Explicitly add poetry export plugin; change docker compose example; s…

…peedup docker cache

.github/workflows/build_docker_image.yml

@@ -23,6 +23,9 @@ jobs:
       - name: Install poetry
         run: curl -sSL https://install.python-poetry.org | python3 -
 
+      - name: Install poetry export plugin
+        run: poetry self add poetry-plugin-export
+
       - name: Export requirements
         run: poetry export -o requirements.txt --only main,deploy
 

Dockerfile

@@ -5,7 +5,7 @@ WORKDIR /usr/src/app
 RUN apt-get update && apt-get install -y ffmpeg
 
 COPY requirements.txt ./
-COPY code ./
 RUN pip install --no-cache-dir -r requirements.txt
 
+COPY code ./
 CMD [ "python", "./app.py" ]

README.md

@@ -11,6 +11,9 @@ Most voices perform best when narrating medium-length sentences with medium-leng
 
 ![This paragraph is a good example of appropriate text input.](/examples/1708512151_ME2_f-player_f-Shepard.ogg)
 
+## Deployment
+GitHub Actions automatically produces a fresh image on every push to the `main` branch. See `docker-compose.example.yml` on how it can be deployed.
+
 ## History (and other experiments)
 I initially [fine-tuned SpeechT5](https://huggingface.co/learn/audio-course/chapter6/fine-tuning), but the results were disappointing. That model very frequently produced garbage and/or hallucinated output for most voices. Interestingly, it also had a very strong bias towards female speakers. 
 

code/app.py

@@ -1,16 +1,14 @@
 """Gradio web interface for ShepardTTS."""
 
-import os
 import time
 from pathlib import Path
 
 import gradio as gr
 import numpy as np
-import torch
-from torchaudio.io import CodecConfig, StreamWriter
-
 import settings
+import torch
 from app_helpers import description, examples, links
+from torchaudio.io import CodecConfig, StreamWriter
 from utils import load_checkpoint, normalize_line
 
 model = load_checkpoint()
@@ -83,7 +81,7 @@ def predict(
     for sentence in out["wav"]:
         waveform = torch.cat((waveform, QUARTER_SECOND_PAUSE, sentence, QUARTER_SECOND_PAUSE))
 
-    base_filename = f"{os.environ['GRADIO_EXAMPLES_CACHE']}/{int(time.time())}_{character}"
+    base_filename = f"{settings.GRADIO_EXAMPLES_CACHE}/{int(time.time())}_{character}"
 
     if codec_format == "mp3":
         # Write compressed mp3
@@ -231,4 +229,4 @@ def predict(
     )
 
 demo.queue(max_size=10)
-demo.launch(debug=False, show_api=True, share=False, auth=("shepard", os.environ["SECRET_KEY"]))
+demo.launch(debug=False, show_api=True, share=False, auth=("shepard", settings.SECRET_KEY))

code/create_character_embeddings.py

@@ -3,7 +3,6 @@
 from pathlib import Path
 
 import torch
-
 from utils import load_checkpoint
 
 

code/create_dataset.py

@@ -5,7 +5,6 @@
 import pandas as pd
 import soundfile
 from datasets import Audio, Dataset, concatenate_datasets
-
 from utils import normalize_line
 
 

code/inference.py

@@ -5,7 +5,6 @@
 import numpy as np
 import torch
 from torchaudio.io import StreamWriter
-
 from utils import load_checkpoint, normalize_line
 
 model = load_checkpoint()

code/settings.py

@@ -1,8 +1,12 @@
 """Contains some universal settings."""
+import os
 
 import torch
 
-CHECKPOINTS_CONFIG_JSON = "./current_model/config.json"
-CHECKPOINT_DIR = "./current_model"
-CHECKPOINT_VOCAB = "./current_model/vocab.json"
+CHECKPOINTS_CONFIG_JSON = "/xtts_model/config.json"
+CHECKPOINT_DIR = "/xtts_model/current_model"
+CHECKPOINT_VOCAB = "/xtts_model/current_model/vocab.json"
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+GRADIO_EXAMPLES_CACHE = os.environ["GRADIO_EXAMPLES_CACHE"]
+SECRET_KEY = os.environ["SECRET_KEY"]

code/utils.py

@@ -2,12 +2,11 @@
 
 import re
 
+import settings
 from cleantext import clean
 from num2words import num2words
-from TTS.tts.configs.xtts_config import XttsConfig
-
-import settings
 from overrides import ShepardXtts
+from TTS.tts.configs.xtts_config import XttsConfig
 
 
 def load_checkpoint():

docker-compose.example.yml

@@ -2,11 +2,11 @@ version: '3.8'
 
 services:
   gradio:
-    build: .
+    image: "ghcr.io/darwinkel/shepardtts:main"
     ports:
       - "1337:1337"
     volumes:
-      - .:/usr/src/app
+      - ./current_model/:/xtts_model/:ro
     environment:
       GRADIO_EXAMPLES_CACHE: "/tmp/"
       GRADIO_ANALYTICS_ENABLED: False