laumerf
commented
1 year ago Did you use the configuration specified here: source/configs/generative_model/dhb.yml?
I see mainly one issue: The predicted frequency is half the actual frequency. I think I experienced this behaviour a couple of times and therefore introduced the regularizer described at equation 7. This helped in our case to prevent it. I honestly don’t know why it still happened in your case.
In your case you can try two things:
- Just continue training and hope that it manages to recover the correct frequency.
- Start training the MVA again from scratch and train it until you see a similar frequency/phase pattern as in figure 6. I think it took my around 24 to 48 hours of GPU training with the parameters specified in the current configuration.
I've currently trained every model on a Windows 10 computer with two NVIDIA Geforce GTX 1080 Ti GPUs using the instructions in the readme.md file. The final EF prediction by cycleGAN seemed to be under fitting, which may have been the reason I continued with the subsequent steps when the MVA had not yet demonstrated a nice correlation (after 104 epochs).
After carefully searching in the article and not finding related information, I would like to ask the following questions:
How many epochs did you train the MVA model with in order to show the association as you did in Fig. 6 of the article?
Could you like to show some details about the implementation setup and working environment? Thank you!