Update app.py

app.py

@@ -30,8 +30,8 @@ def Text_to_Music(input_title, input_num_tokens, input_prompt_type):
 
     print('Loading model...')
 
-    SEQ_LEN = 8192 # Models seq len
-    PAD_IDX = 767 # Models pad index
+    SEQ_LEN = 4096 # Models seq len
+    PAD_IDX = 2571 # Models pad index
     DEVICE = 'cuda' # 'cuda'
 
     # instantiate the model
@@ -39,7 +39,7 @@ def Text_to_Music(input_title, input_num_tokens, input_prompt_type):
     model = TransformerWrapper(
         num_tokens = PAD_IDX+1,
         max_seq_len = SEQ_LEN,
-        attn_layers = Decoder(dim = 2048, depth = 4, heads = 16, attn_flash = True)
+        attn_layers = Decoder(dim = 2048, depth = 8, heads = 16, attn_flash = True)
         )
     
     model = AutoregressiveWrapper(model, ignore_index = PAD_IDX)
@@ -50,7 +50,7 @@ def Text_to_Music(input_title, input_num_tokens, input_prompt_type):
     print('Loading model checkpoint...')
 
     model.load_state_dict(
-        torch.load('Ultimate_Accompaniment_Transformer_Small_Improved_Trained_Model_13649_steps_0.3229_loss_0.898_acc.pth',
+        torch.load('Text_to_Music_Transformer_Medium_Trained_Model_33934_steps_0.6093_loss_0.813_acc.pth',
                    map_location=DEVICE))
     print('=' * 70)
 
@@ -66,253 +66,186 @@ def Text_to_Music(input_title, input_num_tokens, input_prompt_type):
     print('Done!')
     print('=' * 70)
 
-    fn = os.path.basename(input_midi.name)
-    fn1 = fn.split('.')[0]
-
-    input_num_tokens = max(4, min(128, input_num_tokens))
+    input_num_tokens = max(8, min(2048, input_num_tokens))
 
     print('-' * 70)
-    print('Input file name:', fn)
+    print('Input title:', input_title)
     print('Req num toks:', input_num_tokens)
-    print('Force acc:', input_acc_type)
+    print('Open-ended prompt:', input_prompt_type)
     print('-' * 70)
 
     #===============================================================================
-    raw_score = TMIDIX.midi2single_track_ms_score(input_midi.name)
-    
-    #===============================================================================
-    # Enhanced score notes
-    
-    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
-    
-    escore_notes = [e for e in escore_notes if e[3] != 9]
-    
-    if len(escore_notes) > 0:
-    
-      #=======================================================
-      # PRE-PROCESSING
-    
-      #===============================================================================
-      # Augmented enhanced score notes
-    
-      escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes, timings_divider=32)
-    
-      cscore = TMIDIX.chordify_score([1000, escore_notes])
-    
-      melody = TMIDIX.fix_monophonic_score_durations([sorted(e, key=lambda x: x[4], reverse=True)[0] for e in cscore])
-    
-      #=======================================================
-      # FINAL PROCESSING
-    
-      melody_chords = []
-    
-      #=======================================================
-      # MAIN PROCESSING CYCLE
-      #=======================================================
-    
-      pe = cscore[0][0]
-    
-      mpe = melody[0]
-    
-      midx = 1
-    
-      for i, c in enumerate(cscore):
-    
-          c.sort(key=lambda x: (x[3], x[4]), reverse=True)
-    
-          # Next melody note
-    
-          if midx < len(melody):
-    
-            # Time
-            mtime = melody[midx][1]-mpe[1]
-    
-            mdur = melody[midx][2]
-    
-            mdelta_time = max(0, min(127, mtime))
-    
-            # Durations
-            mdur = max(0, min(127, mdur))
-    
-            # Pitch
-            mptc = melody[midx][4]
-    
-          else:
-            mtime = 127-mpe[1]
-    
-            mdur = mpe[2]
-    
-            mdelta_time = max(0, min(127, mtime))
-    
-            # Durations
-            mdur = max(0, min(127, mdur))
-    
-            # Pitch
-            mptc = mpe[4]
-    
-    
-          e = melody[i]
-    
-          #=======================================================
-          # Timings...
-    
-          time = e[1]-pe[1]
-    
-          dur = e[2]
-    
-          delta_time = max(0, min(127, time))
+
+    print('Setting up model patches and loading helper functions...')
+
+    # @title Setup and load model channels MIDI patches
+
+    model_channel_0_piano_family = "Acoustic Grand" # @param ["Acoustic Grand", "Bright Acoustic", "Electric Grand", "Honky-Tonk", "Electric Piano 1", "Electric Piano 2", "Harpsichord", "Clav"]
+    model_channel_1_chromatic_percussion_family = "Music Box" # @param ["Celesta", "Glockenspiel", "Music Box", "Vibraphone", "Marimba", "Xylophone", "Tubular Bells", "Dulcimer"]
+    model_channel_2_organ_family = "Church Organ" # @param ["Drawbar Organ", "Percussive Organ", "Rock Organ", "Church Organ", "Reed Organ", "Accordion", "Harmonica", "Tango Accordion"]
+    model_channel_3_guitar_family = "Acoustic Guitar(nylon)" # @param ["Acoustic Guitar(nylon)", "Acoustic Guitar(steel)", "Electric Guitar(jazz)", "Electric Guitar(clean)", "Electric Guitar(muted)", "Overdriven Guitar", "Distortion Guitar", "Guitar Harmonics"]
+    model_channel_4_bass_family = "Fretless Bass" # @param ["Acoustic Bass", "Electric Bass(finger)", "Electric Bass(pick)", "Fretless Bass", "Slap Bass 1", "Slap Bass 2", "Synth Bass 1", "Synth Bass 2"]
+    model_channel_5_strings_family = "Violin" # @param ["Violin", "Viola", "Cello", "Contrabass", "Tremolo Strings", "Pizzicato Strings", "Orchestral Harp", "Timpani"]
+    model_channel_6_ensemble_family = "Choir Aahs" # @param ["String Ensemble 1", "String Ensemble 2", "SynthStrings 1", "SynthStrings 2", "Choir Aahs", "Voice Oohs", "Synth Voice", "Orchestra Hit"]
+    model_channel_7_brass_family = "Trumpet" # @param ["Trumpet", "Trombone", "Tuba", "Muted Trumpet", "French Horn", "Brass Section", "SynthBrass 1", "SynthBrass 2"]
+    model_channel_8_reed_family = "Alto Sax" # @param ["Soprano Sax", "Alto Sax", "Tenor Sax", "Baritone Sax", "Oboe", "English Horn", "Bassoon", "Clarinet"]
+    model_channel_9_pipe_family = "Flute" # @param ["Piccolo", "Flute", "Recorder", "Pan Flute", "Blown Bottle", "Skakuhachi", "Whistle", "Ocarina"]
+    model_channel_10_synth_lead_family = "Lead 8 (bass+lead)" # @param ["Lead 1 (square)", "Lead 2 (sawtooth)", "Lead 3 (calliope)", "Lead 4 (chiff)", "Lead 5 (charang)", "Lead 6 (voice)", "Lead 7 (fifths)", "Lead 8 (bass+lead)"]
+    model_channel_11_synth_pad_family = "Pad 2 (warm)" # @param ["Pad 1 (new age)", "Pad 2 (warm)", "Pad 3 (polysynth)", "Pad 4 (choir)", "Pad 5 (bowed)", "Pad 6 (metallic)", "Pad 7 (halo)", "Pad 8 (sweep)"]
+    model_channel_12_synth_effects_family = "FX 3 (crystal)" # @param ["FX 1 (rain)", "FX 2 (soundtrack)", "FX 3 (crystal)", "FX 4 (atmosphere)", "FX 5 (brightness)", "FX 6 (goblins)", "FX 7 (echoes)", "FX 8 (sci-fi)"]
+    model_channel_13_ethnic_family = "Banjo" # @param ["Sitar", "Banjo", "Shamisen", "Koto", "Kalimba", "Bagpipe", "Fiddle", "Shanai"]
+    model_channel_14_percussive_family = "Melodic Tom" # @param ["Tinkle Bell", "Agogo", "Steel Drums", "Woodblock", "Taiko Drum", "Melodic Tom", "Synth Drum", "Reverse Cymbal"]
+    model_channel_15_sound_effects_family = "Bird Tweet" # @param ["Guitar Fret Noise", "Breath Noise", "Seashore", "Bird Tweet", "Telephone Ring", "Helicopter", "Applause", "Gunshot"]
+    model_channel_16_drums_family = "Drums" # @param ["Drums"]
+
+    print('=' * 70)
+    print('Loading helper functions...')
+
+    def txt2tokens(txt):
+        return [ord(char)+2440 if 0 < ord(char) < 128 else 0+2440 for char in txt.lower()]
     
-          # Durations
+    def tokens2txt(tokens):
+        return [chr(tok-2440) for tok in tokens if 0+2440 < tok < 128+2440 ]
     
-          dur = max(0, min(127, dur))
+    print('=' * 70)
+    print('Setting up patches...')
+    print('=' * 70)
     
-          # Pitches
+    instruments = [v[1] for v in TMIDIX.Number2patch.items()]
+    
+    patches = [instruments.index(model_channel_0_piano_family),
+                           instruments.index(model_channel_1_chromatic_percussion_family),
+                           instruments.index(model_channel_2_organ_family),
+                           instruments.index(model_channel_3_guitar_family),
+                           instruments.index(model_channel_4_bass_family),
+                           instruments.index(model_channel_5_strings_family),
+                           instruments.index(model_channel_6_ensemble_family),
+                           instruments.index(model_channel_7_brass_family),
+                           instruments.index(model_channel_8_reed_family),
+                           9, # Drums patch
+                           instruments.index(model_channel_9_pipe_family),
+                           instruments.index(model_channel_10_synth_lead_family),
+                           instruments.index(model_channel_11_synth_pad_family),
+                           instruments.index(model_channel_12_synth_effects_family),
+                           instruments.index(model_channel_13_ethnic_family),
+                           instruments.index(model_channel_15_sound_effects_family)
+                           ]
     
-          ptc = max(1, min(127, e[4]))
+    print('Done!')
+    print('=' * 70)
+
+    print('Generating...')
     
-          if ptc < 60:
-            ptc = 60 + (ptc % 12)
+    #@title Standard Text-to-Music Generator
     
-          cha = e[3]
+    #@markdown Prompt settings
     
-          #=======================================================
-          # FINAL NOTE SEQ
+    song_title_prompt = input_title
+    open_ended_prompt = input_prompt_type
     
-          if midx < len(melody):
-            melody_chords.append([delta_time, dur+128, ptc+384, mdelta_time+512, mptc+640])
-            mpe = melody[midx]
-            midx += 1
-          else:
-            melody_chords.append([delta_time, dur+128, ptc+384, mdelta_time+512, mptc+640])
+    #@markdown Generation settings
     
-          pe = e
-
-    #===============================================================================
+    number_of_tokens_to_generate = input_num_tokens
+    number_of_batches_to_generate = 1 #@param {type:"slider", min:1, max:16, step:1}
+    temperature = 0.9 # @param {type:"slider", min:0.1, max:1, step:0.05}
 
     print('=' * 70)
-    
-    print('Sample output events', melody_chords[:5])
+    print('Text-to-Music Model Generator')
     print('=' * 70)
-    print('Generating...')
-
-    output = []
     
-    force_acc = input_acc_type
-    num_toks_per_note = 32
-    temperature=0.9
-    max_drums_limit=4
-    num_memory_tokens=4096
-
-    output1 = []
-    output2 = []
+    if song_title_prompt == '':
+      outy = [2569]
     
-
-    for m in melody_chords[:input_num_tokens]:
-    
-        output1.extend(m)
-
-        input_seq = output1
-    
-        if force_acc:
-          x = torch.LongTensor([input_seq+[0]]).cuda()
-        else:
-          x = torch.LongTensor([input_seq]).cuda()
+    else:
+      if open_ended_prompt:
+        outy = [2569] + txt2tokens(song_title_prompt)
     
-        time = input_seq[-2]-512
+      else:
+        outy = [2569] + txt2tokens(song_title_prompt) + [2570]
     
-        cur_time = 0
+    print('Selected prompt sequence:')
+    print(outy[:12])
+    print('=' * 70)
     
-        for _ in range(num_toks_per_note):
-          with ctx:
-            out = model.generate(x[-num_memory_tokens:],
-                                1,
-                                temperature=temperature,
-                                return_prime=False,
-                                verbose=False)
+    torch.cuda.empty_cache()
     
-          o = out.tolist()[0][0]
+    inp = [outy] * number_of_batches_to_generate
     
-          if 0 <= o < 128:
-            cur_time += o
+    inp = torch.LongTensor(inp).cuda()
     
-          if cur_time < time and o < 384:
+    with ctx:
+      out = model.generate(inp,
+                            number_of_tokens_to_generate,
+                            temperature=temperature,
+                            return_prime=True,
+                            verbose=False)
     
-            out = torch.LongTensor([[o]]).cuda()
-            x = torch.cat((x, out), 1)
-          else:
-            break
+    out0 = out.tolist()
     
-        outy =  x.tolist()[0][len(input_seq):]
-
-        output1.extend(outy)
-        output2.append(outy)
-
     print('=' * 70)
     print('Done!')
     print('=' * 70)
     
     #===============================================================================
     print('Rendering results...')
-    
-    print('=' * 70)
-    print('Sample INTs', output1[:12])
     print('=' * 70)
+ 
+    out1 = out0[i]
     
-    out1 = output2
-
-    accompaniment_MIDI_patch_number = 0
-    melody_MIDI_patch_number = 40
+    print('Sample INTs', out1[:12])
+    print('=' * 70)
 
-    if len(out1) != 0:
+    generated_song_title = ''.join(tokens2txt(out1)).title()
     
-        song = out1
-        song_f = []
+    print('Generated song title:', generated_song_title)
+    print('=' * 70)
     
-        time = 0
-        ntime = 0
-        ndur = 0
-        vel = 90
-        npitch = 0
-        channel = 0
+    if len(out1) != 0:
     
-        patches = [0] * 16
-        patches[0] = accompaniment_MIDI_patch_number
-        patches[3] = melody_MIDI_patch_number
+      song = out1
+      song_f = []
     
-        for i, ss in enumerate(song):
+      time = 0
+      dur = 0
+      vel = 90
+      pitch = 0
+      channel = 0
     
-                ntime += melody_chords[i][0] * 32
-                ndur = (melody_chords[i][1]-128) * 32
-                nchannel = 1
-                npitch = (melody_chords[i][2]-256) % 128
-                vel = max(40, npitch)+20
+      for ss in song:
     
-                song_f.append(['note', ntime, ndur, 3, npitch, vel, melody_MIDI_patch_number ])
+          if 0 <= ss < 128:
     
-                time = ntime
+              time += ss * 32
     
-                for s in ss:
+          if 128 <= ss < 256:
     
-                  if 0 <= s < 128:
+              dur = (ss-128) * 32
     
-                      time += s * 32
+          if 256 <= ss < 2432:
     
-                  if 128 <= s < 256:
+              chan = (ss-256) // 128
     
-                      dur = (s-128) * 32
+              if chan < 9:
+                  channel = chan
+              elif 9 < chan < 15:
+                  channel = chan+1
+              elif chan == 15:
+                  channel = 15
+              elif chan == 16:
+                  channel = 9
     
-                  if 256 <= s < 384:
+              pitch = (ss-256) % 128
     
-                      pitch = (s-256)
+          if 2432 <= ss < 2440:
     
-                      vel = max(40, pitch)
+              vel = (((ss-2432)+1) * 15)-1
     
-                      song_f.append(['note', time, dur, 0, pitch, vel, accompaniment_MIDI_patch_number])
+              song_f.append(['note', time, dur, channel, pitch, vel, chan*8 ])
 
-    fn1 = "Ultimate-Accompaniment-Transformer-Composition"
+    fn1 = "Text-to-Music-Transformer-Composition"
     
     detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
-                                                              output_signature = 'Ultimate Accompaniment Transformer',
+                                                              output_signature = 'Text-to-Music Transformer',
                                                               output_file_name = fn1,
                                                               track_name='Project Los Angeles',
                                                               list_of_MIDI_patches=patches
@@ -333,7 +266,7 @@ def Text_to_Music(input_title, input_num_tokens, input_prompt_type):
 
     #========================================================
 
-    output_midi_title = str(fn1)
+    output_midi_title = generated_song_title
     output_midi_summary = str(song_f[:3])
     output_midi = str(new_fn)
     output_audio = (16000, audio)
@@ -342,7 +275,7 @@ def Text_to_Music(input_title, input_num_tokens, input_prompt_type):
 
     print('Output MIDI file name:', output_midi)
     print('Output MIDI title:', output_midi_title)
-    print('Output MIDI summary:', '')
+    print('Output MIDI summary:', output_midi_summary)
     print('=' * 70) 
     
 
@@ -390,7 +323,7 @@ if __name__ == "__main__":
 
         gr.Markdown("## Generation results")
 
-        output_midi_title = gr.Textbox(label="Output MIDI title")
+        output_midi_title = gr.Textbox(label="Generated MIDI title")
         output_midi_summary = gr.Textbox(label="Output MIDI summary")
         output_audio = gr.Audio(label="Output MIDI audio", format="wav", elem_id="midi_audio")
         output_plot = gr.Plot(label="Output MIDI score plot")