view demo here

• collect some midi files and put them in the midi folder
• run prepare.py, we will count the frequency at which notes are pressed

• the model uses deep one-dimensional residual dense convolution network to extract time series characteristics

• the model predicts the state of 88 keys in 1/64 note time at a time
• so the generation process is shown in the following figure

• run train.py
• in the case of small amount of data, it is recommended that the sequence length be set to 128

• run predict.py
• give the model a short melody, which will produce a melody of similar style

• tensorflow-gpu 1.12.0
• Keras 2.2.4