myay / SPICE-Torch

A framework for connecting SPICE simulations of analog computing neuron circuits with PyTorch accuracy evaluations for Binarized (and soon Quantized) Neural Networks.

SPICE-Torch

A framework for connecting SPICE simulations of analog computing neuron circuits with PyTorch accuracy evaluations for Binarized (and soon Quantized) Neural Networks.

Tested setups:

• Python 3.10.6 (pip + venv), PyTorch 1.13.1, GeForce RTX 3060 Ti 8GB (Driver Version: 525.60.13, CUDA Version: 12.0), Ubuntu 22.04.1
• Python 3.6.9, PyTorch 1.5.0, GeForce GTX 1060 6GB (Driver Version: 440.100, CUDA Version: 10.2), Ubuntu 18.04.4
• Python 3.6.13 (conda), 1.7.0+cu110, GeForce GTX 1080 8GB (Driver Version: 450.102.04, CUDA Version: 11.0), Ubuntu 18.04.6
• Python 3.9.7, PyTorch 1.9.0, GeForce GTX 3080 10GB (Driver Version: 512.15, CUDA Version: 11.6)
• Python 3.10.6, PyTorch 2.01+cu118, GeForce RTX 3060 12 GB (Driver Version: 531.14, CUDA Version: 11.71), Windows 10

Supported features:

• BNN models for FC, VGG3, VGG7, ResNet18
• Datasets: FashionMNIST, KMNIST, SVHN, CIFAR10, IMAGENETTE
• Error model application for Linear and Conv2d layers (direct mapping and distribution-based mapping)
• Variable crossbar array size
• Training with error models
• Performance mode (MAC engine with error application are fused into one GPU-kernel)

TODOs:

• Optimize MAC engine for execution time
• Support for QNNs
• More NN models

CUDA-based Error Model Application and Binarization/Quantization

First, install PyTorch. For fast binarization/quantization and error injection during training, CUDA support is needed. To enable it, install pybind11 and CUDA toolkit.

Then, to install the CUDA-kernels, go to the folder code/cuda/ and run

./install_kernels.sh

After successful installation of all kernels, run the binarization/quantization-aware training with error injection for BNNs using

python3 run.py --model=VGG3 --dataset=FMNIST --batch-size=256 --epochs=5 --lr=0.001 --step-size=2 --gamma=0.5 --train-model=1 --save-model=vgg3_test.

Here is a list of the command line parameters for running the error evaluations with SPICE-Torch:

Command line parameter	Options
--model	FC, VGG3, VGG7, ResNet
--dataset	MNIST, FMNIST, QMNIST, SVHN, CIFAR10, IMAGENETTE
--an-sim	int, whether to turn on the mapping from SPICE, default: None
--mapping	string, loads a direct mapping from the specified path, default: None
--mapping-distr	string, loads a distribution-based mapping from the specified path, default: None
--array-size	int, specifies the size of the crossbar array, default: None
--performance-mode	int, specify whether to activate the faster and more memory-efficient performance mode (when using this sub-MAC results can only be changed in cuda-kernel!), default: None
--print-accuracy	int, specifies whether to print inference accuracy, default: None
--test-error-distr	int, specifies the number of repetitions to perform in accuracy evaluations for distribution based evaluation, default: None
--train-model	bool, whether to train a model, default: None
--epochs	int, number of epochs to train, default: 10
--lr	float, learning rate, default: 1.0
--gamma	float, learning rate step, default: 0.5
--step-size	int, learning rate step site, default: 5
--batch-size	int, specifies the batch size in training, default: 64
--test-batch-size	int, specifies the batch size in testing, default: 1000
--save-model	string, saves a trained model with the specified name in the string, default:None
--load-model-path	string, loads a model from the specified path in the string, default: None
--load-training-state	string, saves a training state with the specified name in the string, default:None
--save-training-state	string, loads a training state from the specified path in the string, default: None
--gpu-num	int, specifies the GPU on which the training should be performed, default: 0
--profile-time	int, Specify whether to profile the execution time by specifying the repetitions, default: None
--extract-absfreq	int, Specify whether to extract the absolute frequency of MAC values, default: None
--extract-absfreq-resnet	int, Specify whether to extract the absolute frequency of MAC values for resnet, default: None

More information on the command line parameters can be found here.

Please contact me if you have any questions: [e-mail].