DeepLib

A unified PyTorch library for computer vision tasks, focusing on object detection, semantic segmentation, and anomaly detection.

Installation

Prerequisites

DeepLib requires PyTorch and torchvision to be installed first. For optimal performance, CUDA 11.8 or above is recommended.

You can install PyTorch with CUDA support using:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

For other installation options (CPU-only, different CUDA versions), see the PyTorch installation guide.

Installing DeepLib

Full Installation

Install DeepLib with all optional dependencies (recommended):

pip install deeplib

Core Installation

Install only the core functionality (no logging backends):

pip install deeplib[core]

Documentation

Full documentation is available at https://jonleinena.github.io/deeplib/

Features

• Semantic Segmentation Models (✅ Implemented)

  • UNet
  • DeepLabV3
  • DeepLabV3+

• Experiment Tracking (✅ Implemented)

  • TensorBoard
  • MLflow
  • Weights & Biases (W&B)

• Object Detection Models (🚧 In Progress)

  • YOLOv4
  • YOLOv5
  • YOLOX
  • YOLOv7 and YOLOv9
  • Faster R-CNN

• Anomaly Detection Models (🚧 In Progress)

  • PatchCore
  • FastFlow
  • PADIM
  • Other anomalib implementations

Quick Start - Semantic Segmentation

from deeplib.models.segmentation import UNet
from deeplib.trainers import SegmentationTrainer
from deeplib.datasets import SegmentationDataset
from deeplib.loggers import WandbLogger  # or TensorBoardLogger, MLFlowLogger
from torch.utils.data import DataLoader

# Initialize model
model = UNet(num_classes=4)

# Prepare dataset
train_dataset = SegmentationDataset(
    root="path/to/data",
    images_dir="images",
    masks_dir="masks",
    num_classes=4,
    split="train"
)
val_dataset = SegmentationDataset(
    root="path/to/data",
    images_dir="images",
    masks_dir="masks",
    num_classes=4,
    split="val"
)

# Create dataloaders
train_loader = DataLoader(train_dataset, batch_size=8, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=8)

# Configure loss function (e.g., Dice Loss)
model.configure_loss('dice', {'ignore_index': 255})

# Initialize logger (choose one)
logger = WandbLogger(
    experiment_name="segmentation_experiment",
    project="deeplib-segmentation"
)
# Or use TensorBoard:
# logger = TensorBoardLogger(experiment_name="segmentation_experiment")
# Or use MLflow:
# logger = MLFlowLogger(experiment_name="segmentation_experiment")

# Initialize trainer with logger
trainer = SegmentationTrainer(
    model=model,
    train_loader=train_loader,
    val_loader=val_loader,
    monitor_metric='iou',  # Monitor IoU for LR scheduling
    logger=logger
)

# Train model
trainer.train(
    num_epochs=100,
    save_path='best_model.pth'
)

For more examples and detailed usage, check the examples directory.

Project Structure

deeplib/
├── models/
│   ├── segmentation/  # ✅ Semantic segmentation models
│   ├── detection/     # 🚧 Object detection models (TODO)
│   └── anomaly/       # 🚧 Anomaly detection models (TODO)
├── trainers/          # Training logic and utilities
├── datasets/          # Dataset implementations
├── loggers/           # ✅ Experiment tracking (TensorBoard, MLflow, W&B)
└── utils/             # Utility functions

Development Roadmap

High Priority

• Implement experiment tracking
  • TensorBoard support
  • MLflow support
  • W&B support
• Add comprehensive documentation
  • API Reference
  • Examples
  • Installation Guide
• Implement object detection models
  • YOLOv4
  • YOLOv5
  • Faster R-CNN
• Add anomaly detection support
  • PatchCore
  • FastFlow
  • PADIM
• Add data augmentation pipeline
• Add model export (ONNX, TorchScript)

Medium Priority

• Add more segmentation models
  • FPN
  • SegFormer
  • BEiT
• Add test suite
• Add model benchmarks
• Add visualization tools

Low Priority

• Add multi-GPU training support
• Add quantization support
• Add model pruning
• Add hyperparameter tuning

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

This library is inspired by the following projects:

• torchvision
• anomalib
• segmentation-models-pytorch
• YOLOMIT