Summary
To properly evaluate effects of changes on performance of the model we want to setup a proper ablation study.
Current ideas what to evaluate are:
- different embeddings
- angle represetations for IMU data (e.g. quaternion vs euler-angles)
- different vision encoders (e.g. ResNet vs DINO)
- different hyperparameters
- context length of sequences
- number of prediction steps (timeframe) into the future
- Robot type (on/off)
hidden_dim, num_decoder_layers, *encoder_layers
- datasets
- wolfgang / wolfgang + nao
- with / without simulation
Summary
To properly evaluate effects of changes on performance of the model we want to setup a proper ablation study.
Current ideas what to evaluate are:
hidden_dim,num_decoder_layers,*encoder_layers