Auto Regressive models -> AR-CNN

Amazon DeepComposer

Adarsha Regmi
1 min readAug 1, 2021

what it is ??

The models that study the system which are updated with time are auto regressive models. The future task are performed based upon the historical events and their outcome.

Amazon Deep Composer

The AR-CNN is common thing that comes in topic when using deep composer. It trains model over time to result into better inference. The model generates the output with small change so that the outputs modules towards the input music. The chords notes can be added removed or adjusted throughout.

time in x-axis and pitch in y-axis >>>. images of musical score and piano roll

How AR-CNN model works ?

The iteration begins to compare the input track with the training datasets. After the process the notes are either added or removed. The process also involves detecting the variance between the modified roll with the original roll .


The AR-CNN model adds a new note.


move input track to a state where it sounds like the music is constrain on. The model creates another edit event to minimize the loss function or set another way continue to make input track more like training set.