Pytorch Image Classification Trainer (V2)

This repository provides a robust and flexible framework for training image classification models using PyTorch. It's designed to be highly customizable and easy to use, allowing you to run experiments with different models, data augmentation techniques, and training configurations.

📂 Project Structure

text ├── dataset │ └── README.md ├── training_eng │ ├── core │ │ ├── device.py │ │ ├── misc.py │ │ └── callback_arg.py │ ├── train_utils │ │ ├── early_stopping.py │ │ └── model_eval.py │ ├── data_utils │ │ ├── data_proc.py │ │ └── data_loader.py │ └── trainer.py ├── train_exper.py ├── tensorboard.cmd ├── logs ├── cache ├── uv.lock ├── pyproject.toml ├── tensorboard.sh ├── run_expers.py ├── GIT_COMMIT.md ├── models └── expers.toml

🌶️ Features

• Experiment Management: Easily define and run multiple experiments using a simple TOML configuration file (expers.toml).
• Data Loading and Processing: Efficient data loading and augmentation pipelines with support for various backends (opencv, pil, turbojpeg).
• Flexible Training Loop: The core training loop in training_eng/trainer.py supports:
  • Mixed precision training
  • Gradient accumulation
  • Learning rate schedulers
  • Early stopping
  • TensorBoard logging
  • Model compilation with torch.compile
• Extensible Model Support: Easily integrate any PyTorch model. The current example uses efficientnet-pytorch.
• Rich Console Output: Uses the rich library for beautiful and informative console output.

🚀 Getting Started

Prerequisites

• Python 3.11+
• PyTorch
• Other dependencies listed in pyproject.toml

Installation

1. Clone the repository:

   bash git clone https://github.com/AidinHamedi/Pytorch-Img-Classification-Trainer-V2.git cd Pytorch-Img-Classification-Trainer-V2

2. Install dependencies: This project uses uv for package management.

   bash pip install uv uv sync

   If you want to use turbojpeg

   bash uv sync --extra tjpeg

Dataset Setup

Place your training and validation datasets in the dataset/train and dataset/validation directories, respectively. The data should be organized in subdirectories, where each subdirectory represents a class.

text dataset/ ├── train/ │ ├── class_a/ │ │ ├── image1.jpg │ │ └── image2.jpg │ └── class_b/ │ ├── image3.jpg │ └── image4.jpg └── validation/ ├── class_a/ │ ├── image5.jpg │ └── image6.jpg └── class_b/ ├── image7.jpg └── image8.jpg

🤔 How to Run Experiments

1. Define your experiments in expers.toml:

   Each section in expers.toml represents a separate experiment. You can specify the model name and other parameters for each experiment.

   Example expers.toml:

   ```toml ["Test"] model_name = "efficientnet-b0"

   ["Experiment2"] modelname = "efficientnet-b1" ```

2. Configure training parameters in train_exper.py:

   This file contains the main configuration for the training process, including:

- Dataset paths
- Image resolution
- Batch size
- Data augmentation settings
- Optimizer and loss function
- And other training-related hyperparameters.

1. Run the experiments:

   Execute the run_expers.py script to start training all the experiments defined in expers.toml.

   bash python run_expers.py

   The script will iterate through each experiment, train the model, and save the results.

🎛️ Monitoring and Results

• TensorBoard: Monitor the training process in real-time using TensorBoard.
  • On Windows, run tensorboard.cmd.
  • On Linux/macOS, run tensorboard.sh.
• Saved Models: The best and latest models for each experiment are saved in the models directory.
• Logs: Training logs are stored in the logs directory.

🧪 How it Works

1. run_expers.py: This is the main entry point. It reads the expers.toml file and iterates through each experiment defined in it.
2. train_exper.py: For each experiment, this script sets up the data loaders, model, optimizer, and loss function based on the configuration. It then calls the fit function from training_eng/trainer.py. (can be modified to suit your needs)
3. training_eng/trainer.py: This file contains the core fit function that implements the training loop. It handles all the complexities of training, including mixed precision, gradient accumulation, early stopping, and logging.
4. training_eng/data_utils: These modules handle the creation of data pairs, data loading, and data augmentation.
5. training_eng/train_utils: These modules provide utilities for model evaluation and early stopping.
6. training_eng/core: These modules provide core functionalities like device management and callback arguments.

📷 Example Output

🤝 Contributing

Contributions are welcome! Please feel free to submit a pull request or open an issue.

📝 License

 Copyright (c) 2025 Aidin Hamedi

 This software is released under the MIT License.
 https://opensource.org/licenses/MIT