Update app.py

app.py

@@ -71,7 +71,7 @@ dtype = 'bfloat16'
 ptdtype = {'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
 ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)
 
-SEQ_LEN = 8192
+SEQ_LEN = 1668
 PAD_IDX = 18819
 
 model = TransformerWrapper(num_tokens = PAD_IDX+1,
@@ -110,56 +110,79 @@ def load_midi(input_midi):
 
     raw_score = TMIDIX.midi2single_track_ms_score(input_midi)
 
-    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
-    
-    sp_escore_notes = TMIDIX.solo_piano_escore_notes(escore_notes)
-    zscore = TMIDIX.recalculate_score_timings(sp_escore_notes)
-    
-    escore = TMIDIX.augment_enhanced_score_notes(zscore, timings_divider=32)
-    
-    escore = TMIDIX.fix_escore_notes_durations(escore)
-    
-    cscore = TMIDIX.chordify_score([1000, escore])
-    
-    score = []
-    chords = []
-    
-    pc = cscore[0]
-    
-    for c in cscore:
-    
-        tones_chord = sorted(set([e[4] % 12 for e in c]))
-    
-        if tones_chord not in TMIDIX.ALL_CHORDS_SORTED:
-            tones_chord = TMIDIX.check_and_fix_tones_chord(tones_chord, use_full_chords=False)
-    
-        chord_tok = TMIDIX.ALL_CHORDS_SORTED.index(tones_chord)
-        chords.append(chord_tok+384)
-    
-        score.append(chord_tok+384)
-        score.append(max(0, min(127, c[0][1]-pc[0][1])))
-    
-        for n in c:
-            score.extend([max(1, min(127, n[2]))+128, max(1, min(127, n[4]))+256])
-    
-        pc = c
-    
-    print('Done!')
-    print('=' * 70)
-    print('Score has', len(chords), 'chords')
-    print('Score hss', len(score), 'tokens')
-    print('=' * 70)
+    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True, apply_sustain=True)
+
+    if escore_notes:
+    
+        escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes[0], sort_drums_last=True)
+        
+        dscore = TMIDIX.delta_score_notes(escore_notes)
+        
+        dcscore = TMIDIX.chordify_score([d[1:] for d in dscore])
+        
+        melody_chords = [18816]
+        
+        #=======================================================
+        # MAIN PROCESSING CYCLE
+        #=======================================================
+        
+        for i, c in enumerate(dcscore):
+        
+            delta_time = c[0][0]
+        
+            melody_chords.append(delta_time)
+        
+            for e in c:
+            
+                #=======================================================
+                
+                # Durations
+                dur = max(1, min(255, e[1]))
+        
+                # Patches
+                pat = max(0, min(128, e[5]))
+                
+                # Pitches
+                ptc = max(1, min(127, e[3]))
+                
+                # Velocities
+                # Calculating octo-velocity
+                
+                vel = max(8, min(127, e[4]))
+                velocity = round(vel / 15)-1
+                
+                #=======================================================
+                # FINAL NOTE SEQ
+                #=======================================================
+                
+                # Writing final note
+                pat_ptc = (128 * pat) + ptc 
+                dur_vel = (8 * dur) + velocity
+        
+                melody_chords.extend([pat_ptc+256, dur_vel+16768]) # 18816
+    
+        
+        print('Done!')
+        print('=' * 70)
+        print('Score hss', len(melody_chords), 'tokens')
+        print('=' * 70)
+
+        if len(melody_chords) > SEQ_LEN:
+            return melody_chords
+
+        else:
+            return None
 
-    return score, chords
+    else:
+        return None
 
 #==================================================================================
 
 @spaces.GPU
-def Generate_Chords_Textures(input_midi,
-                             prime_with_first_note,
-                             model_temperature,
-                             model_sampling_top_p
-                            ):
+def Generate_Music_Bridge(input_midi,
+                          model_temperature,
+                          model_sampling_top_p
+                         ):
 
     #===============================================================================
 
@@ -174,7 +197,6 @@ def Generate_Chords_Textures(input_midi,
     fn = os.path.basename(input_midi)
     fn1 = fn.split('.')[0]
     print('Input MIDI file name:', fn)
-    print('Prime with first note:', prime_with_first_note)
     print('Model temperature:', model_temperature)
     print('Model top p:', model_sampling_top_p)
    
@@ -182,145 +204,153 @@ def Generate_Chords_Textures(input_midi,
 
     #==================================================================
 
-    print('Loading MIDI...')
-
-    score, chords = load_midi(input_midi.name)
-
-    print('Sample score chords', chords[:10])
-    print('Sample score tokens', score[:10])
-
-    #==================================================================
-
-    chords_len = 128
-
-    if len(chords) < 128:
-        chords_len = len(chords)
-        chords = chords * ((128 // chords_len)+1)
-
-    chords = chords[:128]
-
-    #==================================================================
-    
-    print('=' * 70)
-    print('Generating...')
-
-    if prime_with_first_note:
-        seq = [705] + chords[:128] + [706] + score[:4]
-
-    else:
-        seq = [705] + chords[:128] + [706]
-
-    x = torch.LongTensor(seq).to(device_type)
-    
-    with ctx:
-        out = model.generate(x,
-                             1024,
-                             temperature=model_temperature,
-                             filter_logits_fn=top_p,
-                             filter_kwargs={'thres': model_sampling_top_p},
-                             return_prime=True,
-                             eos_token=707,
-                             verbose=False)
-
-    if prime_with_first_note:
-        final_song = out.tolist()[len(seq)-4:]
-
-    else:
-        final_song = out.tolist()[len(seq):]
-
-    #==================================================================
-
-    ccount = -1
-
-    for i, f in enumerate(final_song):
-
-        if f > 384:
-            ccount += 1
+    if input_midi is not None:
+
+        print('Loading MIDI...')
+    
+        score = load_midi(input_midi.name)
+    
+        if score is not None:
+            print('Sample score tokens', score[:10])
+        
+            #==================================================================
+        
+            chords_len = 128
+        
+            if len(chords) < 128:
+                chords_len = len(chords)
+                chords = chords * ((128 // chords_len)+1)
+        
+            chords = chords[:128]
+        
+            #==================================================================
             
-        if ccount == chords_len:
-            break
+            print('=' * 70)
+            print('Generating...')
+        
+            if prime_with_first_note:
+                seq = [705] + chords[:128] + [706] + score[:4]
+        
+            else:
+                seq = [705] + chords[:128] + [706]
+        
+            x = torch.LongTensor(seq).to(device_type)
             
-    final_song = final_song[:i]
-    
-    #==================================================================
-   
-    print('=' * 70)
-    print('Done!')
-    print('=' * 70)
-    
-    #===============================================================================
-    
-    print('Rendering results...')
-    
-    print('=' * 70)
-    print('Sample INTs', final_song[:15])
-    print('=' * 70)
-
-    song_f = []
-    
-    if len(final_song) != 0:
-    
-        time = 0
-        dur = 1
-        vel = 90
-        pitch = 60
-        channel = 0
-        patch = 0
-    
-        patches = [0] * 16
-    
-        for m in final_song:
-    
-            if 0 <= m < 128:
-                time += m * 32
-    
-            elif 128 < m < 256:
-                dur = (m-128) * 32
-    
-            elif 256 < m < 384:
-                pitch = (m-256)
-    
-                song_f.append(['note', time, dur, 0, pitch, max(40, pitch), 0])
-
-    fn1 = "Orpheus-Bridge-Music-Transformer-Composition"
-    
-    detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
-                                                              output_signature = 'Orpheus Bridge Music Transformer',
-                                                              output_file_name = fn1,
-                                                              track_name='Project Los Angeles',
-                                                              list_of_MIDI_patches=patches
-                                                              )
-    
-    new_fn = fn1+'.mid'
+            with ctx:
+                out = model.generate(x,
+                                     1024,
+                                     temperature=model_temperature,
+                                     filter_logits_fn=top_p,
+                                     filter_kwargs={'thres': model_sampling_top_p},
+                                     return_prime=True,
+                                     eos_token=707,
+                                     verbose=False)
+        
+            if prime_with_first_note:
+                final_song = out.tolist()[len(seq)-4:]
+        
+            else:
+                final_song = out.tolist()[len(seq):]
+        
+            #==================================================================
+        
+            ccount = -1
+        
+            for i, f in enumerate(final_song):
+        
+                if f > 384:
+                    ccount += 1
+                    
+                if ccount == chords_len:
+                    break
+                    
+            final_song = final_song[:i]
             
+            #==================================================================
+           
+            print('=' * 70)
+            print('Done!')
+            print('=' * 70)
+            
+            #===============================================================================
+            
+            print('Rendering results...')
+            
+            print('=' * 70)
+            print('Sample INTs', final_song[:15])
+            print('=' * 70)
+        
+            song_f = []
+            
+            if len(final_song) != 0:
+            
+                time = 0
+                dur = 1
+                vel = 90
+                pitch = 60
+                channel = 0
+                patch = 0
+            
+                patches = [0] * 16
+            
+                for m in final_song:
+            
+                    if 0 <= m < 128:
+                        time += m * 32
+            
+                    elif 128 < m < 256:
+                        dur = (m-128) * 32
+            
+                    elif 256 < m < 384:
+                        pitch = (m-256)
+            
+                        song_f.append(['note', time, dur, 0, pitch, max(40, pitch), 0])
+        
+            fn1 = "Orpheus-Bridge-Music-Transformer-Composition"
+            
+            detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
+                                                                      output_signature = 'Orpheus Bridge Music Transformer',
+                                                                      output_file_name = fn1,
+                                                                      track_name='Project Los Angeles',
+                                                                      list_of_MIDI_patches=patches
+                                                                      )
+            
+            new_fn = fn1+'.mid'
+                    
+            
+            audio = midi_to_colab_audio(new_fn, 
+                                soundfont_path=SOUDFONT_PATH,
+                                sample_rate=16000,
+                                volume_scale=10,
+                                output_for_gradio=True
+                                )
+            
+            print('Done!')
+            print('=' * 70)
+        
+            #========================================================
+        
+            output_midi = str(new_fn)
+            output_audio = (16000, audio)
+            output_plot = TMIDIX.plot_ms_SONG(song_f, plot_title=output_midi, return_plt=True)
+        
+            print('Output MIDI file name:', output_midi)
+            print('=' * 70) 
+            
+            #========================================================
     
-    audio = midi_to_colab_audio(new_fn, 
-                        soundfont_path=SOUDFONT_PATH,
-                        sample_rate=16000,
-                        volume_scale=10,
-                        output_for_gradio=True
-                        )
-    
-    print('Done!')
-    print('=' * 70)
-
-    #========================================================
-
-    output_midi = str(new_fn)
-    output_audio = (16000, audio)
-    output_plot = TMIDIX.plot_ms_SONG(song_f, plot_title=output_midi, return_plt=True)
-
-    print('Output MIDI file name:', output_midi)
-    print('=' * 70) 
+        else:
+            return None, None, None
     
-    #========================================================
+        print('-' * 70)
+        print('Req end time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
+        print('-' * 70)
+        print('Req execution time:', (reqtime.time() - start_time), 'sec')
     
-    print('-' * 70)
-    print('Req end time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
-    print('-' * 70)
-    print('Req execution time:', (reqtime.time() - start_time), 'sec')
+        return output_audio, output_plot, output_midi
 
-    return output_audio, output_plot, output_midi
+    return
+        None, None, None
     
 #==================================================================================
 
@@ -351,7 +381,7 @@ 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 demo is limited and will only texture first 128 chords of the MIDI file")
+    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"]
@@ -359,7 +389,6 @@ with gr.Blocks() as demo:
     
     gr.Markdown("## Generation options")
 
-    prime_with_first_note = gr.Checkbox(value=True, label="Use the first note onset, duration and pitch to prime the model")
     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")
     
@@ -372,9 +401,8 @@ with gr.Blocks() as demo:
     output_plot = gr.Plot(label="MIDI score plot")
     output_midi = gr.File(label="MIDI file", file_types=[".mid"])
 
-    generate_btn.click(Generate_Chords_Textures, 
+    generate_btn.click(Generate_Music_Bridge, 
                        [input_midi,
-                        prime_with_first_note,
                         model_temperature,
                         model_sampling_top_p
                        ], 
@@ -385,10 +413,9 @@ with gr.Blocks() as demo:
                       )
 
     gr.Examples(
-                [["Sharing The Night Together.kar", True, 0.9, 0.96]
+                [["Sharing The Night Together.kar", 0.9, 0.96]
                 ],
                 [input_midi,
-                 prime_with_first_note,
                  model_temperature,
                  model_sampling_top_p
                 ],
@@ -396,7 +423,7 @@ with gr.Blocks() as demo:
                  output_plot,
                  output_midi
                 ],
-                Generate_Chords_Textures
+                Generate_Music_Bridge
     )
     
 #==================================================================================