This is a Debugged version of Implementation of MidiNet by pytorch.

MidiNet paper : https://arxiv.org/abs/1703.10847

MidiNet code : https://github.com/RichardYang40148/MidiNet

dataset is from theorytab : https://www.hooktheory.com/theorytab

You can find crawler here : https://github.com/wayne391/Symbolic-Musical-Datasets

---

Prepare the data

file to run	function
get_data.py	get melody and chord matrix from xml
get_train_and_test_data.py	seperate the melody data into training set and testing set (chord preparation not included)

Or get the processed data from the author.

Note: The processed data from the author combines the training and testing data together. You may need to re-separate them.
Another Note: Whether get_data.py and get_train_and_test_data.py are bug-free is unknown (download the processed data from the author is recommended).

---

After you have the data (placed in data folder),

1. Make sure you have toolkits in the requirement.txt

2. Run main.py ,
   is_train = 1 for training,
   is_draw = 1 for drawing loss,
   is_sample = 1 for generating music after finishing training.

3. If you would like to turn the output into real midi for listening
   Run demo.py

---

file	purposes
requirement.txt	toolkits used in the whole work
main.py	training setting, drawing setting, generation setting.
ops.py	some functions used in model
model.py	Generator and Discriminator. (Based on model 3 in the MidiNet paper)
demo.py	transform matrix into midi. (input : melody and chord matrix, output : midi)