.

app.py

@@ -14,7 +14,7 @@ device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
 # API URLs and headers
 AUDIO_API_URL = "https://api-inference.huggingface.co/models/MIT/ast-finetuned-audioset-10-10-0.4593"
-LYRICS_API_URL = "https://api-inference.huggingface.co/models/gpt2-xl"
+LYRICS_API_URL = "https://api-inference.huggingface.co/models/gpt2-medium"
 headers = {"Authorization": f"Bearer {os.environ.get('HF_TOKEN')}"}
 
 def get_audio_duration(audio_path):
@@ -62,19 +62,22 @@ def create_lyrics_prompt(classification_results, song_structure):
     main_confidence = float(classification_results[0]['score'].strip('%'))
     secondary_elements = [result['label'] for result in classification_results[1:3]]
     
-    # Create a more specific songwriting prompt
-    prompt = f"""Write a song in the style of {main_style} music.
+    # Create a simpler prompt with example structure
+    prompt = f"""Here's a {main_style} song with {', '.join(secondary_elements)} elements:
 
-Style Guide:
-- Primary style: {main_style} ({main_confidence:.1f}% confidence)
-- Incorporate elements of: {', '.join(secondary_elements)}
-- Structure: {song_structure['verses']} verses and {song_structure['choruses']} choruses
-- Each verse and chorus should be exactly 4 lines
-- Write in proper song lyric format
-- Focus on musical rhythm and flow
-- Include rhyming where appropriate
+[Verse 1]
+The melody rings through the air tonight
+Like a gentle whisper in the light
+Every note tells a story so clear
+Creating magic for all to hear
 
-Start with the first verse:
+[Chorus]
+Let the rhythm flow and shine
+Feel the music so divine
+Every moment, every sound
+Brings the joy that we have found
+
+Now continue with your own lyrics in this style:
 
 [Verse 1]"""
     return prompt
@@ -137,7 +140,7 @@ def format_lyrics(generated_text, song_structure):
     return "\n".join(lines)
 
 def generate_lyrics_with_retry(prompt, song_structure, max_retries=5, initial_wait=2):
-    """Generate lyrics using GPT2-XL with retry logic"""
+    """Generate lyrics using GPT2-Medium with retry logic"""
     wait_time = initial_wait
     
     for attempt in range(max_retries):
@@ -149,13 +152,13 @@ def generate_lyrics_with_retry(prompt, song_structure, max_retries=5, initial_wa
                     "inputs": prompt,
                     "parameters": {
                         "max_new_tokens": song_structure['tokens'],
-                        "temperature": 0.85,  # Balanced between creativity and coherence
-                        "top_p": 0.92,
+                        "temperature": 0.7,  # Lower temperature for more focused output
+                        "top_p": 0.85,
                         "do_sample": True,
                         "return_full_text": False,
-                        "repetition_penalty": 1.2,
-                        "presence_penalty": 0.5,  # Encourage diverse vocabulary
-                        "frequency_penalty": 0.5  # Discourage repetitive phrases
+                        "repetition_penalty": 1.1,  # Reduced repetition penalty
+                        "presence_penalty": 0.3,
+                        "frequency_penalty": 0.3
                     }
                 }
             )
@@ -173,8 +176,8 @@ def generate_lyrics_with_retry(prompt, song_structure, max_retries=5, initial_wa
                     content_lines = [l for l in formatted_lyrics.split('\n') 
                                    if l.strip() and not l.strip().startswith('[') and l.strip() != '...']
                     
-                    # Check if the content looks like lyrics (short lines, not prose)
-                    if len(content_lines) < 4 or any(len(line.split()) > 15 for line in content_lines):
+                    # More lenient line length check
+                    if len(content_lines) < 4 or any(len(line.split()) > 20 for line in content_lines):
                         if attempt < max_retries - 1:
                             print("Generated text doesn't look like lyrics, retrying...")
                             continue