L

Lwm V1.1

Developed by wi-lab
LWM 1.1 is an upgraded pre-trained model specifically designed for wireless channel feature extraction, supporting diverse channel configurations to enhance feature extraction quality and generalization capabilities.
Downloads 277
Release Time : 4/25/2025

Model Overview

LWM 1.1 is a Transformer-based foundational model for wireless channels, learning spatial and frequency dependencies of wireless channels through Masked Channel Modeling (MCM) pre-training, suitable for various wireless communication and sensing tasks.

Model Features

Enhanced Input Flexibility
Supports various (N, SC) configurations with sequence length increased to 512, adapting to the diverse configurations of real-world wireless systems.
Dataset and Pre-training Enhancement
Training scenarios increased from 15 to 140, masking ratio raised to 40%, with pre-training samples reaching 1.05 million, significantly improving cross-environment generalization.
Model Architecture Optimization
Parameter count increased to 2.5 million, employing 2D block processing technology to cover both antenna and subcarrier dimensions, enhancing spatial-frequency feature learning.
Training and Efficiency Optimization
Utilizes AdamW optimizer with cosine decay strategy and bucket batching to optimize memory usage, balancing computational cost and feature extraction capability.
Task Adaptability
Supports freezing specific layers for targeted fine-tuning, provides default classification and regression heads, and allows user-defined modules.

Model Capabilities

Wireless Channel Feature Extraction
Spatial and Frequency Dependency Modeling
LoS/NLoS Classification
Beam Prediction
Few-shot Learning

Use Cases

Wireless Communication
LoS/NLoS Classification
Classification task based on (32, 32) channels between BS 3 and 8,299 users in DeepMIMO's Denver dense scenario.
Embedding features based on LWM show significant advantages compared to raw channel data.
Beam Prediction
Predicts optimal beam direction using pre-trained embedding features.
Wireless Sensing
Environmental Sensing
Extracts environmental features from channel data for scene recognition or user localization.
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
© 2025AIbase