BinHD: A Binary Learning Framework for Hyperdimensional Computing

BinHD is a Python implementation based on "A Binary Learning Framework for Hyperdimensional Computing" paper.

Usage

Prerequisites

First, install PyTorch using their installation instructions. Then, use the following command to install Torchhd:

pip install torch-hd

Then, use the following command to install BinHD:

pip install binhd

Requirements: PyTorch, Torchhd and ucimlrepo to load datasets from UCI repository.

Quick Start

Iris Example

To quickly get started with BinHD, here's an example using the Iris dataset. Full training code is available in the examples/iris.py file.

import torch
import torch.nn as nn
import torchhd
from torchhd import embeddings
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

from binhd.embeddings import ScatterCode
from binhd.datasets.iris import Iris
from binhd.classifiers import BinHD

# Use the GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Using {} device".format(device))

iris = Iris()
dimension = 1000
num_levels = 100

min_val, max_val = iris.get_min_max_values()
print(min_val, max_val)

class RecordEncoder(nn.Module):
    def __init__(self, out_features, size, levels, low, high):
        super(RecordEncoder, self).__init__() 
        self.position = embeddings.Random(size, out_features, vsa="BSC", dtype=torch.uint8)
        self.value = ScatterCode(levels, out_features, low = low, high = high)
    
    def forward(self, x):
        sample_hv = torchhd.bind(self.position.weight, self.value(x))
        sample_hv = torchhd.multiset(sample_hv)
        return sample_hv

X = iris.features
y = list(iris.labels)

record_encode = RecordEncoder(dimension, iris.num_features, num_levels, min_val, max_val)
record_encode = record_encode.to(device)

with torch.no_grad():
    samples = torch.tensor(X.values).to(device)
    labels = torch.tensor(y).squeeze().to(device)

    X_hv = record_encode(samples)

X_train, X_test, y_train, y_test = train_test_split(X_hv, labels, test_size=0.3, random_state = 0)  

model = BinHD(dimension, iris.num_classes)

with torch.no_grad():
    model.fit(X_train,y_train)
    predictions = model.predict(X_test)  
    acc = accuracy_score(predictions, y_test)
    print("BinHD: Accuracy = ", acc)

    model.fit_adapt(X_train,y_train)
    predictions = model.predict(X_test)  
    acc = accuracy_score(predictions, y_test)
    print("BinHD - Adapt: Accuracy = ", acc)
    

Supported HDC/VSA models

Currently, the library supports the following HDC/VSA models:

• BinHD.