update examples, update upload ckpt scripts

app.py

@@ -24,7 +24,7 @@ DEFAULT_MAX_NEW_TOKENS = 500
 DEFAULT_TEMPERATURE = 0.9
 DEFAULT_TOP_K = 50
 MAX_PROMPT_CHARS = 500
-CHECKPOINT_PATH = ROOT / "checkpoints" / "nanotts_ckpt_030000_inference.pt"
+CHECKPOINT_PATH = ROOT / "checkpoints" / "ckpt_025000_inference.pt"
 TEXT_TOKENIZER_PATH = ROOT / "text_tokenizer" / "libritts_bpe.json"
 WAVTOKENIZER_REPO_ID = "novateur/WavTokenizer"
 WAVTOKENIZER_CONFIG = "wavtokenizer_smalldata_frame40_3s_nq1_code4096_dim512_kmeans200_attn.yaml"
@@ -57,7 +57,7 @@ def load_bundle() -> ModelBundle:
     if not CHECKPOINT_PATH.exists():
         raise FileNotFoundError(
             f"Missing nanoTTS checkpoint at {CHECKPOINT_PATH}. "
-            "Upload checkpoints/nanotts_ckpt_030000_inference.pt to this Space first."
+            "Upload checkpoints/ckpt_025000_inference.pt to this Space first."
         )
 
     device = _device()
@@ -151,7 +151,7 @@ with gr.Blocks(title="nanoTTS") as demo:
                 placeholder="Type a short sentence...",
                 lines=4,
                 max_lines=8,
-                value="Hello again! This is nano text to speech.",
+                value="Hello friend! This is nano text to speech.",
             )
             generate_button = gr.Button("Generate", variant="primary")
         with gr.Column(scale=2):
@@ -183,9 +183,11 @@ with gr.Blocks(title="nanoTTS") as demo:
 
     gr.Examples(
         examples=[
+            ["Nano text to speech", 500, 0.9, 50, 0],
+            ["Created at the University of Maryland", 500, 0.9, 50, 0],
+            ["Trained using only two hundred and forty five hours of speech data", 600, 0.9, 50, 0],
+            ["In the year two thousand and twenty six", 500, 0.9, 50, 0],
             ["The quick brown fox jumps over the lazy dog.", 500, 0.9, 50, 0],
-            ["My name is Pedro Sandoval Segura.", 500, 0.9, 50, 0],
-            ["Nano text to speech was created at the University of Maryland.", 600, 0.9, 50, 0],
         ],
         inputs=[text_input, max_new_tokens, temperature, top_k, seed],
         outputs=audio_output,

scripts/export_inference_checkpoint.py

@@ -5,15 +5,9 @@ from pathlib import Path
 
 import torch
 
-DEFAULT_SOURCE = Path("/fs/nexus-scratch/psando/nanotts-05-10/gpt2/ckpt_030000.pt")
-DEFAULT_OUTPUT = Path("checkpoints/nanotts_ckpt_030000_inference.pt")
-
-
-def to_python_number(value):
-    if isinstance(value, torch.Tensor):
-        return value.detach().cpu().item()
-    return value
-
+DEFAULT_SOURCE = Path("/fs/nexus-scratch/psando/nanotts-05-10/gpt2/ckpt_025000.pt")
+stem = DEFAULT_SOURCE.stem
+DEFAULT_OUTPUT = Path(f"checkpoints/{stem}_inference.pt")
 
 def main() -> None:
     parser = argparse.ArgumentParser(description="Export a slim nanoTTS inference checkpoint.")
@@ -32,8 +26,8 @@ def main() -> None:
         "model": checkpoint["model"],
         "model_args": checkpoint["model_args"],
         "iter_num": checkpoint.get("iter_num"),
-        "train_loss": to_python_number(checkpoint.get("train_loss")),
-        "val_loss": to_python_number(checkpoint.get("val_loss")),
+        "train_loss": checkpoint.get("train_loss"),
+        "val_loss": checkpoint.get("val_loss"),
         "source_checkpoint": str(args.source),
     }
 
@@ -43,6 +37,7 @@ def main() -> None:
     size_mib = args.output.stat().st_size / 1024 / 1024
     print(f"Wrote {args.output} ({size_mib:.1f} MiB)")
     print(f"Keys: {sorted(inference_checkpoint.keys())}")
+    print(f"Val loss: {inference_checkpoint.get('val_loss')}")
 
 
 if __name__ == "__main__":

scripts/upload_checkpoint.py

@@ -8,8 +8,7 @@ from pathlib import Path
 from huggingface_hub import HfApi
 from huggingface_hub.errors import HfHubHTTPError
 
-DEFAULT_PATH = Path("checkpoints/nanotts_ckpt_030000_inference.pt")
-
+DEFAULT_PATH = Path("checkpoints/ckpt_025000_inference.pt")
 
 def main() -> None:
     parser = argparse.ArgumentParser(description="Upload the nanoTTS inference checkpoint to the Space.")