app.py

import subprocess
import sys

subprocess.check_call([sys.executable,"-m","pip","install",'causal-conv1d'])
subprocess.check_call([sys.executable, "-m", "pip", "install", 'miditok','mamba-ssm','gradio'])
subprocess.check_call(["apt-get", "install", "timidity", "-y"])

# !pip install pretty_midi midi2audio
# !pip install miditok
# !apt-get install fluidsynth
# !apt install timidity -y
# !pip install causal-conv1d>=1.1.0
# !pip install mamba-ssm
# !pip install gradio



# !export LC_ALL="en_US.UTF-8"
# !export LD_LIBRARY_PATH="/usr/lib64-nvidia"
# !export LIBRARY_PATH="/usr/local/cuda/lib64/stubs"

# subprocess.check_call(['export', 'LC_ALL="en_US.UTF-8"'])
# subprocess.check_call(['export', 'LD_LIBRARY_PATH="/usr/lib64-nvidia"'])
# subprocess.check_call(['export', 'LIBRARY_PATH="/usr/local/cuda/lib64/stubs"'])
import os

os.environ['LC_ALL'] = "en_US.UTF-8"
os.environ['LD_LIBRARY_PATH'] = "/usr/lib64-nvidia"
os.environ['LIBRARY_PATH'] = "/usr/local/cuda/lib64/stubs"



import gradio as gr
import torch
from mamba_ssm import Mamba
from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
from mamba_ssm.models.config_mamba import MambaConfig
import numpy as np

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
if torch.cuda.is_available():
    subprocess.check_call(['ldconfig', '/usr/lib64-nvidia'])
    # !ldconfig /usr/lib64-nvidia

# !wget "https://huggingface.co/krystv/MIDI_Mamba-159M/resolve/main/MIDI_Mamba-159M_1536VS.pt"
# !wget "https://huggingface.co/krystv/MIDI_Mamba-159M/resolve/main/tokenizer_1536mix_BPE.json"
if os.path.isfile("MIDI_Mamba-159M_1536VS.pt") == False:
    subprocess.check_call(['wget', 'https://huggingface.co/krystv/MIDI_Mamba-159M/resolve/main/MIDI_Mamba-159M_1536VS.pt'])

if os.path.isfile("tokenizer_1536mix_BPE.json") == False:
    subprocess.check_call(['wget', 'https://huggingface.co/krystv/MIDI_Mamba-159M/resolve/main/tokenizer_1536mix_BPE.json'])



mc = MambaConfig()
mc.d_model = 768
mc.n_layer = 42
mc.vocab_size = 1536

from miditok import MIDILike,REMI,TokenizerConfig
from pathlib import Path
import torch

tokenizer = REMI(params='tokenizer_1536mix_BPE.json')



mf = MambaLMHeadModel(config=mc,device=device)
mf.load_state_dict(torch.load("/content/MIDI_Mamba-159M_1536VS.pt",map_location=device))



twitter_follow_link = "https://twitter.com/iamhemantindia"
instagram_follow_link = "https://instagram.com/iamhemantindia"

custom_html = f"""
<div style='text-align: center;'>
    <a href="{twitter_follow_link}" target="_blank" style="margin-right: 5px;">
        <img src="https://img.icons8.com/fluent/24/000000/twitter.png" alt="Follow on Twitter"/>
    </a>
    <a href="{instagram_follow_link}" target="_blank">
        <img src="https://img.icons8.com/fluent/24/000000/instagram-new.png" alt="Follow on Instagram"/>
    </a>
</div>
"""


@spaces.GPU(duration=120)
def generate(number,top_k_selector,top_p_selector, temperature_selector):
    input_ids = torch.tensor([[1,]]).to(device)
    out = mf.generate(
    input_ids=input_ids,
    max_length=int(number),
    temperature=temperature_selector,
    top_p=top_p_selector,
    top_k=top_k_selector,

    eos_token_id=2,)
    m = tokenizer.decode(np.array(out[0].to('cpu')))
    np.array(out.to('cpu')).shape
    m.dump_midi('output.mid')
    # !timidity output.mid -Ow -o - | ffmpeg -y -f wav -i - output.mp3
    timidity_cmd = ['timidity', 'output.mid', '-Ow', '-o', 'output.wav']
    subprocess.check_call(timidity_cmd)

    # Then convert the WAV to MP3 using ffmpeg
    ffmpeg_cmd = ['ffmpeg', '-y', '-f', 'wav', '-i', 'output.wav', 'output.mp3']
    subprocess.check_call(ffmpeg_cmd)

    return "output.mp3"


# text_box = gr.Textbox(label="Enter Text")


def generate_and_save(number,top_k_selector,top_p_selector, temperature_selector,generate_button,custom_html_wid):
    output_audio = generate(number,top_k_selector,top_p_selector, temperature_selector)
    return gr.Audio(output_audio,autoplay=True),gr.File(label="Download MIDI",value="output.mid"),generate_button






# iface = gr.Interface(fn=generate_and_save,
#  inputs=[number_selector,top_k_selector,top_p_selector, temperature_selector,generate_button,custom_html_wid],
#   outputs=[output_box,download_midi_button],
#    title="MIDI Mamba-159M",submit_btn=False,
#    clear_btn=False,
#    description="MIDI Mamba is a Mamba based model trained on MIDI data collected from open internet to train music model.",
#    allow_flagging=False,)

with gr.Blocks() as b1:
    gr.Markdown("<h1 style='text-align: center;'>MIDI Mamba-159M <h1/> ")
    gr.Markdown("<h3 style='text-align: center;'>MIDI Mamba is a Mamba based model trained on MIDI data collected from open internet to train music model. <br> by Hemant Kumar<h3/>")
    with gr.Row():
        with gr.Column():
            number_selector = gr.Number(label="Select Length of output",value=512)
            top_p_selector = gr.Slider(label="Select Top P", minimum=0, maximum=1.0, step=0.05, value=0.9)
            temperature_selector = gr.Slider(label="Select Temperature", minimum=0, maximum=1.0, step=0.1, value=0.9)
            top_k_selector = gr.Slider(label="Select Top K", minimum=1, maximum=1536, step=1, value=30)
            generate_button = gr.Button(value="Generate",variant="primary")
            custom_html_wid = gr.HTML(custom_html)
        with gr.Column():
            output_box = gr.Audio("output.mp3",autoplay=True,)
            download_midi_button = gr.File(label="Download MIDI")
            generate_button.click(generate_and_save,inputs=[number_selector,top_k_selector,top_p_selector, temperature_selector,generate_button,custom_html_wid],outputs=[output_box,download_midi_button,generate_button])




b1.launch(share=True)