wineslab / deepbeam

M. Polese, F. Restuccia, and T. Melodia, "DeepBeam: Deep Waveform Learning for Coordination-Free Beam Management in mmWave Networks", Proc. of ACM Intl. Symp. on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing (ACM MobiHoc), July 2021.

DeepBeam data and repo

The code in this repository has been used to generate the results for the paper

M. Polese, F. Restuccia, and T. Melodia, "DeepBeam: Deep Waveform Learning for Coordination-Free Beam Management in mmWave Networks", Proc. of ACM Intl. Symp. on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing (ACM MobiHoc), July 2021.

The associated dataset can be found at this link.

Please reference the paper if you use the code or data from the dataset: bibtex entry

Dataset structure

The DeepBeam dataset can be found at this link.

It contains 19 HDF5 files that represent a data collection campaign run on the NI mmWave Transceiver System with four SiBeam 60 GHz radio heads and on two Pi-Radio digital 60 GHz radios. The data collection campaign is described in Section 4 of the DeepBeam paper.

The data is organized as follows.

NI-based dataset for TXB classification

Each HDF5 file contains I/Q samples corresponding to 3 (parameter num_gains in the scripts) receiver gain values (40 dB, 50 dB, 60 dB) to represent three different received SNR values (i.e., in a range between -15 dB and 20 dB) and 24 TX beams or 12 TX beams.

The files are organized using HDF5 datasets. Each file contains four datasets

• iq contains the I/Q samples (one column for the in-phase (I) samples, the other for the quadrature (Q) samples)
• tx_beam contains a label with the transmit beam used for the corresponding I/Q sample (i.e., entry N in the tx_beam dataset corresponds to entry N in the iq dataset)
• rx_beam contains a label with the receive beam used for the corresponding I/Q sample (i.e., entry N in the rx_beam dataset corresponds to entry N in the iq dataset)
• gain contains a label with the receiver gain value for the corresponding I/Q sample (i.e., entry N in the gain dataset corresponds to entry N in the iq dataset)

The total number of entries in each dataset depends on whether 24 or 12 TX beams are used (parameter num_beams in the scripts). For each (gain, tx_beam) pair, we collected 10000 frames (parameter num_frames_for_gain_tx_beam_pair in the scripts). Each frame contains 15 blocks (parameter num_blocks_per_frame in the scripts). Each frame contains 2048 I/Q samples (parameter num_samples_per_block in the scripts). Therefore, the total number of entries is num_gains * num_beams * num_frames_for_gain_tx_beam_pair * num_blocks_per_frame * num_samples_per_block.

The I/Q samples are arranged sequentially, according to the following logic:

For gain in [40, 50, 60]:
    For tx_beam in 0:num_beams:
        Store num_frames_for_gain_tx_beam_pair * num_blocks_per_frame * num_samples_per_block of the (gain, tx_beam) pair

The receive beam is fixed (boresight of the antenna array).

Basic configuration (see Figure 7 of the DeepBeam paper)

For the basic configuration, we provide the 24 and 12 TX beams data sets with 4 different configurations of the SiBeam 60 GHz heads:

• 24 TX beams

  • TX antenna 0, RX antenna 1 srf-basic-config-24-beams-tx-ant-0-rx-ant-1.h5
  • TX antenna 1, RX antenna 0 srf-basic-config-24-beams-tx-ant-1-rx-ant-0.h5
  • TX antenna 2, RX antenna 1 srf-basic-config-24-beams-tx-ant-2-rx-ant-1.h5
  • TX antenna 3, RX antenna 1 srf-basic-config-24-beams-tx-ant-3-rx-ant-1.h5

• 12 TX beams

  • TX antenna 0, RX antenna 1 srf-basic-config-12-beams-tx-ant-0-rx-ant-1.h5
  • TX antenna 1, RX antenna 0 srf-basic-config-12-beams-tx-ant-1-rx-ant-0.h5
  • TX antenna 2, RX antenna 1 srf-basic-config-12-beams-tx-ant-2-rx-ant-1.h5
  • TX antenna 3, RX antenna 1 srf-basic-config-12-beams-tx-ant-3-rx-ant-1.h5

Diagonal configuration (see Figure 7 of the DeepBeam paper)

For the diagonal configuration, we provide the 24 and 12 TX beams data sets with one configuration of the SiBeam 60 GHz heads:

• 24 TX beams

  • TX antenna 0, RX antenna 1 srf-diagonal-config-24-beams-tx-ant-0-rx-ant-1.h5

• 12 TX beams

  • TX antenna 0, RX antenna 1 srf-diagonal-config-12-beams-tx-ant-0-rx-ant-1.h5

Obstacle configuration (see Figure 7 of the DeepBeam paper)

For the obstacle configuration, we provide the 24 and 12 TX beams data sets with one configuration of the SiBeam 60 GHz heads:

• 24 TX beams

  • TX antenna 0, RX antenna 1 srf-obstacle-config-24-beams-tx-ant-0-rx-ant-1.h5

• 12 TX beams

  • TX antenna 0, RX antenna 1 srf-obstacle-config-12-beams-tx-ant-0-rx-ant-1.h5

Pi-Radio-based dataset for TXB classification

A single HDF5 file is available for the Pi-Radio TXB data set. It contains I/Q samples corresponding to31 (parameter num_gains in the scripts) transmitter gain values and 5 TX beams (parameter num_beams in the scripts).

The files are organized using HDF5 datasets. Each file contains four datasets

• iq contains the I/Q samples (one column for the in-phase (I) samples, the other for the quadrature (Q) samples)
• tx_beam contains a label with the transmit beam used for the corresponding I/Q sample (i.e., entry N in the tx_beam dataset corresponds to entry N in the iq dataset)
• rx_beam contains a label with the receive beam used for the corresponding I/Q sample (i.e., entry N in the rx_beam dataset corresponds to entry N in the iq dataset)
• gain contains a label with the receiver gain value for the corresponding I/Q sample (i.e., entry N in the gain dataset corresponds to entry N in the iq dataset)

For each (gain, tx_beam) pair, we collected 10000 frames (parameter num_frames_for_gain_tx_beam_pair in the scripts). Each frame contains 5 blocks (parameter num_blocks_per_frame in the scripts). Each frame contains 2048 I/Q samples (parameter num_samples_per_block in the scripts). Therefore, the total number of entries is num_gains * num_beams * num_frames_for_gain_tx_beam_pair * num_blocks_per_frame * num_samples_per_block.

The I/Q samples are arranged sequentially, according to the following logic:

For gain in ['att-tx-0-0-', 'att-tx-5-0-', 'att-tx-5-4-']: # these values correspond to increasing attenuation values
    For tx_beam in 0:num_beams:
        Store num_frames_for_gain_tx_beam_pair * num_blocks_per_frame * num_samples_per_block of the (gain, tx_beam) pair

The receive beam is fixed (boresight of the antenna array).

The Pi-Radio-based dataset is in the file mrf-basic-config-5-beams.h5 for the configuration shown in Figure 8 of the DeepBeam paper.

NI-based dataset for AoA classification

Each HDF5 file contains I/Q samples corresponding to 3 (parameter num_gains in the scripts) receiver gain values (40 dB, 50 dB, 60 dB) to represent three different received SNR values (i.e., in a range between -15 dB and 20 dB), 3 TX beams for the 24 TX beams codebook, and 3 AoA values (-45, 0, 45) (parameter num_angles in the scripts).

The files are organized using HDF5 datasets. Each file contains four datasets

• iq contains the I/Q samples (one column for the in-phase (I) samples, the other for the quadrature (Q) samples)
• tx_beam contains a label with the transmit beam used for the corresponding I/Q sample (i.e., entry N in the tx_beam dataset corresponds to entry N in the iq dataset)
• rx_beam contains a label with the receive beam used for the corresponding I/Q sample (i.e., entry N in the rx_beam dataset corresponds to entry N in the iq dataset)
• gain contains a label with the receiver gain value for the corresponding I/Q sample (i.e., entry N in the gain dataset corresponds to entry N in the iq dataset)
• angle contains a label with the receiver angle value for the corresponding I/Q sample (i.e., entry N in the angle dataset corresponds to entry N in the iq dataset)

In this case, num_beams is 3. For each (gain, tx_beam, angle) tuple, we collected 10000 frames (parameter num_frames_for_gain_tx_beam_pair in the scripts). Each frame contains 15 blocks (parameter num_blocks_per_frame in the scripts). Each frame contains 2048 I/Q samples (parameter num_samples_per_block in the scripts). Therefore, the total number of entries is num_gains * num_beams * num_angles * num_frames_for_gain_tx_beam_pair * num_blocks_per_frame * num_samples_per_block.

The I/Q samples are arranged sequentially, according to the following logic:

For gain in [40, 50, 60]:
    For tx_beam in [4, 12, 20]:
        For angle in [-45, 0, 45]:
            Store num_frames_for_gain_tx_beam_pair * num_blocks_per_frame * num_samples_per_block of the (gain, tx_beam, angle) tuple

The receive beam is fixed (boresight of the antenna array).

Basic configuration (see Figure 7 of the DeepBeam paper)

For the basic configuration, we provide the AoA data sets with 4 different configurations of the SiBeam 60 GHz heads:

• 24 TX beams
  • TX antenna 0, RX antenna 1 srf-basic-config-24-beams-aoa-tx-ant-0-rx-ant-1.h5
  • TX antenna 1, RX antenna 0 srf-basic-config-24-beams-aoa-tx-ant-1-rx-ant-0.h5
  • TX antenna 2, RX antenna 1 srf-basic-config-24-beams-aoa-tx-ant-2-rx-ant-1.h5
  • TX antenna 3, RX antenna 1 srf-basic-config-24-beams-aoa-tx-ant-3-rx-ant-1.h5

Diagonal configuration (see Figure 7 of the DeepBeam paper)

For the diagonal configuration, we provide the AoA data sets with one configuration of the SiBeam 60 GHz heads:

• 24 TX beams
  • TX antenna 0, RX antenna 1 srf-diagonal-config-24-beams-aoa-tx-ant-0-rx-ant-1.h5

Obstacle configuration (see Figure 7 of the DeepBeam paper)

For the obstacle configuration, we provide the AoA data sets with one configuration of the SiBeam 60 GHz heads:

• 24 TX beams
  • TX antenna 0, RX antenna 1 srf-obstacle-config-24-beams-aoa-tx-ant-0-rx-ant-1.h5

Requirements

See the requirements.txt file.

Source code structure

The source code can be found in the keras_code folder. We provide bash and Python scripts to run training and testing, as well as auxiliary Python code with DataGenerators to automate the parsing of the data.

TXB Classification Training and Testing

• Training for the TXB classification involves the launch_deepbeam.sh bash script, which sets the relevant variables and calls DeepBeam.py. This Python script accepts as input a number of parameters (documented in the DeepBeam.py file), runs training over a specific HDF5 file, and saves the model. It is possible to specify the input dataset (which will be parsed by the DataGenerator.py) and the folder to store the model parameters. Please refer to launch_deepbeam.sh and DeepBeam.py for a complete list of parameters.

• Testing is performed by calling launch_deepbeam_testing.sh. It is possible to specify the input dataset (which will be parsed by the DataGenerator.py) and the folder to store the output accuracy. Please refer to launch_deepbeam_testing.sh and DeepBeamTesting.py for a complete list of parameters.

AoA Classification Training and Testing

• Training for the AoA classification involves the launch_deepbeam_aoa.sh bash script, which sets the relevant variables and calls DeepBeamAoa.py. This Python script accepts as input a number of parameters (documented in the DeepBeamAoa.py file), runs training over a specific HDF5 file, and saves the model. It is possible to specify the input dataset (which will be parsed by the DataGeneratorAoa.py) and the folder to store the model parameters. Please refer to launch_deepbeam_aoa.sh and DeepBeamAoa.py for a complete list of parameters.

• Testing is performed by calling launch_deepbeam_testing.sh. It is possible to specify the input dataset (which will be parsed by the DataGeneratorAoa.py) and the folder to store the output accuracy. Please refer to launch_deepbeam_aoa_testing.sh and DeepBeamTesting.py for a complete list of parameters.

Mixed TXB/AoA Training and Testing

For training and testing over a combination of multiple HDF5 files, we rely on DataGeneratorCross.py as DataGenerator. It combines multiple generators to extract data consistently from different HDF5 files.

• Training for the TXB classification involves the launch_deepbeam_mixed.sh bash script, which sets the relevant variables and calls DeepBeamMixed.py. This Python script accepts as input a number of parameters (documented in the DeepBeamMixed.py file), runs training over multiple HDF5 file, and saves the model. It is possible to specify the input datasets (which will be parsed by the DataGeneratorCross.py) and the folder to store the model parameters. Please refer to launch_deepbeam_mixed.sh and DeepBeamMixed.py for a complete list of parameters.

• Testing is performed by calling launch_deepbeam_mixed_testing.sh. It is possible to specify multiple input datasets (which will be parsed by the DataGeneratorCross.py) and the folder to store the output accuracy. Please refer to launch_deepbeam_mixed_testing.sh and DeepBeamMixedTesting.py for a complete list of parameters.

Other files

The DataGeneratorTesting.py (or similar) can be used to test the DataGenerator classes. PlotConfusionMatrix.py plots the confusion matrix. Utils.py contains functions to create models. launch_filters.sh and DeepBeamGetFilters.py extract the filter values from a specific model.