CIFAR10 Deep Learning Speed Benchmark

For informally benchmarking CUDA hardware.

IMPORTANT NOTES ABOUT ACCURACY:

• This program outputs the accuracy only as a sanity check.
• Accuracy should not be compared across batch sizes, since the batch size influences the total number of iterations, which in turn influences the accuracy.

Usage

This script looks for the CIFAR data inside $CIFAR; if the environment variable does not exist, it downloads the dataset into ./data. You can manually download the data here and put it inside the directory.

usage: benchmark.py [-h] [--gpus GPUS] [--progressive]
                    [--measurements MEASUREMENTS] [--size SIZE]
                    [--epochs EPOCHS] [--batches BATCHES]
                    [--batch-size BATCH_SIZE]
                    {densenet,wideresnet}

Image classification speed benchmark

positional arguments:
  {densenet,wideresnet}

optional arguments:
  -h, --help                     show this help message and exit
  --gpus GPUS                    Number of gpus to use. Default: all
  --progressive                  Try 1 gpus, 2 gpus, 3 gpus, etc.
  --measurements MEASUREMENTS    Num measurements for avg and std
  --size SIZE                    image size multiplier
  --epochs EPOCHS
  --batches BATCHES              stop early for testing
  --batch-size BATCH_SIZE        Batch size PER GPU