ADLR Project

• run build.sh to install requirements

• Results of both models with different DoFs are stored in src/figures/evaluation/results/

Hyper-parameters to tune

cVAE

• learning rate (<= 0.01)
• batch_size
• amount of hidden layers per encoder/decoder (>= 1)
• amount of neurons per hidden layer (>= 100)
• variational beta
• (loss weightings)

INN

• learning rate (<= 0.001)
• batch_size
• amount of coupling layers (>= 2)
• amount of hidden layers per subnetwork (>= 1)
• amount of neurons per hidden layer (>= 100)
• (loss weightings)

TODOs:

• implement cVAE model
• implement planar robot simulation with 2 and 3 DoF
• implement paper's robot simulation
• implement rejection sampling
• implement random search for hyperparameter optimization (https://docs.ray.io/en/master/tune/)
• implement basic INN model
• implement backward training of INN model
• debug MMD loss