app.py: add locker

app.py

@@ -8,24 +8,28 @@ import os
 import glob
 import logging
 from typing import cast
+from threading import Lock
 
 import gradio as gr
 from balacoon_tts import TTS
 from huggingface_hub import hf_hub_download, list_repo_files
 
+# locker that disallow access to the tts object from more then one thread
+locker = Lock()
 # global tts module, initialized from a model selected
 tts = None
 # path to the model that is currently used in tts
 cur_model_path = None
 # cache of speakers, maps model name to speaker list
 model_to_speakers = dict()
-model_repo_dir = "data"
+model_repo_dir = "/data"
 for name in list_repo_files(repo_id="balacoon/tts"):
-    hf_hub_download(
-        repo_id="balacoon/tts",
-        filename=name,
-        local_dir=model_repo_dir,
-    )
+    if not os.path.isfile(os.path.join(model_repo_dir, name)):
+        hf_hub_download(
+            repo_id="balacoon/tts",
+            filename=name,
+            local_dir=model_repo_dir,
+        )
 
 
 def main():
@@ -70,13 +74,16 @@ def main():
                 if model_name_str in model_to_speakers:
                     speakers = model_to_speakers[model_name_str]
                 else:
-                    global tts, cur_model_path
-                    # need to load this model to learn the list of speakers
-                    model_path = os.path.join(model_repo_dir, model_name_str)
-                    tts = TTS(model_path)
-                    cur_model_path = model_path
-                    speakers = tts.get_speakers()
-                    model_to_speakers[model_name_str] = speakers
+                    global tts, cur_model_path, model_to_speakers, locker
+                    with locker:
+                        # need to load this model to learn the list of speakers
+                        model_path = os.path.join(model_repo_dir, model_name_str)
+                        if tts is not None:
+                            del tts
+                        tts = TTS(model_path)
+                        cur_model_path = model_path
+                        speakers = tts.get_speakers()
+                        model_to_speakers[model_name_str] = speakers
                 
                 value = speakers[-1]
                 return gr.Dropdown.update(
@@ -101,15 +108,18 @@ def main():
                 logging.info("text, model name or speaker are not provided")
                 return None
             expected_model_path = os.path.join(model_repo_dir, model_name_str)
-            global tts, cur_model_path
-            if expected_model_path != cur_model_path:
-                # reload model
-                tts = TTS(expected_model_path)
-                cur_model_path = expected_model_path
-            if len(text_str) > 1024:
-                # truncate the text
-                text_str = text_str[:1024]
-            samples = tts.synthesize(text_str, speaker_str)
+            global tts, cur_model_path, locker
+            with locker:
+                if expected_model_path != cur_model_path:
+                    # reload model
+                    if tts is not None:
+                        del tts
+                    tts = TTS(expected_model_path)
+                    cur_model_path = expected_model_path
+                if len(text_str) > 1024:
+                    # truncate the text
+                    text_str = text_str[:1024]
+                samples = tts.synthesize(text_str, speaker_str)
             return gr.Audio.update(value=(tts.get_sampling_rate(), samples))
 
         generate.click(synthesize_audio, inputs=[text, model_name, speaker], outputs=audio)