/h-net-dynamic-chunking

Implementation of the dynamic chunking mechanism in H-net by Hwang et al. of Carnegie Mellon

Primary LanguagePythonMIT LicenseMIT

H-Net Dynamic Chunking

Implementation of the dynamic chunking mechanism in H-net by Hwang et al. of Carnegie Mellon

Install

$ pip install h-net-dynamic-chunking

Usage

import torch
from h_net_dynamic_chunking import DynamicSequenceChunker

downsampler = DynamicSequenceChunker(512)

tokens = torch.randn(3, 1024, 512).requires_grad_()

downsampled, upsample_fn, *_ = downsampler(tokens)

assert upsample_fn(downsampled).shape == tokens.shape

3 layers hierarchy

import torch
from h_net_dynamic_chunking import DynamicSequenceChunker

downsampler1 = DynamicSequenceChunker(512)
downsampler2 = DynamicSequenceChunker(512)
downsampler3 = DynamicSequenceChunker(512)

tokens = torch.randn(3, 1024, 512).requires_grad_()

downsampled1, upsample_fn1, aux_loss1 = downsampler1(tokens)

# hierarchical network 1 ...

downsampled2, upsample_fn2, aux_loss2 = downsampler2(downsampled1)

# hierarchical network 2 ...

downsampled3, upsample_fn3, aux_loss3 = downsampler3(downsampled2)

# inner most network

# reconstituting

assert upsample_fn1(upsample_fn2(upsample_fn3(downsampled3))).shape == tokens.shape

HNet wrapper

import torch
from torch import nn
from h_net_dynamic_chunking.h_net import HNet

# 3 hierarchies, from 512 -> 1024 -> 2048 inner

net = HNet(
    nn.Identity(),
    HNet(
        nn.Identity(),
        HNet(
            nn.Identity(),
            nn.Identity(),
            nn.Identity(),
            dim = 2048
        ),
        nn.Identity(),
        dim = 1024,
        dim_inner = 2048
    ),
    nn.Identity(),
    dim = 512,
    dim_inner = 1024,
)

tokens = torch.randn(1, 1024, 512)

out, aux_loss = net(tokens) # (1, 1024, 512), (1,)

Example

Enwik8 with 2 hierarchies

$ pip install '.[examples]'

Then

$ python train.py

Citations

@misc{hwang2025dynamicchunkingendtoendhierarchical,
    title   = {Dynamic Chunking for End-to-End Hierarchical Sequence Modeling},
    author  = {Sukjun Hwang and Brandon Wang and Albert Gu},
    year    = {2025},
    eprint  = {2507.07955},
    archivePrefix = {arXiv},
    primaryClass = {cs.LG},
    url     = {https://arxiv.org/abs/2507.07955},
}