ImgTorch is a minimal yet powerful image importer and preprocessor tailored for classification tasks in PyTorch. It supports both common and RAW image formats, applies consistent preprocessing, and enables fast dataset creation and visualization — all with minimal dependencies.

• Directory-based class labeling — each subfolder = one class
• Supports RAW and standard formats: .jpg, .png, .cr2, .nef, .dng, etc.
• Aspect-preserving resize and padding to uniform shape
• Converts to PyTorch tensors ready for training
• Save/load dataset as .pt files for fast reuse
• Live previews via matplotlib and terminal-friendly ASCII art
• Graceful handling of unreadable or corrupted files
• Minimal dependencies: Only uses PyTorch, Pillow, rawpy, matplotlib, tqdm

Your dataset should be organized by class subfolders:

your_dataset/
├── ClassA/
│   ├── iMg1.jpg
│   ├── ige2.cr2
│   ├── imG3.cr3
│   └── imag4.png
├── ClassB/
│   ├── imAg5.cr2
│   ├── imag6.nef
│   └── imge7.jpeg
├── ClassC/
│   ├── img8.dng
│   └── img9.jpeg

from imgtorch import ImgTorch

imp = ImgTorch(
    baseDir="your_dataset",
    classDir=["ClassA", "ClassB", "ClassC"]
)

imp.collect_images()     # Scan all images
imp.shuffle_images()     # Optional: randomize order
imp.process_images(imageSize=(128,256))     # Load, resize, convert to tensor

imp.preview_images(max_images=6)         # Matplotlib preview
imp.preview_ASCII(count=3, contrast=1.2) # Terminal-friendly ASCII visualization

imp.save_dataset("dataset.pt")    # Save tensors to disk
X, Y = imp.get_dataset()          # Retrieve processed data
print(X.shape, Y.shape)

• RAW formats are decoded using rawpy and converted to RGB using Pillow.
• Aspect ratio is preserved using thumbnail() and centered padding.
• Corrupted or unreadable files are skipped and listed.

Install with:

pip install torch torchvision pillow rawpy matplotlib tqdm

Jesse Hng, 2025
A practical tool for quick dataset preparation in terminal or notebook environments.