Small fixes for finetuning

aria/datasets.py

@@ -1060,7 +1060,7 @@ def _get_onset_adjusted_msg(
 
         return _temp_note_msg
 
-    _note_msgs = midi_dict.note_msgs
+    _note_msgs = copy.deepcopy(midi_dict.note_msgs)
 
     # Remove notes
     if random.random() < config["remove_notes"]["activation_prob"]:
@@ -1270,10 +1270,6 @@ def _build_epoch(_save_path, _midi_dataset):
                     if _idx % 250 == 0:
                         logger.info(f"Finished processing {_idx}")
 
-                    # DEBUG
-                    if _idx == 1000:
-                        break
-
         logger = setup_logger()
         assert max_seq_len > 0, "max_seq_len must be greater than 0"
         assert num_epochs > 0, "num_epochs must be greater than 0"

aria/run.py

@@ -152,17 +152,20 @@ def sample(args):
         guidance_midi_dict=guidance_midi_dict,
     )
 
-    if guidance_seq:
-        tokenizer.detokenize(guidance_seq).to_midi().save(
-            os.path.join(samples_dir, f"guidance.mid")
-        )
     if len(prompt_seq) + args.l > model_config.max_seq_len:
         print(
             "WARNING: Required context exceeds max_seq_len supported by model"
         )
     prompts = [prompt_seq for _ in range(num_variations)]
 
-    if args.cfg is not None:
+    samples_dir = os.path.join(os.path.dirname(__file__), "..", "samples")
+    if os.path.isdir(samples_dir) is False:
+        os.mkdir(samples_dir)
+    if guidance_seq:
+        tokenizer.detokenize(guidance_seq).to_midi().save(
+            os.path.join(samples_dir, f"guidance.mid")
+        )
+    if args.cfg is not None and guidance_seq is not None:
         results = sample_batch_cfg(
             model=model,
             tokenizer=tokenizer,
@@ -186,10 +189,6 @@ def sample(args):
             compile=args.compile,
         )
 
-    samples_dir = os.path.join(os.path.dirname(__file__), "..", "samples")
-    if os.path.isdir(samples_dir) is False:
-        os.mkdir(samples_dir)
-
     for idx, tokenized_seq in enumerate(results):
         res_midi_dict = tokenizer.detokenize(tokenized_seq)
         res_midi = res_midi_dict.to_midi()

aria/sample.py

@@ -303,6 +303,7 @@ def sample_batch_cfg(
         logits_cfg[:, tokenizer.tok_to_id[tokenizer.prompt_start_tok]] = float(
             "-inf"
         )
+        logits_cfg[:, tokenizer.tok_to_id[tokenizer.dim_tok]] = float("-inf")
 
         if temperature > 0.0:
             probs = torch.softmax(logits_cfg / temperature, dim=-1)
@@ -389,7 +390,10 @@ def get_inference_prompt(
     if tokenizer.dim_tok in prompt_seq:
         prompt_seq.remove(tokenizer.dim_tok)
 
-    if guidance_midi_dict is not None:
+    if (
+        guidance_midi_dict is not None
+        and tokenizer.guidance_start_tok in prompt_seq
+    ):
         guidance_seq = copy.deepcopy(prompt_seq)
         guidance_seq = guidance_seq[
             : guidance_seq.index(tokenizer.guidance_end_tok)

aria/tokenizer.py

@@ -174,8 +174,8 @@ def detokenize(self, tokenized_seq: list, **kwargs):
 
     def export_data_aug(self):
         return [
-            self.export_guidance_tempo_aug(max_tempo_aug=0.25, mixup=True),
-            self.export_guidance_pitch_aug(4),
+            self.export_guidance_tempo_aug(max_tempo_aug=0.2, mixup=True),
+            self.export_guidance_pitch_aug(3),
             self.export_guidance_velocity_aug(2),
         ]
 

aria/train.py

@@ -61,9 +61,6 @@
 #   -bs 32 \
 #   -workers 8
 
-# TODO:
-# - Test that everything works on a distributed setup
-
 
 def setup_logger(project_dir: str):
     # Get logger and reset all handlers
@@ -196,7 +193,7 @@ def get_optim(
     num_epochs: int,
     steps_per_epoch: int,
 ):
-    LR = 3e-5
+    LR = 3e-4
     END_RATIO = 0.1
     WARMUP_STEPS = 200
 

config/config.json

@@ -182,31 +182,29 @@
             "min_clean_interval_ms": 60000,
             "max_clean_interval_ms": 200000,
             "noising": {
-                "activation_prob": 0.8,
+                "activation_prob": 0.5,
                 "remove_notes": {
-                    "activation_prob": 0.5,
+                    "activation_prob": 0.25,
                     "min_ratio": 0.0,
-                    "max_ratio": 0.3
+                    "max_ratio": 0.15
                 },
                 "adjust_velocity": {
-                    "activation_prob": 0.3,
+                    "activation_prob": 0.25,
                     "min_adjust": 1,
-                    "max_adjust": 30,
-                    "max_ratio": 0.1,
-                    "min_ratio": 0.30
+                    "max_adjust": 20
                 },
                 "adjust_onsets": {
                     "activation_prob": 0.25,
-                    "min_adjust_s": 0.01,
-                    "max_adjust_s": 0.07,
-                    "max_ratio": 0.15,
-                    "min_ratio": 0.3
+                    "min_adjust_s": 0.005,
+                    "max_adjust_s": 0.05,
+                    "max_ratio": 0.0,
+                    "min_ratio": 0.2
                 },
                 "quantize_onsets": {
-                    "activation_prob": 0.15,
+                    "activation_prob": 0.05,
                     "min_quant_s": 0.05,
-                    "max_quant_s": 0.15,
-                    "max_vel_delta": 45
+                    "max_quant_s": 0.1,
+                    "max_vel_delta": 30
                 }
             }
         }
@@ -215,7 +213,7 @@
         "inference_abs": {
             "guidance": {
                 "min_ms": 5000,
-                "max_ms": 30000
+                "max_ms": 40000
             }