and Videos.

A simple algorithm for extracting tree height in sparse scene from point cloud data.

Locally Differentially Private Distributed Deep Learning via Knowledge Distillation (LDP-DL)

A PyTorch implementation of DenseNet.

Unified file system operation experience for different backend

Plenoxels: Radiance Fields without Neural Networks

PyTorch implementation of a collections of scalable Video Transformer Benchmarks.

Open source implementation of AceNAS: Learning to Rank Ace Neural Architectures with Weak Supervision of Weight Sharing