Improve: Make ONNX inference tests more rigorous

style_bert_vits2/nlp/chinese/bert_feature.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any, cast, Optional, Sequence, Union
+from typing import TYPE_CHECKING, Any, Optional, Sequence, Union, cast
 
 import numpy as np
 from numpy.typing import NDArray
@@ -108,9 +108,9 @@ def extract_bert_feature_onnx(
     res = session.run(
         [output_name],
         {
-            "input_ids": cast(NDArray[Any], inputs["input_ids"]).astype(np.int64),
+            "input_ids": inputs["input_ids"].astype(np.int64),  # type: ignore
-            "token_type_ids": cast(NDArray[Any], inputs["token_type_ids"]).astype(np.int64),
+            "token_type_ids": inputs["token_type_ids"].astype(np.int64),  # type: ignore
-            "attention_mask": cast(NDArray[Any], inputs["attention_mask"]).astype(np.int64),
+            "attention_mask": inputs["attention_mask"].astype(np.int64),  # type: ignore
         },
     )[0]
 
@@ -120,9 +120,9 @@ def extract_bert_feature_onnx(
         style_res = session.run(
             [output_name],
             {
-                "input_ids": cast(NDArray[Any], style_inputs["input_ids"]).astype(np.int64),
+                "input_ids": style_inputs["input_ids"].astype(np.int64),  # type: ignore
-                "token_type_ids": cast(NDArray[Any], style_inputs["token_type_ids"]).astype(np.int64),
+                "token_type_ids": style_inputs["token_type_ids"].astype(np.int64),  # type: ignore
-                "attention_mask": cast(NDArray[Any], style_inputs["attention_mask"]).astype(np.int64),
+                "attention_mask": style_inputs["attention_mask"].astype(np.int64),  # type: ignore
             },
         )[0]
         style_res_mean = np.mean(style_res, axis=0)

style_bert_vits2/nlp/english/bert_feature.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any, cast, Optional, Sequence, Union
+from typing import TYPE_CHECKING, Any, Optional, Sequence, Union, cast
 
 import numpy as np
 from numpy.typing import NDArray
@@ -108,8 +108,8 @@ def extract_bert_feature_onnx(
     res = session.run(
         [output_name],
         {
-            "input_ids": cast(NDArray[Any], inputs["input_ids"]).astype(np.int64),
+            "input_ids": inputs["input_ids"].astype(np.int64),  # type: ignore
-            "attention_mask": cast(NDArray[Any], inputs["attention_mask"]).astype(np.int64),
+            "attention_mask": inputs["attention_mask"].astype(np.int64),  # type: ignore
         },
     )[0]
 
@@ -119,8 +119,8 @@ def extract_bert_feature_onnx(
         style_res = session.run(
             [output_name],
             {
-                "input_ids": cast(NDArray[Any], style_inputs["input_ids"]).astype(np.int64),
+                "input_ids": style_inputs["input_ids"].astype(np.int64),  # type: ignore
-                "attention_mask": cast(NDArray[Any], style_inputs["attention_mask"]).astype(np.int64),
+                "attention_mask": style_inputs["attention_mask"].astype(np.int64),  # type: ignore
             },
         )[0]
         style_res_mean = np.mean(style_res, axis=0)

style_bert_vits2/nlp/japanese/bert_feature.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any, cast, Optional, Sequence, Union
+from typing import TYPE_CHECKING, Any, Optional, Sequence, Union, cast
 
 import numpy as np
 from numpy.typing import NDArray
@@ -121,8 +121,8 @@ def extract_bert_feature_onnx(
     res = session.run(
         [output_name],
         {
-            "input_ids": cast(NDArray[Any], inputs["input_ids"]).astype(np.int64),
+            "input_ids": inputs["input_ids"].astype(np.int64),  # type: ignore
-            "attention_mask": cast(NDArray[Any], inputs["attention_mask"]).astype(np.int64),
+            "attention_mask": inputs["attention_mask"].astype(np.int64),  # type: ignore
         },
     )[0]
 
@@ -132,8 +132,8 @@ def extract_bert_feature_onnx(
         style_res = session.run(
             [output_name],
             {
-                "input_ids": cast(NDArray[Any], style_inputs["input_ids"]).astype(np.int64),
+                "input_ids": style_inputs["input_ids"].astype(np.int64),  # type: ignore
-                "attention_mask": cast(NDArray[Any], style_inputs["attention_mask"]).astype(np.int64),
+                "attention_mask": style_inputs["attention_mask"].astype(np.int64),  # type: ignore
             },
         )[0]
         style_res_mean = np.mean(style_res, axis=0)

tests/test_main.py

@@ -1,4 +1,4 @@
-from typing import Any, Literal, Sequence, Union
+from typing import Any, Literal, Sequence
 
 import pytest
 from scipy.io import wavfile
@@ -10,8 +10,8 @@ from style_bert_vits2.tts_model import TTSModelHolder
 def synthesize(
     inference_type: Literal["torch", "onnx"] = "torch",
     device: str = "cpu",
-    onnx_providers: Sequence[Union[str, tuple[str, dict[str, Any]]]] = [
+    onnx_providers: Sequence[tuple[str, dict[str, Any]]] = [
-        "CPUExecutionProvider"
+        ("CPUExecutionProvider", {}),
     ],
 ):
 
@@ -45,6 +45,11 @@ def synthesize(
                 model = model_holder.get_model(model_info.name, model_files[0])
                 model.load()
 
+                # ロードされた InferenceSession の ExecutionProvider が一致するか確認
+                # 一致しない場合、指定された ExecutionProvider で推論できない状態
+                assert model.onnx_session is not None
+                assert model.onnx_session.get_providers()[0] == onnx_providers[0][0]
+
                 # 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:
 
@@ -87,23 +92,35 @@ def test_synthesize_cpu():
 
 
 def test_synthesize_cuda():
-    pytest.importorskip("torch.cuda")
     synthesize(inference_type="torch", device="cuda")
 
 
 def test_synthesize_onnx_cpu():
-    synthesize(inference_type="onnx", onnx_providers=["CPUExecutionProvider"])
+    synthesize(
+        inference_type="onnx",
+        onnx_providers=[
+            ("CPUExecutionProvider", {}),
+        ],
+    )
 
 
 def test_synthesize_onnx_cuda():
     synthesize(
         inference_type="onnx",
         onnx_providers=[
-            ("CUDAExecutionProvider", {"cudnn_conv_algo_search": "DEFAULT"})
+            ("CUDAExecutionProvider", {"cudnn_conv_algo_search": "DEFAULT"}),
         ],
     )
 
 
 def test_synthesize_onnx_directml():
-    pytest.importorskip("onnxruntime_directml")
+    synthesize(
-    synthesize(inference_type="onnx", onnx_providers=["DmlExecutionProvider"])
+        inference_type="onnx",
+        onnx_providers=[
+            # device_id: 0 は、システムにインストールされているプライマリディスプレイ用 GPU に対応する
+            # プライマリディスプレイ用 GPU (GPU 0) よりも性能の高い GPU が接続されている環境では、
+            # 適宜 device_id を変更する必要がある
+            # ref: https://github.com/w-okada/voice-changer/issues/410#issuecomment-1627994911
+            ("DmlExecutionProvider", {"device_id": 0}),
+        ],
+    )