GPUScheduling-for-Sparsity-and-Reliability

Integrating sparsity for CNNs and GPU scheduling for reliability study.

Logs

All output logs are stored in a directory named based on model-name-dataset.

Testing an original model:

python main.py --model=model-name --dataset=dataset-name --batch-size=xx

model-name is set based on: https://github.com/chenyaofo/pytorch-cifar-models/

dataset: cifar10, or cifar100, or imagenet

batch-size is an integer number

Applying sparsity and fine-tuning a model

`python main.py --model=model-name --dataset=dataset-name --batch-size=xx --is-sparsity --is-finetune --epoch=xx

epoch is an integer number used for fine-tuning. recommended: 20 for cifar10, and 50 for cifar100

Testing a sparse model with semi-structured accuracy

python main.py --model=model-name --dataset=dataset-name --batch-size=xx --is-sparsity

Exporting tiled values for inputs, weights, and sparsity metadata

python main.py --model=model-name --dataset=dataset-name --batch-size=1 --is-tiling --is-sparsity

It outputs the information for 100 tiles.

Fault injection

into metadata:

python main.py --model=model-name --dataset=dataset-name --batch-size=100 --is-sparsity --is-FI --FI-location=metadata --FI-iteration=10 --FI-study=random

into weights:

python main.py --model=model-name --dataset=dataset-name --batch-size=100 --is-sparsity --is-FI --FI-location=weight --FI-iteration=10 --FI-study=random

NB! Due to memory usage, execute fault injection with a maximum of 10 iterations.

NB! After running an FI experiment, you can obtain the overall results using the log-reader.py file and editing it, taking into account the name of the FI log file.