Melodies sometimes cannot be extended due to a key error in the backend when translating from the network's encoded representation to the time series representation.
Frist assumptions are that is caused by a missing key-value pair,
As the network's number mappings are constructed from the dataset, and this map is where the error is occuring.
Is the LSTM Generating a number for a mapping which was never seen? this makes zero sense as it shouldn't be able to generate something not found in its training data.
Therefore i've got no real idea of what this may actually be!
Solution?
Quick Fix:
Add a try/catch block to the mapping function, simply retrying a note generation if an exception is raised.
Actual Fix:
Use a better model structure and better Music representation.
Melodies sometimes cannot be extended due to a
key errorin the backend when translating from the network's encoded representation to the time series representation.Frist assumptions are that is caused by a missing key-value pair, As the network's number mappings are constructed from the dataset, and this map is where the error is occuring. Is the LSTM Generating a number for a mapping which was never seen? this makes zero sense as it shouldn't be able to generate something not found in its training data.
Therefore i've got no real idea of what this may actually be!
Solution? Quick Fix: Add a try/catch block to the mapping function, simply retrying a note generation if an exception is raised.
Actual Fix: Use a better model structure and better Music representation.