Question/Issue:
Is there a way to reduce the feature input from the spectrogram or increase the allowed model training time?
Project ID:
Large dataset problem currently using spectral features: 120514
Reduced dataset using spectrogram: 120926
Context/Use case:
Hello, my current project seeks to classify heart rate in a regression model based on 4 sensor inputs. When testing features between spectral features and spectrogram, I found that the spectrogram worked almost perfectly at separating each HR into a distinct bin. The spectral features struggle to separate features and end with a high loss outcome, however, due to the low feature input to the model, it is possible to attempt a range of model architectures. The spectrogram on the other hand ends with over 100k features being input into the model. I even attempted a dataset with fewer samples with the same result. Due to the very high feature set, it makes it impossible to train a model more complex than a single layer (<20 neuron) neural network.
Hi @jsindorf,
I’ll send you a DM to discuss your use case. We can increase compute time for some specific projects.
Aurelien