change download mechanism

app.py

@@ -1,38 +1,25 @@
 import spaces
 import gradio as gr
-import torch
 import librosa
-import numpy as np
-import subprocess
-import sys
+import soundfile as sf
 import os
 import glob
 from pathlib import Path
 from huggingface_hub import snapshot_download
-import tempfile
-import uuid
+import shutil
+import sys
+import subprocess
+import numpy as np
 
-token = os.getenv("HF_TOKEN")
+# Install madmom from GitHub
+def install_madmom():
+    subprocess.check_call([
+        sys.executable, "-m", "pip", "install", 
+        "git+https://github.com/CPJKU/madmom", "--no-cache-dir"
+    ])
+    print("madmom installed from GitHub")
 
-# Install dependencies
-def install_dependencies():
-    dependencies = [
-        ("madmom", "git+https://github.com/CPJKU/madmom"),
-        ("soundfile", "soundfile")
-    ]
-    
-    for name, package in dependencies:
-        try:
-            __import__(name.replace('-', '_'))
-            print(f"{name} already installed")
-        except ImportError:
-            print(f"Installing {name}...")
-            subprocess.check_call([
-                sys.executable, "-m", "pip", "install", 
-                package, "--no-cache-dir"
-            ])
-
-install_dependencies()
+install_madmom()
 
 # Add current directory to Python path for ml_models
 sys.path.insert(0, '.')
@@ -49,7 +36,6 @@ def download_data_from_hub():
     folders_to_check = ["covers80", "ml_models"]
     downloaded_folders = {}
     
-    # Check LFS file
     lfs_file = base_dir / "1005_e_4"
     print(f"Checking LFS file: {lfs_file}")
     if lfs_file.exists():
@@ -60,27 +46,13 @@ def download_data_from_hub():
         print("LFS file not found")
         downloaded_folders["1005_e_4"] = None
     
-    # Check existing folders
     print("=== CHECKING EXISTING FOLDERS ===")
-    for folder in folders_to_check:
-        folder_path = base_dir / folder
-        print(f"Checking {folder} at {folder_path}")
-        if folder_path.exists():
-            if any(folder_path.iterdir()):
-                print(f"  {folder} exists and has content")
-            else:
-                print(f"  {folder} exists but is empty")
-        else:
-            print(f"  {folder} does not exist")
-    
     all_folders_exist = all((base_dir / folder).exists() and any((base_dir / folder).iterdir()) 
                            for folder in folders_to_check)
     print(f"All folders exist: {all_folders_exist}")
     
     if not all_folders_exist:
         print("=== STARTING DOWNLOAD ===")
-        
-        # Download to a temporary directory first
         temp_dir = base_dir / "temp_download"
         print(f"Creating temp directory: {temp_dir}")
         temp_dir.mkdir(exist_ok=True)
@@ -91,44 +63,29 @@ def download_data_from_hub():
             repo_type="dataset",
             local_dir=str(temp_dir),
             local_dir_use_symlinks=False,
-            token=token,
+            token=os.getenv("HF_TOKEN"),
             ignore_patterns=["*.md", "*.txt", ".gitattributes", "README.md"]
         )
         
         print(f"Download completed to: {downloaded_path}")
         
-        # Check what was downloaded
         print("=== CHECKING TEMP DOWNLOAD CONTENTS ===")
         print(f"Temp directory contents:")
         for item in temp_dir.iterdir():
             item_type = "DIR" if item.is_dir() else "FILE"
             print(f"  {item.name} ({item_type})")
-            if item.is_dir():
-                file_count = len([f for f in item.rglob("*") if f.is_file()])
-                print(f"    Contains {file_count} files")
         
-        # Move folders from temp to current directory
         print("=== MOVING FOLDERS ===")
         for folder_name in folders_to_check:
             temp_folder_path = temp_dir / folder_name
             target_folder_path = base_dir / folder_name
             
-            print(f"Processing {folder_name}:")
-            print(f"  Source: {temp_folder_path}")
-            print(f"  Target: {target_folder_path}")
-            print(f"  Source exists: {temp_folder_path.exists()}")
-            
             if temp_folder_path.exists():
-                # Remove existing target if it exists
                 if target_folder_path.exists():
-                    print(f"  Removing existing target directory")
                     shutil.rmtree(target_folder_path)
                 
-                # Move folder
-                print(f"  Moving folder...")
                 shutil.move(str(temp_folder_path), str(target_folder_path))
                 
-                # Verify move
                 if target_folder_path.exists():
                     file_count = len([f for f in target_folder_path.rglob("*") if f.is_file()])
                     print(f"  SUCCESS: {folder_name} moved with {file_count:,} files")
@@ -140,7 +97,6 @@ def download_data_from_hub():
                 print(f"  ERROR: {folder_name} not found in temp download")
                 downloaded_folders[folder_name] = None
         
-        # Clean up temp directory
         print("=== CLEANING UP TEMP DIRECTORY ===")
         if temp_dir.exists():
             shutil.rmtree(temp_dir)
@@ -164,39 +120,21 @@ def download_data_from_hub():
     print("=== DOWNLOAD FUNCTION END ===")
     return downloaded_folders
 
-# Download data and check results
 print("Starting Music Plagiarism Detection App...")
 folders = download_data_from_hub()
 
-# Final verification
-print("=== FINAL VERIFICATION ===")
-current_dir = Path(".")
-print(f"Current directory contents after download:")
-for item in current_dir.iterdir():
-    item_type = "DIR" if item.is_dir() else "FILE"
-    print(f"  {item.name} ({item_type})")
-
-# Check ml_models specifically
-ml_models_path = Path("ml_models")
-print(f"ml_models check:")
-print(f"  Exists: {ml_models_path.exists()}")
-if ml_models_path.exists():
-    print(f"  Is directory: {ml_models_path.is_dir()}")
-    print(f"  Contents:")
-    for item in ml_models_path.iterdir():
-        print(f"    {item.name}")
-
-# Import inference
 print("=== IMPORTING INFERENCE ===")
 from inference import inference
 
+# 오디오 세그먼트를 저장할 임시 디렉토리
+TEMP_AUDIO_DIR = Path("./temp_audio_segments")
+TEMP_AUDIO_DIR.mkdir(exist_ok=True)
+
 def find_song_file_by_title(song_title):
     covers80_path = Path("covers80")
-    
     if not covers80_path.exists():
         return None
     
-    # Try exact match patterns
     exact_patterns = [
         f"{song_title}.mp3",
         f"*{song_title}.mp3",
@@ -208,7 +146,6 @@ def find_song_file_by_title(song_title):
         if matches:
             return str(matches[0])
     
-    # Try partial matches
     song_parts = song_title.replace('_', ' ').split()
     for part in song_parts:
         if len(part) > 3:
@@ -218,29 +155,54 @@ def find_song_file_by_title(song_title):
     
     return None
 
-def format_time(seconds):
-    """Convert seconds to MM:SS format"""
-    if seconds is None or seconds < 0:
-        return "0:00"
-    
-    minutes = int(seconds // 60)
-    seconds = int(seconds % 60)
-    return f"{minutes}:{seconds:02d}"
+def crop_audio_segment_librosa(input_path, start_time, end_time, output_path):
+    """
+    librosa와 soundfile을 사용하여 오디오 파일의 특정 구간을 자르는 함수.
+    """
+    try:
+        # 오디오 파일 로드
+        y, sr = librosa.load(input_path, sr=None)
+        
+        # 시작 및 종료 시간(초)을 샘플 인덱스로 변환
+        start_sample = int(start_time * sr)
+        end_sample = int(end_time * sr)
+        
+        # numpy 배열 슬라이싱으로 오디오 세그먼트 추출
+        cropped_audio = y[start_sample:end_sample]
+        
+        # 잘린 오디오를 WAV 파일로 저장
+        sf.write(output_path, cropped_audio, sr)
+        
+        print(f"Successfully cropped audio from {input_path} to {output_path} ({start_time}-{end_time}s) using librosa.")
+        return output_path
+    except Exception as e:
+        print(f"Error cropping audio with librosa: {e}")
+        return None
+
+def clear_temp_segments():
+    """임시 오디오 세그먼트 디렉토리 정리"""
+    if TEMP_AUDIO_DIR.exists():
+        shutil.rmtree(TEMP_AUDIO_DIR)
+    TEMP_AUDIO_DIR.mkdir(exist_ok=True)
+    print("Temporary audio segments cleared.")
 
 @spaces.GPU(duration=300)
-def process_audio_for_matching(audio_file):
+def process_audio_for_matching_and_crop(audio_file):
     if audio_file is None:
-        return """
+        # 10개의 None과 에러 메시지 반환
+        return [None] * 10, """
         <div style='text-align: center; color: #dc2626; padding: 20px; background: #fef2f2; border-radius: 8px;'>
             <h3>No Audio File</h3>
             <p>Please upload an audio file to get started!</p>
         </div>
         """
 
+    clear_temp_segments()
+
     result = inference(audio_file)
     
     if result.get('message') != 'success':
-        return f"""
+        return [None] * 10, f"""
         <div style="text-align: center; padding: 20px; background: #fefce8; border-radius: 8px;">
             <h3 style="color: #a16207;">No Matches Found</h3>
             <p style="color: #a16207;">{result.get('message', 'Unknown error occurred')}</p>
@@ -249,135 +211,97 @@ def process_audio_for_matching(audio_file):
     
     matches = result.get('matches', [])
     if not matches:
-        return """
+        return [None] * 10, """
         <div style="text-align: center; padding: 20px; background: #fefce8; border-radius: 8px;">
             <h3 style="color: #a16207;">No Matches Found</h3>
             <p style="color: #a16207;">No matching vocals found in the dataset.</p>
         </div>
         """
     
-    # Create custom HTML audio players with timestamp functionality
-    audio_players_html = ""
-    
-    # Input audio player - use Gradio's file serving
-    audio_players_html += f"""
-    <div style="margin-bottom: 20px;">
-        <h4 style="color: #111827;">Your Uploaded Audio</h4>
-        <audio id="inputAudio" controls style="width: 100%;">
-            <source src="data:audio/wav;base64,{audio_file}" type="audio/wav">
-            Your browser does not support the audio element.
-        </audio>
-    </div>
-    """
-    
-    # Match audio players - 일단 오디오 플레이어는 제거하고 정보만
-    match_files_info = []
-    for i, match in enumerate(matches[:3]):
-        song_title = match.get('song_title', 'Unknown Song')
-        song_file_path = find_song_file_by_title(song_title)
-        
-        if song_file_path and os.path.exists(song_file_path):
-            match_files_info.append({
-                'title': song_title,
-                'path': song_file_path,
-                'index': i+1
-            })
-        else:
-            match_files_info.append({
-                'title': song_title,
-                'path': None,
-                'index': i+1
-            })
-    
-    # Generate match results with clickable timestamps
-    matches_html = ""
-    for i, match in enumerate(matches[:3]):
-        rank = match.get('rank', 0)
+    # 10개의 오디오 컴포넌트를 위한 배열 초기화
+    audio_outputs = [None] * 10
+    match_html = ""
+
+    # 원본 오디오를 위한 슬롯 할당 (첫 번째는 업로드된 파일)
+    audio_outputs[0] = audio_file
+
+    # 상위 3개 매치 세그먼트 및 상위 4개 원본 오디오 처리
+    for i, match in enumerate(matches[:4]):  # 상위 4개 매치만 처리
         song_title = match.get('song_title', 'Unknown Song')
-        confidence = match.get('confidence', '0%')
-        test_time = match.get('test_time', 0)
-        library_time = match.get('library_time', 0)
+        library_file_path = find_song_file_by_title(song_title)
+
+        if not library_file_path:
+            continue
         
-        # Ranking colors
-        rank_colors = {1: '#dc2626', 2: '#ea580c', 3: '#16a34a'}
-        rank_color = rank_colors.get(rank, '#6b7280')
+        # 원본 라이브러리 오디오 할당 (1-3번 슬롯)
+        if i < 3:
+            audio_outputs[i + 1] = library_file_path
         
-        matches_html += f"""
-        <div style="background: #ffffff; border-radius: 8px; padding: 15px; margin: 10px 0; 
-                    border-left: 4px solid {rank_color}; box-shadow: 0 2px 8px rgba(0,0,0,0.1);">
-            <div style="display: flex; justify-content: space-between; align-items: center;">
-                <div style="flex: 1;">
-                    <h4 style="color: #111827; margin: 0; font-size: 1.1em;">
-                        <span style="background: {rank_color}; color: white; padding: 2px 6px; border-radius: 10px; font-size: 0.8em; margin-right: 8px;">
-                            #{rank}
-                        </span>
-                        {song_title}
-                    </h4>
-                </div>
-                <div style="display: flex; gap: 15px; align-items: center;">
-                    <div style="text-align: center;">
-                        <small style="color: #6b7280;">Your Audio</small>
-                        <div style="color: #dc2626; font-weight: 600;">
-                            {format_time(test_time)}
-                        </div>
-                        <small style="color: #9ca3af;">@{test_time:.1f}s</small>
+        # 상위 3개 매치에 대해서만 세그먼트 생성 (입력 오디오 세그먼트 및 라이브러리 오디오 세그먼트)
+        if i < 3:
+            # 입력 오디오 세그먼트 생성
+            input_start = match.get('test_time', 0)
+            input_end = match.get('test_time2', input_start + 10)
+            input_segment_path = TEMP_AUDIO_DIR / f"input_seg_{i}.wav"
+            cropped_input_path = crop_audio_segment_librosa(audio_file, input_start, input_end, input_segment_path)
+            
+            # 라이브러리 오디오 세그먼트 생성
+            library_start = match.get('library_time', 0)
+            library_end = match.get('library_time2', library_start + 10)
+            library_segment_path = TEMP_AUDIO_DIR / f"library_seg_{i}.wav"
+            cropped_library_path = crop_audio_segment_librosa(library_file_path, library_start, library_end, library_segment_path)
+
+            # 세그먼트 파일 할당 (4-9번 슬롯)
+            if cropped_input_path and (i*2 + 4) < 10:
+                audio_outputs[i*2 + 4] = cropped_input_path
+            if cropped_library_path and (i*2 + 5) < 10:
+                audio_outputs[i*2 + 5] = cropped_library_path
+
+        # HTML 결과 포맷팅 (상위 3개 매치에 대해서만)
+        if i < 3:
+            rank = match.get('rank', 0)
+            confidence = match.get('confidence', '0%')
+            rank_colors = {1: '#dc2626', 2: '#ea580c', 3: '#16a34a'}
+            rank_color = rank_colors.get(rank, '#6b7280')
+
+            match_html += f"""
+            <div style="background: #ffffff; border-radius: 8px; padding: 15px; margin: 10px 0; 
+                        border-left: 4px solid {rank_color}; box-shadow: 0 2px 8px rgba(0,0,0,0.1);">
+                <div style="display: flex; justify-content: space-between; align-items: center;">
+                    <div style="flex: 1;">
+                        <h4 style="color: #111827; margin: 0; font-size: 1.1em;">
+                            <span style="background: {rank_color}; color: white; padding: 2px 6px; border-radius: 10px; font-size: 0.8em; margin-right: 8px;">
+                                #{rank}
+                            </span>
+                            {song_title}
+                        </h4>
                     </div>
-                    <div style="text-align: center;">
-                        <small style="color: #6b7280;">Matched At</small>
-                        <div style="color: #16a34a; font-weight: 600;">
-                            {format_time(library_time)}
+                    <div style="display: flex; gap: 15px; align-items: center;">
+                        <div style="text-align: center;">
+                            <small style="color: #6b7280;">Confidence</small>
+                            <div style="background: #f3f4f6; color: #111827; padding: 4px 10px; border-radius: 12px; font-weight: 600; font-size: 0.9em;">
+                                {confidence}
+                            </div>
                         </div>
-                        <small style="color: #9ca3af;">@{library_time:.1f}s</small>
-                    </div>
-                    <div style="background: #f3f4f6; color: #111827; padding: 4px 10px; border-radius: 12px; font-weight: 600; font-size: 0.9em;">
-                        {confidence}
                     </div>
                 </div>
             </div>
-        </div>
-        """
-    
-    # Complete HTML with audio players and results
-    complete_html = f"""
+            """
+
+    results_html = f"""
     <div style="background: #ffffff; border-radius: 12px; padding: 20px; 
                 box-shadow: 0 4px 15px rgba(0,0,0,0.08); border: 1px solid #e5e7eb;">
-        
-        <div style="text-align: center; margin-bottom: 30px;">
+        <div style="text-align: center; margin-bottom: 20px;">
             <h3 style="color: #111827; margin: 0;">Vocal Matching Results</h3>
             <p style="color: #6b7280; margin: 5px 0;">Found {len(matches)} similar vocals in Covers80 dataset</p>
-            <p style="color: #2563eb; margin: 5px 0; font-size: 0.9em;">🎵 Timestamps show exact matching points in both audio files</p>
-        </div>
-        
-        <!-- Audio Players -->
-        <div style="background: #f8fafc; padding: 20px; border-radius: 8px; margin-bottom: 20px;">
-            <h3 style="color: #111827; margin-bottom: 15px;">Audio Information</h3>
-            <p style="color: #6b7280; font-size: 0.9em; margin-bottom: 15px;">
-                Due to Hugging Face file access limitations, audio players are not available in this demo. 
-                However, you can see the exact timestamps where matches were found.
-            </p>
-            {audio_players_html}
-        </div>
-        
-        <!-- Match Results -->
-        <div>
-            <h3 style="color: #111827; margin-bottom: 15px;">Detailed Results</h3>
-            {matches_html}
         </div>
-        
-        <script>
-        // JavaScript removed due to compatibility issues
-        console.log('Music Plagiarism Detection - Timestamp display only version');
-        </script>
-        
-        <style>
-        /* Removed interactive styles */
-        </style>
+        {match_html}
     </div>
     """
     
-    return complete_html
+    # 총 10개의 오디오 컴포넌트와 HTML 결과를 반환
+    return audio_outputs + [results_html]
 
-# CSS styles
 custom_css = """
 .gradio-container { 
     background: #f9fafb !important; 
@@ -395,7 +319,6 @@ custom_css = """
 }
 """
 
-# Gradio interface
 with gr.Blocks(css=custom_css, theme=gr.themes.Soft(), title="Music Plagiarism Detection") as demo:
     
     gr.Markdown("""
@@ -406,30 +329,40 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Soft(), title="Music Plagiarism D
             Authors: Seonghyeon Go, Yumin Kim | MIPPIA Inc. | Submitted to ICASSP 2026
         </p>
         <hr style="border: none; border-top: 1px solid #e5e7eb; margin: 15px 0;">
-        <p><strong>Demo Version Notice:</strong> This demo differs from the paper version and focuses exclusively on vocal segment transcription.</p>
-        <p style="font-size: 0.9em; color: #6b7280; margin: 8px 0;">
-            Structure analysis has been excluded for optimization. Results are derived from all downbeats, 
-            which may cause some timestamps to appear less precise than expected.
-        </p>
+        <p><strong>Demo Version Notice:</strong> This demo provides cropped audio segments of matched parts, rather than clickable timestamps.</p>
         <p style="color: #dc2626; font-weight: 600;">Processing can take up to 2 minutes per file</p>
     </div>
     """, elem_classes=["main-container"])
     
-    # Input section
     with gr.Row():
-        audio_input = gr.Audio(type="filepath", label="Upload Your Audio File")
+        audio_input = gr.Audio(type="filepath", label="Upload Your Audio File", elem_id="audio_input")
     
     with gr.Row():
         submit_btn = gr.Button("Analyze Audio", variant="primary", size="lg")
     
-    # Output section
+    audio_outputs = []
     with gr.Row():
-        results = gr.HTML(label="Analysis Results")
+        # 원본 오디오들을 위한 컴포넌트
+        with gr.Column(scale=1):
+            audio_outputs.append(gr.Audio(label=f"Your Original Audio", show_label=True))
+            for i in range(3):
+                audio_outputs.append(gr.Audio(label=f"Library Original Audio (Rank #{i+1})", show_label=True))
+        
+        # 매칭된 세그먼트들을 위한 컴포넌트
+        with gr.Column(scale=1):
+            for i in range(3):
+                audio_outputs.append(gr.Audio(label=f"Your Audio Segment (Rank #{i+1})", show_label=True))
+            for i in range(3):
+                audio_outputs.append(gr.Audio(label=f"Library Audio Segment (Rank #{i+1})", show_label=True))
+    
+    results = gr.HTML(label="Analysis Results")
+
+    all_outputs = audio_outputs + [results]
     
     submit_btn.click(
-        fn=process_audio_for_matching,
+        fn=process_audio_for_matching_and_crop,
         inputs=[audio_input],
-        outputs=[results]
+        outputs=all_outputs
     )
 
 if __name__ == "__main__":