Pre-trained weights for Music Transformer

[Rahul]

Hi, I am working on my masters thesis about generative music. I am experimenting with different architectures and encoding representations made by magenta and various other research teams. The Music Transformer models are really nice and I would like to experiment with it further. I am searching for all the available pre-trained music transformer models released by magenta, and so far I've found these:

• Unconditional Music Transformer with self attention (architecture is a decoder-only transformer) - official download link here
• Melody Conditional Music Transformer with self attention (architecture is a encoder-decoder transformer with both encoder and decoder using self attention) - official download link here

Can you please share info about any other pre-trained music transformer models available and links to download their weights?

 I would especially be very delighted if there was a pre-trained model for the music transformer using relative attention, but there isn't any available :( 

Can you also please update about when the music transformer using relative attention will be released?

[serkan...@inesctec.pt]

By "relative attention", do you mean "cross-attention" or "attention with relative position encodings"? I didn't run any implementations but based on the Music Transformer paper, it seems like it does have the latter.

[Leon Henne]

Hello Rahul, I am a student as well and currently in an Internship dealing with generative music. My goal is to implement a process to generate new music based on Bach's Goldberg variations. Therefore I would like to ask you, how I could integrate your pretrained_model in some repository. At the moment I'm using the following repository, because it is based on tensorflow 2. 
https://github.com/jason9693/MusicTransformer-tensorflow2.0

My problem is that this repository needs a checkpoint, a ckpt.index an a config.json file next to your provided model file, as shown in the image below.

I would be delighted to receive any information about repositories or ways in generell, to generate music in a good quality based on Wave/Midi files.