Update app.py

app.py

@@ -10,48 +10,31 @@ print('Loading core MIDI Loops Mixer modules...')
 
 import os
 import copy
+import statistics
+import random
 
 import time as reqtime
 import datetime
 from pytz import timezone
 
+import tqdm
+
 print('=' * 70)
 print('Loading main MIDI Loops Mixer modules...')
 
-os.environ['USE_FLASH_ATTENTION'] = '1'
-
-import torch
-
-torch.set_float32_matmul_precision('high')
-torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
-torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
-torch.backends.cuda.enable_mem_efficient_sdp(True)
-torch.backends.cuda.enable_math_sdp(True)
-torch.backends.cuda.enable_flash_sdp(True)
-torch.backends.cuda.enable_cudnn_sdp(True)
-
-from huggingface_hub import hf_hub_download
+import numpy as np
 
 import TMIDIX
 
 from midi_to_colab_audio import midi_to_colab_audio
 
-from x_transformer_1_23_2 import *
-
-import random
-
-import tqdm
+import gradio as gr
 
 print('=' * 70)
 print('Loading aux MIDI Loops Mixer modules...')
 
 import matplotlib.pyplot as plt
 
-import gradio as gr
-import spaces
-
-print('=' * 70)
-print('PyTorch version:', torch.__version__)
 print('=' * 70)
 print('Done!')
 print('Enjoy! :)')
@@ -59,182 +42,32 @@ print('=' * 70)
 
 #==================================================================================
 
-MODEL_CHECKPOINT = 'Guided_Accompaniment_Transformer_Trained_Model_36457_steps_0.5384_loss_0.8417_acc.pth'
-
 SOUDFONT_PATH = 'SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2'
 
 #==================================================================================
 
 print('=' * 70)
-print('Instantiating model...')
+print('Loading MIDI Loops Small Dataset model...')
 
-device_type = 'cuda'
-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
-
-model = TransformerWrapper(
-        num_tokens = PAD_IDX+1,
-        max_seq_len = SEQ_LEN,
-        attn_layers = Decoder(dim = 2048,
-                              depth = 4,
-                              heads = 32,
-                              rotary_pos_emb = True,
-                              attn_flash = True
-                              )
-)
-
-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-Loops-Mixer', filename=MODEL_CHECKPOINT)
-
-model.load_state_dict(torch.load(model_checkpoint, map_location='cpu', weights_only=True))
-
-model = torch.compile(model, mode='max-autotune')
+midi_loops_dataset = TMIDIX.Tegridy_Any_Pickle_File_Reader('MIDI-Loops-Dataset-Small-CC-BY-NC-SA.pickle')
 
 print('=' * 70)
 print('Done!')
 print('=' * 70)
-print('Model will use', dtype, 'precision...')
+print('Loaded', len(midi_loops_dataset), 'MIDI Loops')
 print('=' * 70)
 
 #==================================================================================
 
-def load_midi(input_midi, melody_patch=-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)
-
-        else:
-            zscore = TMIDIX.recalculate_score_timings(sp_escore_notes)
-    
-    cscore = TMIDIX.chordify_score([1000, zscore])
-    
-    score = []
-    
-    score_list = []
-    
-    pc = cscore[0]
-    
-    for c in cscore:
-        score.append(max(0, min(127, c[0][1]-pc[0][1])))
-    
-        scl = [[max(0, min(127, c[0][1]-pc[0][1]))]]
-    
-        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])
-    
-        score_list.append(scl)
-    
-        pc = c
-    
-    score_list.append(scl)
-
-    return score, score_list
-
-#==================================================================================
-
-@spaces.GPU
-def Generate_Accompaniment(input_midi, 
-                           generation_type,
-                           melody_patch,
-                           model_temperature
-                          ):
+def Mix_Loops(input_midi, 
+               generation_type,
+               melody_patch,
+               model_temperature
+              ):
 
     #===============================================================================
 
-    def generate_full_seq(input_seq, temperature=0.9, 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:
-    
-            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,
-                                     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
-    
-        return full_seq
-
-    #===============================================================================
 
-    def generate_block_seq(input_seq, trg_dtime, temperature=0.9):
-    
-        inp_seq = copy.deepcopy(input_seq)
-    
-        block_seq = []
-    
-        cur_time = 0
-    
-        while cur_time < trg_dtime:
-    
-            x = torch.LongTensor(inp_seq).cuda()
-            
-            with ctx:
-                out = model.generate(x,
-                                     1,
-                                     temperature=temperature,
-                                     return_prime=False,
-                                     verbose=False)
-            
-            y = out.tolist()[0][0]
-    
-            if y < 128:
-                cur_time += y
-    
-            inp_seq.append(y)
-            block_seq.append(y)
-    
-        if cur_time != trg_dtime:
-            return []
-    
-        else:
-            return block_seq
 
     #===============================================================================
     
@@ -267,69 +100,7 @@ def Generate_Accompaniment(input_midi,
     print('=' * 70)
     print('Generating...')
 
-    model.to(device_type)
-    model.eval()
-
-    #==================================================================
 
-    start_score_seq = [1792] + score + [1793]
-
-    #==================================================================
-
-    if generation_type == 'Guided':
-
-        input_seq = []
-
-        input_seq.extend(start_score_seq)
-        input_seq.extend(score_list[0][0])
-        
-        block_seq_lens = []
-        
-        idx = 0
-        
-        max_retries = 3
-        mrt = 0
-        
-        while idx < len(score_list)-1:
-        
-            if idx % 10 == 0:
-                print('Generating', idx, 'block')
-        
-            input_seq.extend(score_list[idx][1])
-        
-            block_seq = []
-        
-            for _ in range(max_retries):
-        
-                block_seq = generate_block_seq(input_seq, score_list[idx+1][0][0])
-        
-                if block_seq:
-                    break
-        
-            if block_seq:
-                input_seq.extend(block_seq)
-                block_seq_lens.append(len(block_seq))
-                idx += 1
-                mrt = 0
-        
-            else:
-        
-                if block_seq_lens:
-                    input_seq = input_seq[:-(block_seq_lens[-1]+2)]
-                    block_seq_lens.pop()
-                    idx -= 1
-                    mrt += 1
-        
-                else:
-                    break
-        
-                if mrt == max_retries:
-                    break
-
-    else:
-        input_seq = generate_full_seq(start_score_seq, temperature=model_temperature)
-
-    final_song = input_seq[len(start_score_seq):]
    
     print('=' * 70)
     print('Done!')
@@ -432,7 +203,7 @@ with gr.Blocks() as demo:
     #==================================================================================
 
     gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>MIDI Loops Mixer</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'>Mix random MIDI loops into one coherent music composition</h1>")
     gr.HTML("""            
             <p> 
                 <a href="https://huggingface.co/spaces/asigalov61/MIDI-Loops-Mixer?duplicate=true">
@@ -474,24 +245,7 @@ with gr.Blocks() as demo:
                         output_midi                          
                        ]
                       )
-
-    gr.Examples(
-                [["USSR-National-Anthem-Seed-Melody.mid", "Guided", -1, 0.9],
-                 ["Hotel-California-Seed-Melody.mid", "Guided", -1, 0.9],
-                 ["Sparks-Fly-Seed-Melody.mid", "Guided", -1, 0.9]
-                ],
-                [input_midi, 
-                 generation_type,
-                 melody_patch,
-                 model_temperature
-                ],
-                [output_audio,
-                 output_plot,
-                 output_midi
-                ],
-                Generate_Accompaniment
-    )
-        
+ 
 #==================================================================================
 
 demo.launch()