GFNet-Dynn is a hybrid deep learning model that combines the advantages of the GFNet and Dynn architectures. This project aims to significantly enhance the inference efficiency of deep neural networks on edge devices while maintaining high accuracy by introducing a dynamic exit mechanism (i.e., "early exit mechanism").

Please refer to the official guides of GFNet and Dynn to install the necessary Python dependencies.

Place the supported datasets in the data/ folder at the project root directory.

Supported Datasets:

• AID
• NWPU
• PatternNet
• UCM
• NaSC

📁 Note: The paths and structures of the datasets are defined in data_loader_help.py and dataset_config.py. Please ensure they match the actual paths.

Use the provided example script for training:

bash

./example_usage.sh

Or run it manually:

bashpython main_dynn.py \
    --data-set NWPU \
    --data-path /path/to/NWPU \
    --checkpoint-path /path/to/pretrained/checkpoint.pth \
    --batch 64 \
    --num_epoch 200 \
    --ce_ic_tradeoff 0.75 \
    --lr 0.001 \
    --input-size 256

Evaluate a pre-trained model using the --eval parameter:

bashpython main_dynn.py \
    --data-set NWPU \
    --data-path /path/to/NWPU \
    --checkpoint-path /path/to/pretrained/checkpoint.pth \
    --eval \
    --input-size 256

GFNet-Dynn adopts a two-stage training strategy to optimize early exit performance:

• Train each intermediate classifier (exit) one by one.
• Freeze the exit classifier with the best performance based on the validation set accuracy.
• If an exit has not improved for patience consecutive rounds, save its optimal weights and freeze it.
• If all exits have converged, terminate this stage early.

• Train the backbone network and intermediate classifiers (exits) together to further enhance overall performance.
• The final model supports dynamically selecting the exit point during inference based on confidence.

Model checkpoints during training are saved by default in the checkpoint/ directory.

├── checkpoint/
	    └── checkpoint_{dataset}_{ce_ic_tradeoff}_confEE/       # Training results for different experimental configurations
├── checkpoint_warmup/                                          # Checkpoints from the warmup stage
└── checkpoint_result/                                          # Final models manually selected and saved

All dataset configuration information is uniformly managed in the dataset_config.py file.

pythonDATASET_CONFIG = {
    'UCM': {
        'num_classes': 21,
        'default_input_size': 256,
        'train_folder': 'train',
        'val_folder': 'val',
        'use_random_split': False,
        'default_data_path': './data/UCM'
    },
    # ... Other datasets
}

📌 Tip: To add a new dataset or modify an existing configuration, please edit this file.

This project draws on content from the following repositories:

• GFNet
• Dynn

Please refer to these repositories for more detailed information about the underlying architectures and methods.