Update app.py

app.py

@@ -167,11 +167,8 @@ def load_midi(input_midi):
         print('Score hss', len(melody_chords), 'tokens')
         print('=' * 70)
 
-        if len(melody_chords) > SEQ_LEN:
-            return melody_chords
 
-        else:
-            return None
+        return melody_chords
 
     else:
         return None
@@ -215,35 +212,71 @@ def Generate_Music_Bridge(input_midi,
             print('Sample score tokens', score[:10])
         
             #==================================================================
+            
+            print('=' * 70)
+            print('Inpainting...')
         
-            full_chunk = score[:1536]
-            left_chunk = full_chunk[:512]
-            right_chunk = full_chunk[-512:]
+            patches_to_inpaint = [0]
+            inpaint_every_nth_note = 1
+            forbid_original_pitches = True
+            max_inpainted_pitch_dev = 12
+            max_inpaint_tries_per_note = 20
             
-            bridge_chunk = full_chunk[448:1088]
+            num_prime_tokens = 128
+            num_mem_tokens = 4096
             
-            seq = [18815] + left_chunk + [18816] + right_chunk + [18817]
-        
-            #==================================================================
+            notes_counter = 0
             
-            print('=' * 70)
-            print('Generating...')
-        
-            x = torch.LongTensor(seq).to(device_type)
+            inpainted_song = melody_chords[:num_prime_tokens]
             
-            with ctx:
-                out = model.generate(x,
-                                     641,
-                                     temperature=model_temperature,
-                                     filter_logits_fn=top_p,
-                                     filter_kwargs={'thres': model_sampling_top_p},
-                                     return_prime=False,
-                                     eos_token=18818,
-                                     verbose=False)
-
-            y = out.tolist()
-
-            final_song = left_chunk + y[64:-64] + right_chunk
+            for i, t in enumerate(melody_chords[num_prime_tokens:]):
+            
+                if 256 <= t < 16768:
+            
+                    old_patch = (t-256) // 128
+                    old_pitch = (t-256) % 128
+            
+                    if old_patch in patches_to_inpaint and notes_counter % inpaint_every_nth_note == 0:
+                    
+                        x = torch.LongTensor(inpainted_song[-num_mem_tokens:]).cuda()
+            
+                        new_pitch = old_pitch
+                        tries = 0
+            
+                        while (old_pitch == new_pitch or new_pitch > old_pitch + max_inpainted_pitch_dev or new_pitch < old_pitch - max_inpainted_pitch_dev) and tries < max_inpaint_tries_per_note:
+                        
+                            with ctx:
+                                out = model.generate(x,
+                                                     1,
+                                                     temperature=0.9,
+                                                     filter_logits_fn=top_k,
+                                                     filter_kwargs={'k': 2},
+                                                     return_prime=False,
+                                                     verbose=False)
+                            
+                            y = out.tolist()[0]
+                
+                            new_pitch = (y-256) % 128
+            
+                            tries += 1
+            
+                            if not forbid_original_pitches:
+                                break
+            
+                        if tries == max_inpaint_tries_per_note:
+                            new_pitch = old_pitch
+                            
+                        new_patch_pitch_tok = (128 * old_patch) + new_pitch + 256
+                
+                        inpainted_song.append(new_patch_pitch_tok)
+                        
+                    else:
+                        inpainted_song.append(t)
+                        
+                else:
+                    inpainted_song.append(t)
+            
+                notes_counter += 1
         
             #==================================================================
            
@@ -372,12 +405,16 @@ print('=' * 70)
 
 #==================================================================================
 
+patch2number = {v: k for k, v in TMIDIX.Number2patch.items()}
+
+#==================================================================================
+
 with gr.Blocks() as demo:
 
     #==================================================================================
 
     gr.Markdown("<h1 style='text-align: left; margin-bottom: 1rem'>Orpheus Pitches Inpainter</h1>")
-    gr.Markdown("<h1 style='text-align: left; margin-bottom: 1rem'>Seamless music bridges generation with transformers</h1>")
+    gr.Markdown("<h1 style='text-align: left; margin-bottom: 1rem'>Inpaint pitches in any MIDI composition</h1>")
     gr.HTML("""            
             <p> 
                 <a href="https://huggingface.co/spaces/projectlosangeles/Orpheus-Pitches-Inpainter?duplicate=true">
@@ -391,7 +428,6 @@ with gr.Blocks() as demo:
     #==================================================================================
     
     gr.Markdown("## Upload source MIDI or select a sample MIDI on the bottom of the page")
-    gr.Markdown("### PLEASE NOTE: The MIDI file MUST HAVE at least 800 MIDI pitches for the demo to work properly!")
     
     input_midi = gr.File(label="Input MIDI", 
                          file_types=[".midi", ".mid", ".kar"]
@@ -399,8 +435,24 @@ with gr.Blocks() as demo:
     
     gr.Markdown("## Generation options")
 
+
+    patches_to_inpaint = [0]
+    inpaint_every_nth_note = 1
+    forbid_original_pitches = True
+    max_inpainted_pitch_dev = 12
+    max_inpaint_tries_per_note = 20
+    
+    num_prime_tokens = 128
+    num_mem_tokens = 4096
+
+
+    inpaint_instruments = gr.Dropdown(label="Select instruments to inpaint", choices=list(patch2number.keys()),
+                                      multiselect=True, type="value",
+                                      info="Instruments MUST be present in the composition. For best results select a single instrument."
+                                     )
+
     model_temperature = gr.Slider(0.1, 1, value=0.9, step=0.01, label="Model temperature")
-    model_sampling_top_p = gr.Slider(0.1, 0.99, value=0.96, step=0.01, label="Model sampling top p value")
+    model_sampling_top_k = gr.Slider(1, 100, value=15, step=1, label="Model sampling top k value")
     
     generate_btn = gr.Button("Generate", variant="primary")