Tensorflow implementation of Deep Transfer Learning for Site-Specific Channel Estimation in Low-Resolution mmWave MIMO
Abstract — We consider the problem of channel estimation in low-resolution multiple-input multiple-output (MIMO) systems operating at millimeter wave (mmWave) and present a deep transfer learning (DTL) approach that exploits previously trained models to speed up site adaptation. The proposed model is composed of a feature extractor and a regressor, with only the regressor requiring training for the new environment. The DTL approach is evaluated using two 3D scenarios where ray-tracing is performed to generate the mmWave MIMO channels used in the simulations. Under the defined testing setup, the proposed DTL approach can reduce the computational cost of the training stage without decreasing the estimation accuracy.
First, run python wr_fres.py to train models.
Then, run python pretrained_model.py to apply deep transfer learning
akpy - software to process mimo channels
mimoNrxNt - experiments outputs
models - save models trained
| Filename | Description |
|---|---|
| wr_fres.py | main file to train models |
| pretrained_model.py | apply deep transfer learning in trained models |
| get_channels.py | extract channels data from Raymobtime dataset |
| mimo_channels.py | base file to generate mimo channels |
| mimo_channels_data_generator2.py | this file generates input data for the model training |
| mimo8x32/plot_* | plot graphs shown in paper |
channel_rosslyn60Ghz - 8x8 channels in Rosslyn.
channels_rosslyn_60Ghz_Nr8Nt32_mobile_s004 - 8x32 mobile channels in Rosslyn.
channels_beijing_60Ghz_Nr8Nt32_mobile_s007 - 8x32 mobile channels in Beijing.
See more details in Raymobtime dataset.