app.py

import gradio as gr
import json, urllib
from model import GPT, GPTConfig
from utils import sample
import torch
import pickle

device = torch.device('cpu')


# Create the model
vocab_size=147
block_size=128
mconf = GPTConfig(vocab_size, block_size,
                  n_layer=6, n_head=8, n_embd=256)
model = GPT(mconf)

# Load checkpoint
model.load_state_dict(torch.load('another_epoch_1.75total.ckpt', map_location=device))

# Vocab
stoi = pickle.load(open('stoi.pkl', 'rb'))
itos = pickle.load(open('itos.pkl', 'rb'))

# Post-process generation
# Completion
def completion_to_song(c):
  lines = c.split('\n')
  kept_lines = []
  notes = False
  for l in lines:

    # Record if we've hit music
    if '|' in l:
      notes = True

    # Stop if we then go back to the start of another song
    if 'T' in l and notes:
      break
    if 'T' in l and notes:
      break

    # Stop on an empty line
    if len(l.strip()) < 2 and notes:
      break
    
    # Otherwise keep the line
    kept_lines.append(l)

  return '\n'.join(kept_lines)

# Generate function
def generate_song(randomize, title, nu, ks, key):
  # Start sequence
  context = b"""T:"""

  if not randomize:
    context += bytes(title+'\n', 'utf-8')
    context += bytes('M:'+ks+'\n', 'utf-8')
    context += bytes('K:'+key+'\n', 'utf-8')
    context += bytes('L:'+nu+'\n', 'utf-8')


  # Model inputs
  x = torch.tensor([stoi[s] for s in context], dtype=torch.long)[None,...].to(device)

  # Completion
  y = sample(model, x, 400, temperature=1.0, sample=True, top_k=10)[0]
  completion = ''.join([chr(itos[int(i)]) for i in y])

  # Return the first song
  song = completion_to_song(completion)

  html_song = song.replace('\n', '<br>')
  url_song = urllib.parse.quote(song, safe='~@#$&()*!+=:;,?/\'')

  html_text = '<p><a href="https://editor.drawthedots.com?t='+url_song+'" target="_blank"><b>EDIT LINK - click to open abcjs editor (allows download and playback)</b></a></p>'+"<p>"+html_song+'</p>'
  return html_text

# Gradio demo
demo = gr.Blocks()

with demo:
    gr.Markdown("Quick demo for [WhistleGen v2](https://wandb.ai/johnowhitaker/whistlegen_v2/reports/WhistleGen-v2--VmlldzoyMTAwNjAz) which lets you generate folk music using a transformer model. I can't get the javascript needed for rendering and playback working with gradio, so this shows the raw ABC notation from the model and a link to view it properly in an external editor.")
    with gr.Row():
        title = gr.Text(label='Title', value='The March of AI')
        with gr.Column():
          nu = gr.Text(label='Note unit', value='1/8')
    with gr.Row():
        key_signature = gr.Dropdown(['3/4', '4/4', '6/8', 'Random'], value='4/4', label='Time Signature')
        with gr.Column():
          key = gr.Text(label='Key', value='D')
    with gr.Row():
      randomize = gr.Checkbox(label='Randomize (ignores settings above)', value=True)
    with gr.Row():
        out = gr.HTML(label="Output", value='Output should appear here (takes ~30s)')
    btn = gr.Button("Run")
    btn.click(fn=generate_song, inputs=[randomize, title, nu, key_signature, key], outputs=out)
    
    with gr.Row():
        gr.Markdown("![visitor badge](https://visitor-badge.glitch.me/badge?page_id=whistlegen_v2_space)")
        gr.Markdown("This is currently using an early model. See the [report](https://wandb.ai/johnowhitaker/whistlegen_v2/reports/WhistleGen-v2--VmlldzoyMTAwNjAz) for training info and updates.")

demo.launch(enable_queue=True)