changes

embedding_converter/README.md

@@ -26,9 +26,8 @@ Setup
 This `config.ini` utilizes the MegaFace dataset to train the Embedding Converter for SimSwap.
 
 ```
-[preparing.dataset]
-dataset_path = .datasets/images
-image_pattern = {}/*.*g
+[training.dataset]
+dataset_file_pattern = .datasets/images/{}/*.*g
 ```
 
 ```

embedding_converter/config.ini

@@ -1,6 +1,5 @@
-[preparing.dataset]
-dataset_path =
-image_pattern =
+[training.dataset]
+dataset_file_pattern =
 
 [training.loader]
 split_ratio =

embedding_converter/src/data_loader.py

@@ -6,12 +6,12 @@ import torchvision.transforms as transforms
 from torch.utils.data import Dataset
 
 from .helper import read_image
-from .types import Batch, ImagePathList
+from .types import Batch, Paths
 
 
 class DataLoaderRecognition(Dataset[torch.Tensor]):
-	def __init__(self, dataset_path : str, dataset_image_pattern : str) -> None:
-		self.image_paths = self.prepare_image_paths(dataset_path, dataset_image_pattern)
+	def __init__(self, dataset_file_pattern : str) -> None:
+		self.image_paths = self.prepare_image_paths(dataset_file_pattern)
 		self.dataset_total = len(self.image_paths)
 		self.transforms = self.compose_transforms()
 
@@ -24,9 +24,8 @@ class DataLoaderRecognition(Dataset[torch.Tensor]):
 	def __len__(self) -> int:
 		return self.dataset_total
 
-	def prepare_image_paths(self, dataset_path : str, dataset_image_pattern : str) -> ImagePathList:
-		image_paths = glob.glob(dataset_image_pattern.format(dataset_path))
-		return image_paths
+	def prepare_image_paths(self, dataset_file_pattern : str) -> Paths:
+		return glob.glob(dataset_file_pattern)
 
 	def compose_transforms(self) -> transforms:
 		transform = transforms.Compose(

embedding_converter/src/helper.py

@@ -1,17 +1,7 @@
-import platform
-
 import cv2
-import numpy
 
 from .types import VisionFrame
 
 
-def is_windows() -> bool:
-	return platform.system().lower() == 'windows'
-
-
 def read_image(image_path : str) -> VisionFrame:
-	if is_windows():
-		image_buffer = numpy.fromfile(image_path, dtype = numpy.uint8)
-		return cv2.imdecode(image_buffer, cv2.IMREAD_COLOR)
 	return cv2.imread(image_path)

embedding_converter/src/training.py

@@ -115,11 +115,10 @@ def create_trainer() -> Trainer:
 
 
 def train() -> None:
-	dataset_path = CONFIG.get('preparing.dataset', 'dataset_path')
-	dataset_image_pattern = CONFIG.get('preparing.dataset', 'image_pattern')
+	dataset_file_pattern = CONFIG.get('training.dataset', 'image_pattern')
 	resume_file_path = CONFIG.get('training.output', 'resume_file_path')
 
-	dataset = DataLoaderRecognition(dataset_path, dataset_image_pattern)
+	dataset = DataLoaderRecognition(dataset_file_pattern)
 	training_loader, validation_loader = create_loaders(dataset)
 	embedding_converter_trainer = EmbeddingConverterTrainer()
 	trainer = create_trainer()

embedding_converter/src/types.py

@@ -3,7 +3,7 @@ from typing import Any, List, TypeAlias
 from numpy.typing import NDArray
 from torch import Tensor
 
-ImagePathList : TypeAlias = List[str]
+Paths : TypeAlias = List[str]
 Batch : TypeAlias = Tensor
 Embedding : TypeAlias = Tensor
 VisionFrame : TypeAlias = NDArray[Any]

requirements.txt

@@ -2,6 +2,7 @@
 lightning==2.5.0
 numpy==1.26.4
 onnx==1.17.0
+onnxruntime==1.20.1
 opencv-python==4.11.0.86
 pytorch-msssim==1.0.0
 torch==2.6.0