Official repository for:

Pruning vs XNOR-Net: A ComprehensiveStudy on Deep Learning for AudioClassification in Microcontrollers

Note: Please run all the python scripts from a TERMINAL

Audio Classification

Dataset preparation

1. Please follow official repository for ACDNet to prepare the audio datasets for these experiments
2. For AudioEvent dataset, the scripts are in this repository audio/resources/ae_

Training ACDNet, Micro-ACDNet and Mini-ACDNet

Example: python audio/trainer.py --model_path 'path/to/mini/or/micro/acdnet' --dataset 'esc10 or esc50 or us8k' --data '/path/to/dataset/' --xnor 0 --nClasses 50 --model_name 'file_name_tosave_the_trained_model'

1. ACDNet on ESC-10: python audio/trainer.py --model_path '' --dataset 'esc10' --data '/path/to/dataset/' --xnor 0 --nClasses 10 --model_name 'model_name'
2. ACDNet on us8k: python audio/trainer.py --model_path '' --dataset 'us8k' --data '/path/to/dataset/' --xnor 0 --nClasses 10 --model_name 'model_name'
3. ACDNet on AudioEvent (20 class): python audio/trainer.py --model_path '' --dataset 'audioevent' --data '/path/to/dataset/' --xnor 0 --nClasses 20 --model_name 'model_name'
4. Mini-ACDNet on ESC-50: python audio/trainer.py --model_path 'audio/models/mini_acdnet.pt' --dataset 'esc50' --data '/path/to/dataset/' --xnor 0 --nClasses 50 --model_name 'model_name'
5. Micro-ACDNet on ESC-10: python audio/trainer.py --model_path 'audio/models/micro_acdnet.pt' --dataset 'esc10' --data '/path/to/dataset/' --xnor 0 --nClasses 10 --model_name 'model_name'

Training XNOR-Net for ACDNet, Micro-ACDNet and Mini-ACDNet

1. XACDNet on ESC-50: python audio/trainer.py --model_path '' --dataset 'esc50' --data '/path/to/dataset/' --xnor 1 --nClasses 50 --model_name 'xacdnet_esc50'
2. XMicroACDnet on ESC-40: python audio/trainer.py --model_path 'audio/models/micro_acdnet.pt' --dataset 'esc40' --data '/path/to/dataset/' --xnor 1 --nClasses 40 --model_name 'xmicro_acdnet_esc40'
3. XMiniAcdnet on US8k: python audio/trainer.py --model_path 'audio/models/mini_acdnet.pt' --dataset 'us8k' --data '/path/to/dataset/' --xnor 1 --nClasses 10 --folds_to_train '[1,2]' --model_name 'xmini_acdnet_us8k'

Note: You can use the pretrained micro and mini acdnets files or use your own mini and micro acdnet versions using ACDNet repository

Quantization of MicroACDNet

1. Update settings opt.dataset, opt.data, opt.model_path, opt.model_name, opt.split in audio/quantization.py
2. Run: python audio/quantization.py

Image Classification

Dataset preparation

1. Create a folder to store the datasets and copy the path
2. Update args.datasetpath with the copied path in image/resources/prepare_dataset.py
3. Update args.datasetpath with the copied path in image/resources/prepare_dataset_cifar100.py
4. Update args.dataset_path in image/resources/settings.py
5. To prepare CIFAR-10, run in terminal: python image/resources/prepare_dataset.py
6. To prepare CIFAR-100, run in terminal: python image/resources/prepare_dataset_cifar100.py

Train RESNET-18 (Use Terminal)

1. CIFAR-10: python resnet/trainer.py --dataset 'cifar10' --icl 0 --classes 10
2. CIFAR-100: python resnet/trainer.py --dataset 'cifar100' --icl 0 --classes 100
3. For incremental training (e.g. CIFAR-30): python resnet/trainer.py --dataset 'cifar100' --icl 1 --classes 30

Train XNOR-Net version of RESNET-18

1. CIFAR-10: python resnet/bin_trainer.py --dataset 'cifar10' --icl 0 --classes 10
2. CIFAR-100: python resnet/bin_trainer.py --dataset 'cifar100' --icl 0 --classes 100
3. For incremental training (e.g. CIFAR-30): python resnet/bin_trainer.py --dataset 'cifar100' --icl 1 --classes 30