s-omranpour / midi-transformer

Another implementation of the paper "Compound Word Transformer: Learning to Compose Full-Song Music over Dynamic Directed Hypergraphs" in pytorch.

• 📋 Previously worked on Music Information Retrieval, Natural Language Processing, Speech Recognition
• 🔭 Currently working on Graph Learning and Network Science with a touch of Theory
• 🎸 Playing Guitar and MIDI Keyboard

midi-transformer's Issues

About the attention type in the example config file

if x_mask is None:
            if self.self_attention_type.startswith('causal'):
                x_mask = TriangularCausalMask(x.size(1), device=x.device)
            else:
                x_mask = FullMask(x.size(1), device=x.device)

I thought TriangularCausalMask should be used when training a music generation model, but the self attention type in the example config is "linear". Should I add 'causal' before it?

KeyError: None

Hello, thank you for your work. When I train my model for inference, if the parameter max_len in model.generate() is set to greater than 30, the following error will be reported. How to deal with this problem?

Traceback (most recent call last):
File "MIDI-Transformer-main/examples/infer.py", line 86, in
seq = MusicRepr.from_cp(gen)
File "/home/anaconda3/lib/python3.7/site-packages/deepnote/repr.py", line 176, in from_cp
res += utils.sort_bar_beats(bar)
File "/home/anaconda3/lib/python3.7/site-packages/deepnote/utils.py", line 25, in sort_bar_beats
poses[prev_pos] += [e]
KeyError: None