Homepage of OPPO 6G Data Generation Challenge

• H1_32T4R.mat
• H2_32T4R.mat
• Please put the original data in data folder.

• Complex number is special : a+bj has a same square similarity with -a-bj, b-aj and -b+aj
• With the strategy above, you can quadruple the amount of training data compared to the raw.
• We randomly scale the data with a factor between 0.8~1.2, random gaussian noise with mean equals to 0 and std equals to 1e-4 are adopted.

• Auto encoder with reconstruction loss.
• ResNet18 as an Encoder.
• 3D Conv as a Decoder.
• Position Attention Module and Channel Attention Module are important.
• Normalization such as BatchNorm2d after Decoder is important.
• Latent Quantization.

We provide several pretrained models in the folder of saved_models.

• Sim : similarity score tested on the raw data.
• Multi : multi score tested on the raw data.
• Score : tested on the local raw data.
• Feel free to use the pretrained weights or training from scratch.

• Modify the data_type in train.py, maybe you have to choose a suitable GPU id.
• Online validation, only save the models with best scores so far.
• Hints : smaller batch size may result in higher similarity score and higher multi score.
• Epochs : we perform no ablation study on this parameter, you can just let it run.
• Benchmark : data1: local score approx 0.82~0.83
• Benchmark : data2: local score approx 0.76~0.77

• We use adaboost weights to ensemble several models for acquiring performance gain.
• Without model ensembles, you can still achieve an online score up to 0.72 easily.

• You can just use the single model without ensembles which is much easier.
• Without deep ensembles, it is still trivial to achieve a score up to 0.72

• Deep Residual Learning for Image Recognition (CVPR, 2016)
• Dual Attention Network for Scene Segmentation (CVPR, 2019)
• Deep Learning-based Implicit CSI Feedback in Massive MIMO (IEEE Transactions on Communications, 2021)