ngs_test/checkpoints

• Directory description

Directory description

The naming format of each folder under this directory is Hyperparameter1_ Hyperparameter2_ Hyperparameter3, where Hyperparameter1, Hyperparameter2 and Hyperparameter3 represent three hyperparameters respectively, namely λ、 bottleneck_size and recon_points。

In each folder, the optimal model trained under the super parameter combination is stored (refer to the source code for other super parameters).

λ : λ is the hyperparameter to trade-off the rate R and LossCD (S1, S2), and is set as 10, 100, 500, 1000, 2000, 4000 to obtain different rate-distortion (R-D) points and train each model on the entire train dataset. The loss function is described as: L = λ · LossCD (S1, S2) + R

bottleneck_size: The bottleneck size of the encoder output is 256.

recon_points: The number of points in each reconstructed point cloud is 2048 or 2400.