Update app.py

app.py

@@ -61,24 +61,35 @@ print('=' * 70)
 
 #==================================================================================
 
-MODEL_CHECKPOINT = 'MIDI_Genre_Classifier_Trained_Model_36457_steps_0.5384_loss_0.8417_acc.pth'
+MODEL_CHECKPOINT = 'Giant_Music_Transformer_Medium_Trained_Model_42174_steps_0.5211_loss_0.8542_acc.pth'
 
 SOUDFONT_PATH = 'SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2'
 
-MAX_MELODY_NOTES = 64
+#==================================================================================
+
+print('=' * 70)
+print('Loading MIDI GAS processed scores dataset...')
+
+midi_gas_ps_pickle = hf_hub_download(repo_id='asigalov61/MIDI-GAS', 
+                                               filename='MIDI_GAS_Processed_Scores_CC_BY_NC_SA.pickle'
+                                              )
+
+midi_gas_ps = TMIDIX.Tegridy_Any_Pickle_File_Reader(midi_gas_ps_pickle)
 
-MAX_GEN_TOKS = 3072
+print('=' * 70)
+print('Done!')
+print('=' * 70)
 
 #==================================================================================
 
 print('=' * 70)
-print('Loading popular hook melodies dataset...')
+print('Loading MIDI GAS processed scores embeddings dataset...')
 
-popular_hook_melodies_pickle = hf_hub_download(repo_id='asigalov61/MIDI-Genre-Classifier', 
-                                               filename='popular_hook_melodies_24_64_CC_BY_NC_SA.pickle'
+midi_gas_pse_pickle = hf_hub_download(repo_id='asigalov61/MIDI-GAS', 
+                                               filename='MIDI_GAS_Processed_Scores_Embeddings_CC_BY_NC_SA.pickle'
                                               )
 
-popular_hook_melodies = TMIDIX.Tegridy_Any_Pickle_File_Reader(popular_hook_melodies_pickle)
+midi_gas_pse = TMIDIX.Tegridy_Any_Pickle_File_Reader(midi_gas_pse_pickle)
 
 print('=' * 70)
 print('Done!')
@@ -89,20 +100,20 @@ print('=' * 70)
 print('=' * 70)
 print('Instantiating model...')
 
-device_type = 'cuda'
+device_type = 'cpu'
 dtype = 'bfloat16'
 
 ptdtype = {'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
 ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)
 
-SEQ_LEN = 4096
-PAD_IDX = 1794
+SEQ_LEN = 8192
+PAD_IDX = 19463
 
 model = TransformerWrapper(
         num_tokens = PAD_IDX+1,
         max_seq_len = SEQ_LEN,
         attn_layers = Decoder(dim = 2048,
-                              depth = 4,
+                              depth = 8,
                               heads = 32,
                               rotary_pos_emb = True,
                               attn_flash = True
@@ -114,12 +125,10 @@ model = AutoregressiveWrapper(model, ignore_index=PAD_IDX, pad_value=PAD_IDX)
 print('=' * 70)
 print('Loading model checkpoint...')      
 
-model_checkpoint = hf_hub_download(repo_id='asigalov61/MIDI-Genre-Classifier', filename=MODEL_CHECKPOINT)
+model_checkpoint = hf_hub_download(repo_id='asigalov61/Giant-Music-Transformer', filename=MODEL_CHECKPOINT)
 
 model.load_state_dict(torch.load(model_checkpoint, map_location='cpu', weights_only=True))
 
-model = torch.compile(model, mode='max-autotune')
-
 print('=' * 70)
 print('Done!')
 print('=' * 70)
@@ -128,111 +137,46 @@ print('=' * 70)
 
 #==================================================================================
 
-def load_midi(input_midi, melody_patch=-1, use_nth_note=1):
-
-    raw_score = TMIDIX.midi2single_track_ms_score(input_midi)
-
-    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
-    escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes, timings_divider=32)
-    
-    sp_escore_notes = TMIDIX.solo_piano_escore_notes(escore_notes, keep_drums=False)
-
-    if melody_patch == -1:
-        zscore = TMIDIX.recalculate_score_timings(sp_escore_notes)
-
-    else:
-        mel_score = [e for e in sp_escore_notes if e[6] == melody_patch]
-
-        if mel_score:
-            zscore = TMIDIX.recalculate_score_timings(mel_score)
+def load_midi(input_midi):
 
-        else:
-            zscore = TMIDIX.recalculate_score_timings(sp_escore_notes)
-    
-    cscore = TMIDIX.chordify_score([1000, zscore])[:MAX_MELODY_NOTES:use_nth_note]
+    raw_score = TMIDIX.midi2single_track_ms_score(midi_file)
     
-    score = []
+    escore = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
     
-    score_list = []
+    escore_notes = TMIDIX.augment_enhanced_score_notes(escore)
     
-    pc = cscore[0]
+    instruments_list = list(set([y[6] for y in escore_notes]))
     
-    for c in cscore:
-        score.append(max(0, min(127, c[0][1]-pc[0][1])))
+    tok_score = []
     
-        scl = [[max(0, min(127, c[0][1]-pc[0][1]))]]
+    if 128 in instruments_list:
+        drums_present = 19331
     
-        n = c[0]
-        
-        score.extend([max(1, min(127, n[2]))+128, max(1, min(127, n[4]))+256])
-        scl.append([max(1, min(127, n[2]))+128, max(1, min(127, n[4]))+256])
+    else:
+        drums_present = 19330
     
-        score_list.append(scl)
+    pat = escore_notes[0][6]
     
-        pc = c
+    tok_score.extend([19461, drums_present, 19332+pat])
     
-    score_list.append(scl)
+    tok_score.extend(TMIDIX.multi_instrumental_escore_notes_tokenized(escore_notes)[:8190])
 
-    return score, score_list
+    return tok_score
 
 #==================================================================================
 
-@spaces.GPU
-def Generate_Accompaniment(input_midi,
-                           input_melody,
-                           melody_patch,
-                           use_nth_note,
-                           model_temperature,
-                           model_sampling_top_k
-                          ):
+# @spaces.GPU
+def Classify_MIDI_Genre(input_midi,
+                        input_melody,
+                        melody_patch,
+                        use_nth_note,
+                        model_temperature,
+                        model_sampling_top_k
+                       ):
 
     #===============================================================================
 
-    def generate_full_seq(input_seq, 
-                          max_toks=3072, 
-                          temperature=0.9, 
-                          top_k_value=15, 
-                          verbose=True
-                         ):
-    
-        seq_abs_run_time = sum([t for t in input_seq if t < 128])
-    
-        cur_time = 0
-    
-        full_seq = copy.deepcopy(input_seq)
-    
-        toks_counter = 0
-    
-        while cur_time <= seq_abs_run_time+32:
-    
-            if verbose:
-                if toks_counter % 128 == 0:
-                    print('Generated', toks_counter, 'tokens')
-    
-            x = torch.LongTensor(full_seq).cuda()
-    
-            with ctx:
-                out = model.generate(x,
-                                     1,
-                                     filter_logits_fn=top_k,
-                                     filter_kwargs={'k': top_k_value},
-                                     temperature=temperature,
-                                     return_prime=False,
-                                     verbose=False)
-            
-            y = out.tolist()[0][0]
-    
-            if y < 128:
-                cur_time += y
-    
-            full_seq.append(y)
-    
-            toks_counter += 1
 
-            if toks_counter == max_toks:
-                return full_seq
-    
-        return full_seq
 
     #===============================================================================
 
@@ -326,29 +270,51 @@ def Generate_Accompaniment(input_midi,
         vel = 90
         pitch = 0
         channel = 0
-        patch = 0
-    
-        channels_map = [0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 9, 12, 13, 14, 15]
-        patches_map = [40, 0, 10, 19, 24, 35, 40, 52, 56, 9, 65, 73, 0, 0, 0, 0]
-        velocities_map = [125, 80, 100, 80, 90, 100, 100, 80, 110, 110, 110, 110, 80, 80, 80, 80]
-    
-        for m in final_song:
-    
-            if 0 <= m < 128:
-                time += m * 32
-    
-            elif 128 < m < 256:
-                dur = (m-128) * 32
-    
-            elif 256 < m < 1792:
-                cha = (m-256) // 128
-                pitch = (m-256) % 128
-    
-                channel = channels_map[cha]
-                patch = patches_map[channel]
-                vel = velocities_map[channel]
-    
-                song_f.append(['note', time, dur, channel, pitch, vel, patch])
+        
+        patches = [-1] * 16
+        patches[9] = 9
+        
+        channels = [0] * 16
+        channels[9] = 1
+        
+        for ss in song:
+        
+          if 0 <= ss < 256:
+        
+              time += ss * 16
+        
+          if 256 <= ss < 2304:
+        
+              dur = ((ss-256) // 8) * 16
+              vel = (((ss-256) % 8)+1) * 15
+        
+          if 2304 <= ss < 18945:
+        
+              patch = (ss-2304) // 129
+        
+              if patch < 128:
+        
+                  if patch not in patches:
+                    if 0 in channels:
+                        cha = channels.index(0)
+                        channels[cha] = 1
+                    else:
+                        cha = 15
+        
+                    patches[cha] = patch
+                    channel = patches.index(patch)
+                  else:
+                    channel = patches.index(patch)
+        
+              if patch == 128:
+                  channel = 9
+        
+              pitch = (ss-2304) % 129
+        
+              song_f.append(['note', time, dur, channel, pitch, vel, patch ])
+        
+        patches = [0 if x==-1 else x for x in patches]
+
 
     fn1 = "MIDI-Genre-Classifier-Composition"
     
@@ -356,7 +322,7 @@ def Generate_Accompaniment(input_midi,
                                                               output_signature = 'MIDI Genre Classifier',
                                                               output_file_name = fn1,
                                                               track_name='Project Los Angeles',
-                                                              list_of_MIDI_patches=patches_map
+                                                              list_of_MIDI_patches=patches
                                                               )
     
     new_fn = fn1+'.mid'
@@ -408,7 +374,8 @@ with gr.Blocks() as demo:
     #==================================================================================
 
     gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>MIDI Genre Classifier</h1>")
-    gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Guided melody accompaniment generation with transformers</h1>")
+    gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Accurately classify any MIDI by top music genre/h1>")
+    
     gr.HTML("""            
             <p> 
                 <a href="https://huggingface.co/spaces/asigalov61/MIDI-Genre-Classifier?duplicate=true">
@@ -416,7 +383,7 @@ with gr.Blocks() as demo:
                 </a>
             </p>
             
-            for faster execution and endless generation!
+            for faster execution and endless classification!
             """)
     
     #==================================================================================
@@ -462,25 +429,6 @@ with gr.Blocks() as demo:
                         output_midi                          
                        ]
                       )
-
-    gr.Examples(
-                [["USSR-National-Anthem-Seed-Melody.mid", "Custom MIDI", -1, 1, 0.9, 15],
-                 ["Sparks-Fly-Seed-Melody.mid", "Custom MIDI", -1, 1, 0.9, 15]
-                ],
-                [input_midi,
-                 input_melody,
-                 melody_patch,
-                 use_nth_note,
-                 model_temperature,
-                 model_sampling_top_k
-                ],
-                [output_title,
-                 output_audio,
-                 output_plot,
-                 output_midi
-                ],
-                Generate_Accompaniment
-    )
         
 #==================================================================================