resmps's Introduction

ResMPS: Residual Matrix Product State

ResMPS (short for 'Residual Matrix Product State') is a PyTorch based tensor network machine learning architecture.

The idea of ResMPS is inpired by residual networks, and outperforms the state-of-the-art tensor network models in terms of efficiency, stability, and expression power.

For further details, please see our paper SciPost Physics 14.6 (2023): 142.

Installation

Install dependencies.

torch
torchvision
hilbertcurve
matplotlib
sklearn
dill

The cuda version of PyTorch is recommended.

Clone this repo.

git clone https://github.com/YemingMeng/ResMPS.git

Usage/Examples

Print help of usage information.

python main.py -h

Typical command example, here cuda for GPU acceleration and fashion_mnist is the dataset.

python main.py --device cuda --dataset fashion_mnist

To reproduce results in arxiv:2012.11841, run

python examples.py

Use ResMPS in general case, create your python script like this

from ResMPS import ResMPS

nfeatures = 2     # dimension of the feature map function
nchi      = 12    # dimension of the virtual feature
nrow      = 28    # number of rows of the input image
ncol      = 28    # number of columns of the input image
nlength   = 785   # length of ResMPS, needs to match the dimension of input
noutput   = 10    # dimension of the output, i.e., the total number of categories

batch_size = 1000 # input batch size
cf = ResMPS(nfeatures, nchi, nlength, noutput) # create an instance of ResMPS
input = th.rand(batch_size, 1, nrow, ncol, device=cf.device) # generate random input data
output = cf(input)  # data processing by using ResMPS
print(output.shape) # check the shape of output

Citing ResMPS

Include this BibTex Entry to your .bib file

@misc{https://doi.org/10.48550/arxiv.2012.11841,
  doi = {10.48550/ARXIV.2012.11841},
  url = {https://arxiv.org/abs/2012.11841},
  author = {Meng, Ye-Ming and Zhang, Jing and Zhang, Peng and Gao, Chao and Ran, Shi-Ju},
  title = {Residual Matrix Product State for Machine Learning},
  publisher = {arXiv},
  year = {2020},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

License

resmps's People

Contributors

Stargazers

Watchers

resmps's Issues

When setting '--classical' to 'True', an error occurred.

您好，当我想要测试regular MPS性能时发生了错误：

在已有代码的基础上，我仅修改了以下的一行代码，训练结果崩溃了。

# parser.add_argument('--classical', default=False, action='store_true', ...)
parser.add_argument('--classical', default=True, action='store_true', ...)

结果输出：

learning rate=0.0001
starting training process...
(step/total)=(  60/  60), batch accuracy = 10.20%.
starting evaluating process...
train set accuracy = 10.00%
test  set accuracy = 10.00%
train set loss     = 2.30258
test  set loss     = 2.30258
convergence factor = 0.00000
run epoch_step in 25.25 seconds

从step 1开始，所有结果都约为10%，我重复了多次，且尝试了不同的数据集，都得到了类似的结果。当将classical设置为False后，一切恢复正常，请问是什么问题呢？

Recommend Projects