Update app.py

app.py

@@ -23,15 +23,16 @@ in_space = os.getenv("SYSTEM") == "spaces"
 # =================================================================================================
                        
 @spaces.GPU
-def GenerateMusic(input_title, input_num_tokens, input_prompt_type):
+def GenerateMusic(input_title):
+    
     print('=' * 70)
     print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
     start_time = reqtime.time()
 
     print('Loading model...')
 
-    SEQ_LEN = 4096 # Models seq len
-    PAD_IDX = 2571 # Models pad index
+    SEQ_LEN = 2048
+    PAD_IDX = 780
     DEVICE = 'cuda' # 'cuda'
 
     # instantiate the model
@@ -39,10 +40,10 @@ def GenerateMusic(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 = 8, heads = 16, attn_flash = True)
+        attn_layers = Decoder(dim = 1024, depth = 32, heads = 16, attn_flash = True)
         )
     
-    model = AutoregressiveWrapper(model, ignore_index = PAD_IDX)
+    model = AutoregressiveWrapper(model, ignore_index = PAD_IDX, pad_value=PAD_IDX)
 
     model.to(DEVICE)
     print('=' * 70)
@@ -50,7 +51,7 @@ def GenerateMusic(input_title, input_num_tokens, input_prompt_type):
     print('Loading model checkpoint...')
 
     model.load_state_dict(
-        torch.load('Text_to_Music_Transformer_Medium_Trained_Model_33934_steps_0.6093_loss_0.813_acc.pth',
+        torch.load('Descriptive_Music_Transformer_Trained_Model_20631_steps_0.3218_loss_0.8947_acc.pth',
                    map_location=DEVICE))
     print('=' * 70)
 
@@ -66,83 +67,25 @@ def GenerateMusic(input_title, input_num_tokens, input_prompt_type):
     print('Done!')
     print('=' * 70)
 
-    input_num_tokens = max(8, min(2048, input_num_tokens))
-
-    print('-' * 70)
-    print('Input title:', input_title)
-    print('Req num toks:', input_num_tokens)
-    print('Open-ended prompt:', input_prompt_type)
+    input_num_tokens = 2040
     print('-' * 70)
 
     #===============================================================================
 
-    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()]
+        return [ord(char)+648 if 0 < ord(char) < 128 else 0+648 for char in txt.lower()]
     
     def tokens2txt(tokens):
-        return [chr(tok-2440) for tok in tokens if 0+2440 < tok < 128+2440 ]
+        return [chr(tok-648) for tok in tokens if 0+648 < tok < 128+648 ]
     
     print('=' * 70)
-    print('Setting up patches...')
-    print('=' * 70)
-    
-    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)
-                           ]
-    
-    print('Done!')
-    print('=' * 70)
-
     print('Generating...')
     
     #@title Standard Text-to-Music Generator
     
-    #@markdown Prompt settings
-    
-    song_title_prompt = input_title
-    open_ended_prompt = input_prompt_type
-    
     #@markdown Generation settings
     
     number_of_tokens_to_generate = input_num_tokens
@@ -150,22 +93,10 @@ def GenerateMusic(input_title, input_num_tokens, input_prompt_type):
     temperature = 0.9 # @param {type:"slider", min:0.1, max:1, step:0.05}
 
     print('=' * 70)
-    print('Text-to-Music Model Generator')
+    print('Descriptive Music Transformer Model Generator')
     print('=' * 70)
     
-    if song_title_prompt == '':
-      outy = [2569]
-    
-    else:
-      if open_ended_prompt:
-        outy = [2569] + txt2tokens(song_title_prompt)
-    
-      else:
-        outy = [2569] + txt2tokens(song_title_prompt) + [2570]
-    
-    print('Selected prompt sequence:')
-    print(outy[:12])
-    print('=' * 70)
+    outy = [777]
     
     torch.cuda.empty_cache()
     
@@ -313,25 +244,19 @@ if __name__ == "__main__":
             "(https://colab.research.google.com/github/asigalov61/Text-to-Music-Transformer/blob/main/Text_to_Music_Transformer.ipynb)"
             " for faster execution and endless generation"
         )
-        gr.Markdown("## Enter any desired song title")
 
-        input_title = gr.Textbox(value="Nothing Else Matters", label="Song title")
-        input_prompt_type = gr.Checkbox(label="Open-ended prompt")
-        
-        input_num_tokens = gr.Slider(8, 2048, value=512, step=8, label="Number of tokens to generate")
-        
         run_btn = gr.Button("generate", variant="primary")
 
         gr.Markdown("## Generation results")
 
         output_midi_title = gr.Textbox(label="Generated MIDI title")
-        output_midi_summary = gr.Textbox(label="Output MIDI summary")
+        output_midi_summary = gr.Textbox(label="Generated music description")
         output_audio = gr.Audio(label="Output MIDI audio", format="wav", elem_id="midi_audio")
         output_plot = gr.Plot(label="Output MIDI score plot")
         output_midi = gr.File(label="Output MIDI file", file_types=[".mid"])
 
 
-        run_event = run_btn.click(GenerateMusic, [input_title, input_num_tokens, input_prompt_type],
+        run_event = run_btn.click(GenerateMusic, [input_title],
                                   [output_midi_title, output_midi_summary, output_midi, output_audio, output_plot])
 
         app.queue().launch()