app.py

#================================================================
# https://huggingface.co/spaces/asigalov61/MIDI-Identification
#================================================================

import os
import hashlib

import time
import datetime
from pytz import timezone

import copy
from collections import Counter
import random
import statistics
import re

import gradio as gr

from huggingface_hub import InferenceClient

from datasets import load_dataset

import TMIDIX

#==========================================================================================================

HF_TOKEN = os.getenv('HF_TOKEN')

#==========================================================================================================

def format_table_data(data_string):
    
    # Split the string into rows based on newlines
    rows = data_string.strip().split("\n")

    # Initialize a list to store the formatted data
    formatted_data = []

    for row in rows:
        # Split each row into columns based on the separator '|' and strip extra spaces
        columns = row.split("|")
        formatted_row = [cell.strip() for cell in columns]
        
        # Remove cells with only "-" symbols
        formatted_row = [cell for cell in formatted_row if not all(char == '-' for char in cell)]

        # Handle uneven rows by ensuring each row has the same number of columns
        max_columns = max(len(columns) for columns in formatted_data) if formatted_data else len(columns)
        while len(formatted_row) < max_columns:
            formatted_row.append("")  # Add empty strings to fill the row
        
        formatted_data.append(formatted_row)
    
    # Handle case where new rows have more columns than previous rows
    max_columns = max(len(row) for row in formatted_data)
    for row in formatted_data:
        while len(row) < max_columns:
            row.append("")  # Add empty strings to fill the row

    return formatted_data

#==========================================================================================================


MODELS = {'Mistral Nemo Instruct 2407': 'mistralai/Mistral-Nemo-Instruct-2407'
         }

#==========================================================================================================

def ID_MIDI(input_midi, input_model):
    
    print('*' * 70)
    print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
    start_time = time.time()
    
    print('=' * 70)
    print('Loading MIDI...')

    fn = os.path.basename(input_midi)
    fn1 = fn.split('.')[0]

    fdata = open(input_midi, 'rb').read()

    input_midi_md5hash = hashlib.md5(fdata).hexdigest()
    
    print('=' * 70)
    print('Requested settings:')
    print('=' * 70)
    print('Input MIDI file name:', fn)
    print('Input MIDI md5 hash', input_midi_md5hash)
    print('Input model:', input_model)
    print('=' * 70)
    print('Processing MIDI...Please wait...')
    
    #=======================================================
    # START PROCESSING

    new_midi_data = TMIDIX.score2midi(TMIDIX.midi2score(fdata))
    
    new_midi_md5hash = hashlib.md5(new_midi_data).hexdigest()

    print('New md5 hash:', new_midi_md5hash)
    
    print('Done!')
    print('=' * 70)
    print('Processing...Please wait...')

    output_str = 'None'
    output_midi_records_count = 0
    output_midi_src_dataset= 'Unknown'
    output_midi_path_str = 'None'

    raw_score = TMIDIX.midi2single_track_ms_score(fdata)
    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
    escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes, sort_drums_last=True)

    output_midi_src_dataset = 'unknown'
    output_midi_path_str = 'none'

    if new_midi_md5hash in monster_midi_titles['md5_hashes_titles_dict']:
        title = random.choice(monster_midi_titles['md5_hashes_titles_dict'][new_midi_md5hash]).split(' --- ')

        song = title[0]
        artist = title[1]

        song_description = TMIDIX.escore_notes_to_text_description(escore_notes, song_name=song, artist_name=artist)

    else:
        song_description = TMIDIX.escore_notes_to_text_description(escore_notes)
    
    if new_midi_md5hash in midid_md5_hashes:

        midid_entry_idx = midid_md5_hashes.index(new_midi_md5hash)

        MIDID_record = midid_dataset[midid_entry_idx]['midid']

        output_midi_records_count = len(MIDID_record)

        output_entry = random.choice(MIDID_record)

        output_midi_src_dataset = output_entry[0]

        output_midi_path_str = TMIDIX.clean_string(output_entry[1], regex=r'[^a-zA-Z0-9.() \n]')

    client = InferenceClient(api_key=HF_TOKEN)

    prompt = "Please create a summary table for a MIDI file based on the following keywords strings, best possible description and best possible summary fields. Please respond with the table only. Do not say anything else. Thank you."
    
    data = 'Source MIDI dataset: ' + output_midi_src_dataset + '\n\n'
    data += 'MIDI keywords strings:' + '\n'
    data += output_midi_path_str + '\n\n'
    data += 'Music description:' + '\n'
    data += song_description

    
    messages = [
        {
            "role": "user",
            "content": prompt + "\n\n" + data
        }
    ]

    completion = client.chat.completions.create(
        #model="Qwen/Qwen2.5-72B-Instruct",
        model=MODELS[input_model],
        messages=messages, 
        max_tokens=500
    )

    output_str = completion.choices[0].message['content']

    output_table_data = format_table_data(output_str)
    
    print('Done!')
    print('=' * 70)
    print('Original MIDI unique records count', output_midi_records_count)
    print('Original MIDI dataset', output_midi_src_dataset)
    print('Original MIDI path string', data)
    print('=' * 70)
    print(output_str)
    print('=')

    #========================================================

    output_midi_md5 = str(new_midi_md5hash)
   
    #========================================================

    print('Req end time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
    print('-' * 70)
    print('Req execution time:', (time.time() - start_time), 'sec')
    print('*' * 70)
    
    #========================================================
    
    return output_midi_md5, output_midi_records_count, output_midi_src_dataset, data, output_table_data
    
#==========================================================================================================

if __name__ == "__main__":

    PDT = timezone('US/Pacific')
    
    print('=' * 70)
    print('App start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
    print('=' * 70)

    print('Loading MIDID database...')
    midid_dataset = load_dataset("asigalov61/MIDID")['train']

    midid_md5_hashes = midid_dataset['midi_hash']

    print('Done!')
    print('=' * 70)

    print('Loading Monster MIDI titles database...')

    monster_midi_titles = TMIDIX.Tegridy_Any_Pickle_File_Reader('Monster_MIDI_Titles_Database_CC_BY_NC_SA.pickle')
    
    print('Done!')
    print('=' * 70)
    
    app = gr.Blocks()
    
    with app:
        
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>MIDI Identification</h1>")
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Identify any MIDI in a comprehensive database of 2.32M+ MIDI records</h1>")
        
        gr.Markdown("This is a demo for tegridy-tools, MIDID and Monster MIDI dataset\n\n"
                    "Please see [tegridy-tools](https://github.com/asigalov61/tegridy-tools), [MIDID](https://huggingface.co/datasets/asigalov61/MIDID) and [Monster MIDI Dataset](https://github.com/asigalov61/Monster-MIDI-Dataset) repos for more information\n\n"
                   )
        
        gr.Markdown("## Upload your MIDI")
        
        input_midi = gr.File(label="Input MIDI", file_types=[".midi", ".mid", ".kar"], type="filepath")
        input_model = gr.Dropdown(['Mistral Nemo Instruct 2407', 'Mistral Nemo Instruct 2407'], 
                                  value='Mistral Nemo Instruct 2407', 
                                  label='Select model'
                                 )

        submit = gr.Button("Identify MIDI", variant="primary")

        gr.Markdown("## MIDI identification results")
        
        output_midi_md5 = gr.Textbox(label="Monster MIDI dataset md5 hash")
        output_midi_records_count = gr.Textbox(label="Original MIDI unique records count")
        output_midi_src_dataset = gr.Textbox(label="Original MIDI dataset pretty name")
        output_midi_path_str = gr.Textbox(label="Original MIDI raw path string")
        output_MIDID_results_table = gr.Dataframe(label="MIDID database results table", wrap=True, col_count=(3, 'dynamic'))
        
        run_event = submit.click(ID_MIDI, [input_midi,
                                           input_model
                                              ],
                                                [output_midi_md5,
                                                 output_midi_records_count,
                                                 output_midi_src_dataset,
                                                 output_midi_path_str,                                                 
                                                 output_MIDID_results_table
                                                ])
        
    app.queue().launch()