Update app.py

app.py

@@ -71,8 +71,9 @@ 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'
-                                              )
+                                     filename='MIDI_GAS_Processed_Scores_CC_BY_NC_SA.pickle',
+                                     repo_type='dataset'
+                                    )
 
 midi_gas_ps = TMIDIX.Tegridy_Any_Pickle_File_Reader(midi_gas_ps_pickle)
 
@@ -86,10 +87,11 @@ print('=' * 70)
 print('Loading MIDI GAS processed scores embeddings dataset...')
 
 midi_gas_pse_pickle = hf_hub_download(repo_id='asigalov61/MIDI-GAS', 
-                                               filename='MIDI_GAS_Processed_Scores_Embeddings_CC_BY_NC_SA.pickle'
-                                              )
+                                      filename='MIDI_GAS_Processed_Scores_Embeddings_CC_BY_NC_SA.pickle',
+                                      repo_type='dataset'
+                                     )
 
-midi_gas_pse = TMIDIX.Tegridy_Any_Pickle_File_Reader(midi_gas_pse_pickle)
+midi_gas_pse = [a[3] for a in TMIDIX.Tegridy_Any_Pickle_File_Reader(midi_gas_pse_pickle)]
 
 print('=' * 70)
 print('Done!')
@@ -139,7 +141,7 @@ print('=' * 70)
 
 def load_midi(input_midi):
 
-    raw_score = TMIDIX.midi2single_track_ms_score(midi_file)
+    raw_score = TMIDIX.midi2single_track_ms_score(input_midi)
     
     escore = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
     
@@ -213,6 +215,38 @@ def get_embeddings(inputs):
 
 #==================================================================================
 
+def cosine_similarity_numpy(src_array, trg_array):
+   
+    src_norm = np.linalg.norm(src_array)
+
+    trg_norms = np.linalg.norm(trg_array, axis=1)
+
+    dot_products = np.dot(trg_array, src_array)
+
+    cosine_sims = dot_products / (src_norm * trg_norms + 1e-10)
+    
+    return cosine_sims.tolist()
+
+#==================================================================================
+
+def select_best_output(outputs, embeddings, src_embeddings, top_k=10):
+
+    emb_sims = cosine_similarity_numpy(np.array(src_embeddings), np.array(embeddings))
+
+    sorted_emb_sims = sorted(emb_sims, reverse=True)
+
+    hits = []
+    hits_idxs = []
+    
+    for s in sorted_emb_sims[:top_k]:
+        idx = emb_sims.index(s)
+        hits_idxs.append(idx)
+        hits.extend([[str(s)] + outputs[idx][:3]])
+        
+    return hits, hits_idxs
+
+#==================================================================================
+
 # @spaces.GPU
 def Classify_MIDI_Genre(input_midi,
                         input_melody,
@@ -224,10 +258,6 @@ def Classify_MIDI_Genre(input_midi,
 
     #===============================================================================
 
-
-
-    #===============================================================================
-
     print('=' * 70)
     print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
     start_time = reqtime.time()
@@ -250,52 +280,33 @@ def Classify_MIDI_Genre(input_midi,
    
     print('=' * 70)
 
-    #==================================================================
-
-    print('Prepping melody...')
+    #===============================================================================
 
-    if input_midi:
-        inp_mel = 'Custom MIDI'
-        score, score_list = load_midi(input_midi.name, melody_patch, use_nth_note)
+    print('Loadin and prepping source MIDI...')
 
-    else:
-        mel_list = [m[0].lower() for m in popular_hook_melodies]
+    src_score = load_midi(input_midi)
 
-        inp_mel = random.choice(mel_list).title()
+    inp = torch.LongTensor([src_score]).to(device_type)
 
-        for m in mel_list:
-            if input_melody.lower().strip() in m:
-                inp_mel = m.title()
-                break
+    src_emb = get_embeddings(inp).tolist()
 
-        score = popular_hook_melodies[[m[0] for m in popular_hook_melodies].index(inp_mel)][1]
-        score_list = [[[score[i]], score[i+1:i+3]] for i in range(0, len(score)-3, 3)]
+    print('Done!')
 
-        print('Selected melody:', inp_mel)
+    #===============================================================================
         
-    print('Sample score events', score[:12])
+    print('Sample embeddings values', src_emb[:3])
 
-    #==================================================================
+    #===============================================================================
     
     print('=' * 70)
-    print('Generating...')
+    print('Classifying...')
 
     model.to(device_type)
     model.eval()
 
-    #==================================================================
-
-    start_score_seq = [1792] + score + [1793]
-
-    #==================================================================
+    #===============================================================================
 
-    input_seq = generate_full_seq(start_score_seq,
-                                  max_toks=MAX_GEN_TOKS,
-                                  temperature=model_temperature,
-                                  top_k_value=model_sampling_top_k,
-                                 )
 
-    final_song = input_seq[len(start_score_seq):]
    
     print('=' * 70)
     print('Done!')
@@ -386,7 +397,7 @@ def Classify_MIDI_Genre(input_midi,
     print('Done!')
     print('=' * 70)
 
-    #========================================================
+    #===============================================================================
 
     output_title = str(inp_mel)
     output_midi = str(new_fn)
@@ -398,7 +409,7 @@ def Classify_MIDI_Genre(input_midi,
     print('Output MIDI melody title:', output_title)
     print('=' * 70) 
     
-    #========================================================
+    #===============================================================================
     
     print('-' * 70)
     print('Req end time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))