First of all, thank you very much to the OpenDPD team for developing such an excellent open-source platform! OpenDPD provides the neural network DPD research community with valuable datasets, end-to-end learning frameworks, and a wide range of backbone models for both PA and DPD, significantly lowering the barrier to entry for research.
While reading the OpenDPD benchmark report, I noticed that the APA_200MHz dataset uses a GRU PA model, achieving an awesome NMSE performance of -43.52 dB. However, when I trained a GRU PA model with hidden size 23 on the APA_200MHz dataset, I could only achieve an NMSE of approximately -33 dB. I am wondering how I can achieve a PA model with performance under -40 dB, since the PA model's performance is especially critical in the end-to-end learning framework.
I trained the PA using the following command:
python main.py --step train_pa --dataset_name APA_200MHz --PA_backbone gru --PA_hidden_size 23 --n_epochs 300 --opt_type adamw --frame_length 100/200/300
The corresponding training logs are attached below:
PA_S_0_M_GRU_H_23_F_100_P_1911.csv
PA_S_0_M_GRU_H_23_F_200_P_1911.csv
PA_S_0_M_GRU_H_23_F_300_P_1911.csv
Could this performance gap be related to my specific configuration or training setup? Any advice would be greatly appreciated.
First of all, thank you very much to the OpenDPD team for developing such an excellent open-source platform! OpenDPD provides the neural network DPD research community with valuable datasets, end-to-end learning frameworks, and a wide range of backbone models for both PA and DPD, significantly lowering the barrier to entry for research.
While reading the OpenDPD benchmark report, I noticed that the APA_200MHz dataset uses a GRU PA model, achieving an awesome NMSE performance of -43.52 dB. However, when I trained a GRU PA model with hidden size 23 on the APA_200MHz dataset, I could only achieve an NMSE of approximately -33 dB. I am wondering how I can achieve a PA model with performance under -40 dB, since the PA model's performance is especially critical in the end-to-end learning framework.
I trained the PA using the following command:
The corresponding training logs are attached below:
PA_S_0_M_GRU_H_23_F_100_P_1911.csv
PA_S_0_M_GRU_H_23_F_200_P_1911.csv
PA_S_0_M_GRU_H_23_F_300_P_1911.csv
Could this performance gap be related to my specific configuration or training setup? Any advice would be greatly appreciated.