modelzoo

Time Series

Many-to-Many

• vanillaLSTM : Encoder LSTM decoded by a Dense Layer. Single Shot RNN that predict the output sequence once.
• seq2seq : A classical sequence2sequence RNN. Encoder LSTM and Decoder LSTM, so it keeps Time dependancy in the decoder.
• seq2seqAttention : Similar to seq2seq with the add of Luong Attention.

3Ds

PointClouds

• pointnet : Network used to perform classification, semantic-segmentation and part-segmentation tasks with Points Clouds
• pointnet_foldingnet : Autoencoder for points clouds data. Encoder is based on pointnet and decoder on foldingnet.
• foldingnet : Foldingnet, autoencoder for point clouds : (B,N,3) -> (B,M,3), with M a square number

CNN

• ASPP : ASPP for Atrous Spatial Pyramid Pooling
• unet : CNN used for semantic segmentation

Other

Custom Losses

• wcce : Weighted Categorical Cross Entropy, used for imbalanced classification datasets. Take a weights list in parameters.
• dice_loss : Dice Loss function, mainly used in semantic segmentation
• chamfer_loss : Distance used to compare a B point clouds of shape [B,N,3] with B point clouds of shape [B,M,3]