Small fix (#13)

* add more aug

* add multi gpu inference

* small fix

* format

Author: loubbrad

amt/data.py

@@ -56,7 +56,11 @@ def get_wav_mid_segments(
     # Create features
     total_samples = wav.shape[-1]
     res = []
-    for idx in range(0, total_samples, num_samples // stride_factor):
+    for idx in range(
+        0,
+        total_samples - (num_samples - (num_samples // stride_factor)),
+        num_samples // stride_factor,
+    ):
         audio_feature = pad_or_trim(wav[idx:], length=num_samples)
         if midi_dict is not None:
             mid_feature = tokenizer._tokenize_midi_dict(

amt/infer.py

@@ -5,15 +5,12 @@
 import torch.multiprocessing as multiprocessing
 
 from torch.multiprocessing import Queue
-from torch.cuda import device_count, is_available
 from tqdm import tqdm
 
 from amt.model import AmtEncoderDecoder
 from amt.tokenizer import AmtTokenizer
 from amt.audio import AudioTransform
 from amt.data import get_wav_mid_segments
-from amt.config import load_config
-from aria.data.midi import MidiDict
 
 MAX_SEQ_LEN = 4096
 LEN_MS = 30000
@@ -24,8 +21,11 @@
 VEL_TOLERANCE = 50
 
 
+# TODO: Profile and fix gpu util
+
+
 def calculate_vel(
-    logits: torch.tensor,
+    logits: torch.Tensor,
     init_vel: int,
     tokenizer: AmtTokenizer = AmtTokenizer(),
 ):
@@ -51,13 +51,13 @@ def calculate_vel(
 
     vels = torch.tensor(vels).to(probs.device)
     new_vel = torch.sum(vels * probs) / torch.sum(probs)
-    new_vel = round(new_vel.item() / 10) * 10
+    new_vel = round(new_vel.item() / 5) * 5
 
     return tokenizer.tok_to_id[("vel", new_vel)]
 
 
 def calculate_onset(
-    logits: torch.tensor,
+    logits: torch.Tensor,
     init_onset: int,
     tokenizer: AmtTokenizer = AmtTokenizer(),
 ):
@@ -88,6 +88,7 @@ def calculate_onset(
     return tokenizer.tok_to_id[("onset", new_onset)]
 
 
+@torch.autocast("cuda", dtype=torch.bfloat16)
 def process_segments(
     tasks: list,
     model: AmtEncoderDecoder,
@@ -111,14 +112,14 @@ def process_segments(
 
     kv_cache = model.get_empty_cache()
 
-    # for idx in (
-    #     pbar := tqdm(
-    #         range(min_prefix_len, MAX_SEQ_LEN - 1),
-    #         total=MAX_SEQ_LEN - (min_prefix_len + 1),
-    #         leave=False,
-    #     )
-    # ):
-    for idx in range(min_prefix_len, MAX_SEQ_LEN - 1):
+    for idx in (
+        pbar := tqdm(
+            range(min_prefix_len, MAX_SEQ_LEN - 1),
+            total=MAX_SEQ_LEN - (min_prefix_len + 1),
+            leave=False,
+        )
+    ):
+        # for idx in range(min_prefix_len, MAX_SEQ_LEN - 1):
         if idx == min_prefix_len:
             logits = model.decoder(
                 xa=audio_features,
@@ -160,6 +161,12 @@ def process_segments(
         if all(eos_seen):
             break
 
+    if not all(eos_seen):
+        print("WARNING: OVERFLOW")
+        for _idx in range(seq.shape[0]):
+            if eos_seen[_idx] == False:
+                eos_seen[_idx] = MAX_SEQ_LEN
+
     results = [
         tokenizer.decode(seq[_idx, : eos_seen[_idx] + 1])
         for _idx in range(seq.shape[0])
@@ -174,9 +181,7 @@ def gpu_manager(
     model: AmtEncoderDecoder,
     batch_size: int,
 ):
-    model.cuda()
-    model.eval()
-    model.compile()
+    # model.compile()
     audio_transform = AudioTransform().cuda()
     tokenizer = AmtTokenizer(return_tensors=True)
     process_pid = multiprocessing.current_process().pid
@@ -277,9 +282,6 @@ def process_file(
             seq[init_idx : seq.index(tokenizer.eos_tok)],
             idx * CHUNK_LEN_MS,
         )
-        print(
-            f"{process_pid}: Finished {idx+1}/{len(audio_segments)} audio segments"
-        )
 
         if idx == len(audio_segments) - 1:
             break
@@ -310,8 +312,8 @@ def _get_save_path(_file_path: str):
             save_path = os.path.join(
                 save_dir, os.path.splitext(input_rel_path)[0] + ".mid"
             )
-            if not os.path.exists(os.path.dirname(save_path)):
-                os.makedirs(os.path.dirname(save_path))
+            if not os.path.isdir(os.path.dirname(save_path)):
+                os.makedirs(os.path.dirname(save_path), exist_ok=True)
 
         return save_path
 
@@ -361,20 +363,15 @@ def batch_transcribe(
     if gpu_id is not None:
         os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
 
-    model.to("cuda")
+    model.cuda()
+    model.eval()
     file_queue = Queue()
     for file_path in file_paths:
         file_queue.put(file_path)
 
     gpu_task_queue = Queue()
     result_queue = Queue()
 
-    gpu_manager_process = multiprocessing.Process(
-        target=gpu_manager,
-        args=(gpu_task_queue, result_queue, model, batch_size),
-    )
-    gpu_manager_process.start()
-
     worker_processes = [
         multiprocessing.Process(
             target=worker,
@@ -391,6 +388,13 @@ def batch_transcribe(
     for p in worker_processes:
         p.start()
 
+    time.sleep(10)
+    gpu_manager_process = multiprocessing.Process(
+        target=gpu_manager,
+        args=(gpu_task_queue, result_queue, model, batch_size),
+    )
+    gpu_manager_process.start()
+
     for p in worker_processes:
         p.join()
 

amt/run.py

@@ -119,7 +119,9 @@ def transcribe(args):
     assert os.path.isfile(args.cp), "model checkpoint file not found"
     assert args.load_path or args.load_dir, "must give either load path or dir"
     if args.load_path:
-        assert os.path.isfile(args.load_path), "audio file not found"
+        assert os.path.isfile(
+            args.load_path
+        ), f"audio file not found: {args.load_path}"
         trans_mode = "single"
     if args.load_dir:
         assert os.path.isdir(args.load_dir), "load directory doesn't exist"
@@ -201,6 +203,7 @@ def transcribe(args):
             model=model,
             save_dir=args.save_dir,
             batch_size=args.bs,
+            input_dir=args.load_dir,
         )