Make more VRAM friendly

config/LibriTTS/ar.yml

@@ -3,3 +3,4 @@ spkr_name_getter: "lambda p: p.parts[-3]"
 
 model: ar-quarter
 batch_size: 8
+eval_batch_size: 8

config/LibriTTS/nar.yml

@@ -3,3 +3,4 @@ spkr_name_getter: "lambda p: p.parts[-3]"
 
 model: nar-quarter
 batch_size: 8
+eval_batch_size: 8

scripts/run.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+
+until $@; do echo retrying; done

vall_e/__main__.py

@@ -0,0 +1,14 @@
+import argparse
+from pathlib import Path
+
+
+def main():
+    parser = argparse.ArgumentParser("VALL-E TTS")
+    parser.add_argument("text")
+    parser.add_argument("output")
+    parser.add_argument("--reference", type=Path)
+    args = parser.parse_args()
+
+
+if __name__ == "__main__":
+    main()

vall_e/config.py

@@ -17,7 +17,8 @@ class Config(ConfigBase):
     def sample_rate(self):
         return 24_000
 
-    p_additional_prompt: float = 0.5
+    p_additional_prompt: float = 0.8
+    max_prompts: int = 3
 
     token_dim: int = 256
     num_tokens: int = 1024
@@ -37,6 +38,9 @@ class Config(ConfigBase):
     model: str = "ar-quarter"
     spkr_name_getter: str = "lambda p: p.parts[-2]"
 
+    min_phones: int = 10
+    max_phones: int = 50
+
     @cached_property
     def get_spkr(self):
         return eval(self.spkr_name_getter)

vall_e/data.py

@@ -81,8 +81,8 @@ class VALLEDatset(Dataset):
         paths,
         phone_symmap=None,
         spkr_symmap=None,
-        min_phones=10,
+        min_phones=cfg.min_phones,
-        max_phones=100,
+        max_phones=cfg.max_phones,
         training=False,
         extra_paths_by_spkr_name: dict[str, list] = {},
     ):
@@ -141,7 +141,7 @@ class VALLEDatset(Dataset):
             )
             choices = [better_not]
 
-        for _ in range(10):
+        for _ in range(cfg.max_prompts):
             path = random.choice(choices)
             prom_list.append(_load_quants(path))
             if random.random() > cfg.p_additional_prompt: