looks like it's working?

conf/lora/ella-baila-sola.yaml

@@ -0,0 +1,10 @@
+$include:
+  - conf/lora.yml
+
+fine_tune: True
+
+train/AudioLoader.sources:
+  - /media/CHONK/hugo/spotdl/subsets/this-is-charlie-parker/Eslabon Armado - Ella Baila Sola.mp3
+
+val/AudioLoader.sources:
+  - /media/CHONK/hugo/spotdl/subsets/this-is-charlie-parker/Eslabon Armado - Ella Baila Sola.mp3

conf/lora/lora-is-this-charlie-parker.yaml

@@ -1,4 +1,10 @@
 $include:
-  - conf/vampnet.yml
+  - conf/lora.yml
 
-fine_tune: True
+fine_tune: True
+
+train/AudioLoader.sources:
+  - /media/CHONK/hugo/spotdl/subsets/this-is-charlie-parker/Charlie Parker - Yardbird Suite.mp3
+
+val/AudioLoader.sources:
+  - /media/CHONK/hugo/spotdl/subsets/this-is-charlie-parker/Charlie Parker - Yardbird Suite.mp3

conf/lora/lora.yaml

@@ -0,0 +1,16 @@
+$include:
+  - conf/vampnet.yml
+
+fine_tune: True
+
+train/AudioDataset.n_examples: 10000000
+
+val/AudioDataset.n_examples: 10
+
+
+NoamScheduler.warmup: 250
+
+epoch_length: 100
+save_audio_epochs: 2
+
+AdamW.lr: 0.0001

conf/vampnet.yml

@@ -1,5 +1,5 @@
 
-codec_ckpt: /u/home/src/runs/codec-ckpt/lac-768x.pth
+codec_ckpt: /home/hugo/descript/vampnet/models/spotdl/codec.pth
 save_path: ckpt
 max_epochs: 1000
 epoch_length: 1000
@@ -11,8 +11,8 @@ suffix_amt: 0.0
 prefix_dropout: 0.1
 suffix_dropout: 0.1
 
-batch_size: 120
-num_workers: 80
+batch_size: 8
+num_workers: 10
 
 # Optimization
 detect_anomaly: false

env/setup.py

@@ -11,36 +11,7 @@ def run(cmd):
     return subprocess.check_output(shlex.split(cmd)).decode("utf-8")
 
 
-print("1. Setting up Google Cloud access")
-print("---------------------------------")
-gcloud_authorized = "gs://research-data-raw" in run("gsutil ls")
-if not gcloud_authorized:
-    run("gcloud auth login")
 
-run("gcloud config set project lyrebird-research")
-run("gcloud auth configure-docker")
-
-print()
-print("2. Setting up Github access")
-print("---------------------------")
-
-lines = textwrap.wrap(
-    "First, let's get your Github token, so all "
-    "packages can be installed. Create one by going to your "
-    "Github profile -> Developer settings -> Personal access tokens -> "
-    "Generate new token. Copy the token below."
-)
-[print(l) for l in lines]
-
-GITHUB_TOKEN = input("\nGithub token: ") or "undefined"
-
-print()
-print("3. Setting up Jupyter and Tensorboard")
-print("-------------------------------------")
-
-JUPYTER_TOKEN = input("Password for Jupyter server (default:password): ") or "password"
-JUPYTER_PORT = input("Jupyter port to run on (default:8888): ") or "8888"
-TENSORBOARD_PORT = input("Tensorboard port to run on (default:6006): ") or "6006"
 
 print()
 print("4. Setting up paths.")

scripts/exp/train.py

@@ -248,12 +248,12 @@ def train(
     save_path: str = "ckpt",
     max_epochs: int = int(100e3),
     epoch_length: int = 1000,
-    save_audio_epochs: int = 10,
+    save_audio_epochs: int = 2,
     save_epochs: list = [10, 50, 100, 200, 300, 400,],
     batch_size: int = 48,
     grad_acc_steps: int = 1,
-    val_idx: list = [0, 1, 2, 3, 4],
-    num_workers: int = 20,
+    val_idx: list = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
+    num_workers: int = 10,
     detect_anomaly: bool = False,
     grad_clip_val: float = 5.0,
     prefix_amt: float = 0.0,
@@ -530,7 +530,7 @@ def train(
 
                 accel.unwrap(model).metadata = metadata
                 accel.unwrap(model).save_to_folder(
-                    f"{save_path}/{tag}", model_extra
+                    f"{save_path}/{tag}", model_extra,
                 )
 
         def save_sampled(self, z):

scripts/utils/split.py

@@ -0,0 +1,51 @@
+from pathlib import Path
+import random
+import shutil
+
+import argbind
+
+from audiotools.core import util
+
+
+@argbind.bind(without_prefix=True)
+def train_test_split(
+    audio_folder: str = ".", 
+    test_size: float = 0.2,
+    seed: int = 42,
+):
+    audio_files = util.find_audio(audio_folder)
+    
+    # split according to test_size
+    n_test = int(len(audio_files) * test_size)
+    n_train = len(audio_files) - n_test
+
+    # shuffle
+    random.seed(seed)
+    random.shuffle(audio_files)
+
+    train_files = audio_files[:n_train]
+    test_files = audio_files[n_train:]
+
+
+    print(f"Train files: {len(train_files)}")
+    print(f"Test files: {len(test_files)}")
+    continue_ = input("Continue [yn]? ") or "n"
+
+    if continue_ != "y":
+        return
+    
+    for split, files in (
+        ("train", train_files), ("test", test_files)
+    ):
+        for file in files:
+            out_file = Path(file).parent / split / Path(file).name
+            out_file.parent.mkdir(exist_ok=True, parents=True)
+            shutil.copy(file, out_file)
+    
+
+    
+if __name__ == "__main__":
+    args  = argbind.parse_args()
+
+    with argbind.scope(args):
+        train_test_split()

vampnet/modules/__init__.py

@@ -1,4 +1,4 @@
 import audiotools
 
 audiotools.ml.BaseModel.INTERN += ["vampnet.modules.**"]
-audiotools.ml.BaseModel.EXTERN += ["einops", "flash_attn.flash_attention"]
+audiotools.ml.BaseModel.EXTERN += ["einops", "flash_attn.flash_attention", "loralib"]

vampnet/modules/transformer.py

@@ -14,7 +14,7 @@ from .layers import FiLM
 from .layers import SequentialWithFiLM
 from .layers import WNConv1d
 
-LORA_R = 4
+LORA_R = 8
 
 
 class RMSNorm(nn.Module):