Add: Test for ONNX inference code (without PyTorch dependency)

convert_bert_onnx.py

@@ -1,4 +1,4 @@
-# usage: .venv/bin/python convert_bert_onnx.py --language JP
+# Usage: .venv/bin/python convert_bert_onnx.py --language JP
 # ref: https://github.com/tuna2134/sbv2-api/blob/main/convert/convert_deberta.py
 
 import time

convert_onnx.py

@@ -1,5 +1,5 @@
-# usage: .venv/bin/python convert_onnx.py --model model_assets/koharune-ami/koharune-ami.safetensors
+# Usage: .venv/bin/python convert_onnx.py --model model_assets/koharune-ami/koharune-ami.safetensors
-# usage: .venv/bin/python convert_onnx.py --model model_assets/ (All models in the directory will be converted)
+# Usage: .venv/bin/python convert_onnx.py --model model_assets/ (All models in the directory will be converted)
 # ref: https://github.com/tuna2134/sbv2-api/blob/main/convert/convert_model.py
 
 import time

pyproject.toml

@@ -30,7 +30,7 @@ dependencies = [
     "nltk<=3.8.1",
     "num2words",
     "numba",
-    "numpy",
+    "numpy<2",
     "onnxruntime",
     "pydantic>=2.0",
     "pyopenjtalk-dict",
@@ -71,14 +71,17 @@ exclude = [".git", ".gitignore", ".gitattributes"]
 [tool.hatch.build.targets.wheel]
 packages = ["style_bert_vits2"]
 
+# for PyTorch inference
 [tool.hatch.envs.test]
 dependencies = ["coverage[toml]>=6.5", "pytest", "scipy"]
 features = ["torch"]
 [tool.hatch.envs.test.scripts]
 # Usage: `hatch run test:test`
-test = "pytest {args:tests}"
+test = "pytest tests/test_main.py::test_synthesize_cpu"
+# Usage: `hatch run test:test-cuda`
+test-cuda = "pytest tests/test_main.py::test_synthesize_cuda"
 # Usage: `hatch run test:coverage`
-test-cov = "coverage run -m pytest {args:tests}"
+test-cov = "coverage run -m pytest tests/test_main.py::test_synthesize_cpu"
 # Usage: `hatch run test:cov-report`
 cov-report = ["- coverage combine", "coverage report"]
 # Usage: `hatch run test:cov`
@@ -86,6 +89,23 @@ cov = ["test-cov", "cov-report"]
 [[tool.hatch.envs.test.matrix]]
 python = ["3.9", "3.10", "3.11"]
 
+# for ONNX inference (without PyTorch dependency)
+[tool.hatch.envs.test-onnx]
+dependencies = ["coverage[toml]>=6.5", "pytest", "scipy", "onnxruntime-gpu"]
+[tool.hatch.envs.test-onnx.scripts]
+# Usage: `hatch run test-onnx:test`
+test = "pytest tests/test_main.py::test_synthesize_onnx_cpu"
+# Usage: `hatch run test-onnx:test-cuda`
+test-cuda = "pytest tests/test_main.py::test_synthesize_onnx_cuda"
+# Usage: `hatch run test-onnx:coverage`
+test-cov = "coverage run -m pytest tests/test_main.py::test_synthesize_onnx_cpu"
+# Usage: `hatch run test-onnx:cov-report`
+cov-report = ["- coverage combine", "coverage report"]
+# Usage: `hatch run test-onnx:cov`
+cov = ["test-cov", "cov-report"]
+[[tool.hatch.envs.test-onnx.matrix]]
+python = ["3.9", "3.10", "3.11"]
+
 [tool.hatch.envs.style]
 detached = true
 dependencies = ["black[jupyter]", "isort"]

style_bert_vits2/nlp/bert_models.py

@@ -8,11 +8,12 @@ Style-Bert-VITS2 の学習・推論に必要な各言語ごとの BERT モデル
 一度 load_model/tokenizer() で当該言語の BERT モデルがロードされていれば、ライブラリ内部のどこからでもロード済みのモデル/トークナイザーを取得できる。
 """
 
+from __future__ import annotations
+
 import gc
 import time
-from typing import Optional, Union, cast
+from typing import TYPE_CHECKING, Optional, Union, cast
 
-import torch
 from transformers import (
     AutoModelForMaskedLM,
     AutoTokenizer,
@@ -27,6 +28,10 @@ from style_bert_vits2.constants import DEFAULT_BERT_MODEL_PATHS, Languages
 from style_bert_vits2.logging import logger
 
 
+if TYPE_CHECKING:
+    import torch
+
+
 # 各言語ごとのロード済みの BERT モデルを格納する辞書
 __loaded_models: dict[Languages, Union[PreTrainedModel, DebertaV2Model]] = {}
 
@@ -208,6 +213,8 @@ def unload_model(language: Languages) -> None:
         language (Languages): アンロードする BERT モデルの言語
     """
 
+    import torch
+
     if language in __loaded_models:
         del __loaded_models[language]
         gc.collect()
@@ -224,6 +231,8 @@ def unload_tokenizer(language: Languages) -> None:
         language (Languages): アンロードする BERT トークナイザーの言語
     """
 
+    import torch
+
     if language in __loaded_tokenizers:
         del __loaded_tokenizers[language]
         gc.collect()

style_bert_vits2/tts_model.py

@@ -154,6 +154,16 @@ class TTSModel:
                 f"Model loaded successfully from {self.model_path} to {self.onnx_session.get_providers()[0]} ({time.time() - start_time:.2f}s)"
             )
 
+    def unload(self) -> None:
+        """
+        音声合成モデルをデバイスからアンロードする。
+        """
+
+        if self.net_g is not None:
+            self.net_g = None
+        if self.onnx_session is not None:
+            self.onnx_session = None
+
     def get_style_vector(self, style_id: int, weight: float = 1.0) -> NDArray[Any]:
         """
         スタイルベクトルを取得する。

tests/test_main.py

@@ -1,3 +1,5 @@
+from typing import Any, Literal, Sequence, Union
+
 import pytest
 from scipy.io import wavfile
 
@@ -6,54 +8,95 @@ from style_bert_vits2.tts_model import TTSModelHolder
 
 
 def synthesize(
+    inference_type: Literal["torch", "onnx"] = "torch",
     device: str = "cpu",
-    onnx_providers: list[str] = ["CPUExecutionProvider"],
+    onnx_providers: Sequence[Union[str, tuple[str, dict[str, Any]]]] = [
+        "CPUExecutionProvider"
+    ],
 ):
 
     # 音声合成モデルが配置されていれば、音声合成を実行
     model_holder = TTSModelHolder(BASE_DIR / "model_assets", device, onnx_providers)
     if len(model_holder.models_info) > 0:
 
-        # jvnv-F2-jp モデルを探す
+        # "koharune-ami" または "amitaro" モデルを探す
         for model_info in model_holder.models_info:
-            if model_info.name == "jvnv-F2-jp":
+            if model_info.name == "koharune-ami" or model_info.name == "amitaro":
+
+                # Safetensors 形式または ONNX 形式のモデルファイルに絞り込む
+                if inference_type == "torch":
+                    model_files = [
+                        f
+                        for f in model_info.files
+                        if f.endswith(".safetensors") and not f.startswith(".")
+                    ]
+                else:
+                    model_files = [
+                        f
+                        for f in model_info.files
+                        if f.endswith(".onnx") and not f.startswith(".")
+                    ]
+                if len(model_files) == 0:
+                    pytest.skip(f"音声合成モデル \"{model_info.name}\" のモデルファイルが見つかりませんでした。")
+
+                # モデルをロード
+                model = model_holder.get_model(model_info.name, model_files[0])
+                model.load()
+
                 # すべてのスタイルに対して音声合成を実行
                 for style in model_info.styles:
 
                     # 音声合成を実行
-                    model = model_holder.get_model(model_info.name, model_info.files[0])
-                    model.load()
                     sample_rate, audio_data = model.infer(
                         "あらゆる現実を、すべて自分のほうへねじ曲げたのだ。",
                         # 言語 (JP, EN, ZH / JP-Extra モデルの場合は JP のみ)
                         language=Languages.JP,
                         # 話者 ID (音声合成モデルに複数の話者が含まれる場合のみ必須、単一話者のみの場合は 0)
                         speaker_id=0,
-                        # 感情表現の強さ (0.0 〜 1.0)
+                        # テンポの緩急 (0.0 〜 1.0)
                         sdp_ratio=0.4,
                         # スタイル (Neutral, Happy など)
                         style=style,
                         # スタイルの強さ (0.0 〜 100.0)
-                        style_weight=6.0,
+                        style_weight=2.0,
                     )
 
                     # 音声データを保存
-                    (BASE_DIR / "tests/wavs").mkdir(exist_ok=True, parents=True)
+                    (BASE_DIR / f"tests/wavs/{model_info.name}").mkdir(
-                    wav_file_path = BASE_DIR / f"tests/wavs/{style}.wav"
+                        exist_ok=True, parents=True
+                    )
+                    wav_file_path = (
+                        BASE_DIR / f"tests/wavs/{model_info.name}/{style}.wav"
+                    )
                     with open(wav_file_path, "wb") as f:
                         wavfile.write(f, sample_rate, audio_data)
 
                     # 音声データが保存されたことを確認
                     assert wav_file_path.exists()
-                    # wav_file_path.unlink()
+
+                # モデルをアンロード
+                model.unload()
     else:
         pytest.skip("音声合成モデルが見つかりませんでした。")
 
 
 def test_synthesize_cpu():
-    synthesize(device="cpu")
+    synthesize(inference_type="torch", device="cpu")
 
 
-# Windows環境ではtorchのcudaが簡単に入らないため、テストをスキップ
+def test_synthesize_cuda():
-# def test_synthesize_cuda():
+    pytest.importorskip("torch.cuda")
-#     synthesize(device="cuda")
+    synthesize(inference_type="torch", device="cuda")
+
+
+def test_synthesize_onnx_cpu():
+    synthesize(inference_type="onnx", onnx_providers=["CPUExecutionProvider"])
+
+
+def test_synthesize_onnx_cuda():
+    synthesize(
+        inference_type="onnx",
+        onnx_providers=[
+            ("CUDAExecutionProvider", {"cudnn_conv_algo_search": "DEFAULT"})
+        ],
+    )