Fig. 2

The effect of systematic adjustments of single architectural parameters on the Full Reconstruction rate. Shown is the Full Reconstruction on the test split achieved by models trained on the full set A and the 50 k subset B when adjusting hidden & latent size, only latent size, the number of layers and when adding attention. Each model was trained using three different seeds. The metric is presented as the mean of these seeds with error bars indicating the highest and lowest value. The performance of the base model (hidden and latent size of 64, one layer, no attention) is always shown as a reference to the modified architectures. GRU = Gated Recurrent Unit, LSTM = Long Short-Term Memory