Skip to content

Integrations Module Capabilities

This document details the capabilities of the DeepMIMO integrations module, which provides compatibility with external tools and platforms.

Overview

The integrations module currently supports: - Sionna Adapter: Convert DeepMIMO datasets to Sionna format for use with NVIDIA Sionna - Web Export: Export datasets to binary format for the DeepMIMO web visualizer

Sionna Adapter

Converts DeepMIMO datasets to the format expected by Sionna's neural network training pipelines.

Feature Support Notes
Input Formats
Single Dataset ✓ Single TX-RX pair
MacroDataset ✓ Multiple TX-RX pairs
DynamicDataset ✗ [a] Time-varying not supported
Channel Types
Time domain channels ✓ Required format
Frequency domain channels ✗ [b] Not supported by adapter
Data Extraction
Channel coefficients ✓ Returns (num_rx, num_rx_ant, num_tx, num_tx_ant, num_paths, num_time_steps)
Path delays ✓ Returns (num_rx, num_tx, num_paths)
Path powers ✓ Extracted from channel coefficients
AoA/AoD angles ✗ [c] Not included in Sionna format
Configuration
Multi-BS support ✓ All BSs in dataset included
Multi-UE support ✓ All UEs in dataset included
Subset selection ✓ [d] Subset dataset before passing to adapter
Antenna configuration ✓ Extracts num_rx_ant, num_tx_ant from dataset
Output Format
Sionna channel shape ✓ Matches Sionna's expected dimensions
Batch iteration ✓ __getitem__ for batch access
Length property ✓ Total number of samples
NumPy arrays ✓ Direct NumPy output
Validation
Shape validation ✓ Checks channel dimensions
Antenna count validation ✓ Ensures consistent antenna counts
Channel type validation ✓ Raises error if frequency domain

Web Export

Exports DeepMIMO datasets to binary format for the web-based 3D visualizer.

Feature Support Notes
Input Formats
Single Dataset ✓ Single TX-RX pair
MacroDataset ✓ Multiple TX-RX pairs with metadata
DynamicDataset ✗ [e] Requires snapshot selection
Exported Data
Path power ✓ Per-path power values
Path delays ✓ Per-path delay values
AoA azimuth ✓ Angle of arrival (azimuth)
AoA elevation ✓ Angle of arrival (elevation)
AoD azimuth ✓ Angle of departure (azimuth)
AoD elevation ✓ Angle of departure (elevation)
TX positions ✓ 3D coordinates
RX positions ✓ 3D coordinates
Channel matrices ✗ [f] Not included (size constraint)
Interaction positions ✗ [f] Not included (size constraint)
Binary Format
Custom binary encoding ✓ Optimized format with dtype, shape, data
Float32 precision ✓ Default data type
Float64 precision ✓ Optional data type
Shape preservation ✓ Array dimensions preserved
Metadata JSON ✓ TX/RX set information
TX/RX Set Handling
Single TX/RX set ✓ Simple export
Multiple TX/RX sets ✓ Separate files per set
TX/RX set IDs ✓ Included in metadata
Set count tracking ✓ Total sets in metadata
File naming convention ✓ {dataset_name}_tx{tx_id}_rx{rx_id}.bin
Output Structure
Binary data files ✓ One per TX-RX pair
Metadata file ✓ JSON with set information
Directory organization ✓ Organized by dataset name
File size optimization ✓ Efficient binary encoding
Validation
Array validation ✓ Checks for valid NumPy arrays
Path existence ✓ Creates output directories
File write errors ✓ Exception handling

Usage Examples

Sionna Adapter

import deepmimo as dm
from deepmimo.integrations.sionna_adapter import SionnaAdapter

# Load dataset
dataset = dm.generate('O1_60')

# Create adapter
adapter = SionnaAdapter(dataset)

# Get batch of samples
for batch_idx in range(len(adapter)):
    channels, delays = adapter[batch_idx]
    # Use with Sionna neural network training

Web Export

import deepmimo as dm

# Load dataset
dataset = dm.generate('O1_60')

# Export for web visualizer
dm.export_dataset_to_binary(
    dataset,
    dataset_name='O1_60',
    output_dir='./web_datasets'
)

Notes

  • [a] DynamicDataset support requires selecting specific time snapshot before conversion.
  • [b] Sionna adapter requires time-domain channels. Set freq_domain=False in ch_params when generating.
  • [c] Angle information not exported to Sionna format as it's not used in standard Sionna workflows.
  • [d] To export subset, trim dataset before passing to adapter: dataset.trim(idxs=[...]).
  • [e] DynamicDataset can export individual snapshots: dynamic_dataset[snapshot_idx].
  • [f] Channel matrices and interaction positions omitted from web export due to file size constraints.

Common Limitations

These limitations apply to both integrations:

  • Single frequency: Both integrations work with single-frequency datasets only.
  • Memory constraints: Large datasets may require subsetting before export.
  • Data precision: Web export uses float32 by default for file size optimization.
  • Static snapshots: Dynamic/time-varying datasets require manual snapshot selection.