ZhuHouYi
commented
3 months ago I look forward to your reply. I believe the issue lies with my training strategy, which has led to my inability to reproduce the perfect results presented in your paper. I would greatly appreciate your guidance.😀
Dear Author,
I have reimplemented your VAE model using the Disaggregator class from nilmtk_contrib. The experimental hyperparameters are as follows: epoch=10,batch size = 64, window size = 512, learning rate = 3e-4,optimizer=Adam,validation rate = 0.15. The experimental dataset used four months of UK-DALE data for training and two months for testing. The training loss is defined according to your code, including reconstruction loss and KL divergence loss. The checkpoint strategy during training was also adopted from your code, using ModelCheckpoint(monitor="val_mean_absolute_error", mode="min", save_best_only=True) and CustomStopper(monitor='val_loss', mode="auto"). I observed in the TensorFlow command line logs that the losses were decreasing normally (the KL loss approached 0 over time, while the Recon_loss decreased very little).
However, the performance metrics after training were very poor (I trained five devices: Fridge, Kettle, Microwave, Dishwasher, and Washer Machine). Do you have any insights or suggestions for improvement?
Thank you.
The model architecture part was completely copied from your code, with only some necessary modifications to the latent dimensions to adapt to the input data with a window size of 512. Additionally, I used Adam with a learning rate of 3e-4 in the model.compile() section. I personally believe this should not have a significant impact on the training process.