cph-imgclassification 0.1.0

A generic, config-driven PyTorch Lightning pipeline for image classification tasks

CPH Image Classification

A generic, modular, and reusable PyTorch Lightning pipeline for training image classification models using CNNs. This package is fully config-driven, allowing you to train models on any image classification dataset by simply modifying a YAML configuration file.

Installation

pip install cph-imgclassification

Quick Start

1. Install the Package

pip install cph-imgclassification

2. Organize Your Data

Organize your images in class folders:

YourProjectName/data/images/
├── class1/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
├── class2/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
└── class3/
    ├── image1.jpg
    └── ...

3. Create Configuration File

Create config.yaml:

# Image Classification Model Configuration
seed_everything: true

trainer:
  callbacks:
    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
      init_args:
        filename: "{epoch}-{val_acc:.2f}.best"
        monitor: "val_acc"
        mode: "max"
        save_top_k: 1
    - class_path: cph_imgclassification.imageclassification.callbacks.ONNXExportCallback
      init_args:
        output_dir: "models"
        model_name: "my_model"
        input_shape: [3, 224, 224]

  logger:
    class_path: lightning.pytorch.loggers.TensorBoardLogger
    init_args:
      save_dir: "lightning_logs"
      name: "MyProjectTraining"

  max_epochs: 50
  accelerator: auto
  devices: auto
  precision: 16-mixed

model:
  class_path: cph_imgclassification.imageclassification.modelmodule.ModelModuleIMG
  init_args:
    lr: 0.001
    weight_decay: 0.0001
    model:
      class_path: cph_imgclassification.imageclassification.modelfactory.ImageClassificationModel
      init_args:
        input_channels: 3
        num_classes: 0  # Auto-set from datamodule
        architecture: "medium"  # "simple", "medium", "deep", or "custom"
        input_size: [224, 224]

optimizer: 
  class_path: torch.optim.Adam
  init_args:
    lr: 0.001
    weight_decay: 0.0001

data:
  class_path: cph_imgclassification.imageclassification.datamodule.DataModuleIMG
  init_args:
    data_dir: "YourProjectName/data/images"
    image_size: [224, 224]
    batch_size: 32
    num_workers: 4
    val_split: 0.2
    random_seed: 42
    augmentation:
      enabled: true
      rotation: 15
      horizontal_flip: true
    normalization:
      mean: [0.485, 0.456, 0.406]
      std: [0.229, 0.224, 0.225]
    save_preprocessor: true
    preprocessor_path: "models/label_encoder.joblib"

fit:
  ckpt_path: null

test:
  ckpt_path: best

4. Train Your Model

cph-imgclassification fit --config config.yaml

Or for fit + test workflow:

cph-imgclassification fit --config config.yaml
cph-imgclassification test --config config.yaml

Features

• Fully Config-Driven: All settings controlled via YAML files
• Generic & Reusable: Use for any image classification task
• Flexible Data Formats: Supports folder-based and CSV-based datasets
• Auto-Dimension Detection: Automatically detects number of classes
• Configurable CNN Architectures: Choose from preset architectures or define custom layers
• Data Augmentation: Built-in augmentation support
• Production-Ready: Exports models to ONNX format
• PyTorch Lightning: Built on PyTorch Lightning for scalable training
• Comprehensive Metrics: Tracks Accuracy, F1-Score, Precision, and Recall

Usage Examples

Training Only

cph-imgclassification fit --config config.yaml

Testing Only

cph-imgclassification test --config config.yaml

Resume Training

cph-imgclassification fit --config config.yaml --fit.ckpt_path path/to/checkpoint.ckpt

Using CSV Data Format

In your config file, use:

data:
  class_path: cph_imgclassification.imageclassification.datamodule.DataModuleIMG
  init_args:
    csv_path: "data/image_labels.csv"
    image_path_col: "image_path"
    label_col: "label"
    # ... other settings

Configuration Reference

Model Architectures

• simple: 2 conv blocks (32, 64 channels) - Good for small datasets
• medium: 4 conv blocks (64, 128, 256, 512 channels) - Balanced performance
• deep: 6 conv blocks (64, 128, 256, 256, 512, 512 channels) - For large datasets
• custom: Define your own conv layers

Data Augmentation Options

• rotation: Random rotation degrees
• horizontal_flip: Random horizontal flip
• vertical_flip: Random vertical flip
• color_jitter: Color jitter intensity
• random_crop: Crop size for random cropping

Output Files

After training, you'll find:

• ONNX Model: models/your_model_name.onnx
• Label Encoder: models/label_encoder.joblib
• Checkpoints: lightning_logs/.../checkpoints/
• TensorBoard Logs: lightning_logs/

Viewing Training Progress

tensorboard --logdir lightning_logs

Then open http://localhost:6006 in your browser.

Python API

You can also use the package programmatically:

from cph_imgclassification.imageclassification import (
    DataModuleIMG,
    ImageClassificationModel,
    ModelModuleIMG
)

# Create datamodule
datamodule = DataModuleIMG(
    data_dir="data/images",
    image_size=(224, 224),
    batch_size=32
)

# Create model
model = ImageClassificationModel(
    input_channels=3,
    num_classes=10,
    architecture="medium"
)

# Create Lightning module
lightning_model = ModelModuleIMG(model=model, lr=0.001)

Requirements

• Python >= 3.8
• PyTorch >= 2.0.0
• PyTorch Lightning >= 2.1.0

See requirements.txt for full dependency list.

License

MIT License

Author

chandra

Repository

https://github.com/imchandra11/cph-imgclassification

Support

For issues, questions, or contributions, please visit the GitHub repository.