def Run(benchmark_spec):
"""Run MXNet on the cluster for each model specified.
Args:
benchmark_spec: The benchmark specification. Contains all data that is
required to run the benchmark.
Returns:
A list of sample.Sample objects.
"""
_UpdateBenchmarkSpecWithFlags(benchmark_spec)
vm = benchmark_spec.vms[0]
mx_benchmark_dir = 'incubator-mxnet/example/image-classification'
results = []
for model in FLAGS.mx_models:
num_layers = _GetNumLayers(model)
batch_size = _GetBatchSize(model, num_layers)
benchmark_spec.model = model
benchmark_spec.batch_size = batch_size
benchmark_spec.num_layers = num_layers
benchmark_spec.image_shape = _GetImageShape(model)
mx_benchmark_cmd = (
'python train_imagenet.py '
'--benchmark=1 '
'--network={network} '
'--batch-size={batch_size} '
'--image-shape={image_shape} '
'--num-epochs={num_epochs} '
'--dtype={precision} '
'--kv-store={key_value_store}').format(
network=model,
batch_size=batch_size,
image_shape=benchmark_spec.image_shape,
num_epochs=benchmark_spec.num_epochs,
precision=benchmark_spec.precision,
key_value_store=benchmark_spec.key_value_store)
if benchmark_spec.device == GPU:
num_gpus = cuda_toolkit.QueryNumberOfGpus(vm)
mx_benchmark_cmd = '{env} {cmd} --gpus {gpus}'.format(
env=mxnet.GetEnvironmentVars(vm),
cmd=mx_benchmark_cmd,
gpus=','.join(str(n) for n in range(num_gpus)))
elif benchmark_spec.device == CPU:
# Specifies the number of threads to use in CPU test.
# https://mxnet.incubator.apache.org/faq/perf.html
mx_benchmark_cmd = 'OMP_NUM_THREADS={omp_num_threads} {cmd}'.format(
omp_num_threads=vm.NumCpusForBenchmark() // 2,
cmd=mx_benchmark_cmd)
if num_layers:
mx_benchmark_cmd = '%s --num-layers %s' % (mx_benchmark_cmd, num_layers)
run_command = 'cd %s && %s' % (mx_benchmark_dir,
mx_benchmark_cmd)
stdout, stderr = vm.RobustRemoteCommand(run_command, should_log=True)
results.append(_MakeSamplesFromOutput(benchmark_spec, stdout or stderr))
return results
def Run(benchmark_spec):
"""Run MXNet on the cluster for each model specified.
Args:
benchmark_spec: The benchmark specification. Contains all data that is
required to run the benchmark.
Returns:
A list of sample.Sample objects.
"""
_UpdateBenchmarkSpecWithFlags(benchmark_spec)
vm = benchmark_spec.vms[0]
mx_benchmark_dir = 'incubator-mxnet/example/image-classification'
results = []
for model in FLAGS.mx_models:
num_layers = _GetNumLayers(model)
batch_size = _GetBatchSize(model, num_layers)
benchmark_spec.model = model
benchmark_spec.batch_size = batch_size
mx_benchmark_cmd = (
'python train_imagenet.py --benchmark 1 --network %s --batch-size %s '
'--image-shape %s --num-epochs %s --kv-store device') % (
model, batch_size, IMAGENET_SHAPE, benchmark_spec.num_epochs)
if benchmark_spec.device == GPU:
gpus = cuda_toolkit_8.QueryNumberOfGpus(vm)
mx_benchmark_cmd = '%s %s --gpus %s' % (mxnet.GetEnvironmentVars(
vm), mx_benchmark_cmd, ','.join(str(n) for n in range(gpus)))
if num_layers:
mx_benchmark_cmd = '%s --num-layers %s' % (mx_benchmark_cmd,
num_layers)
benchmark_spec.num_layers = num_layers
run_command = 'cd %s && %s' % (mx_benchmark_dir, mx_benchmark_cmd)
stdout, stderr = vm.RobustRemoteCommand(run_command, should_log=True)
results.append(_MakeSamplesFromOutput(benchmark_spec, stdout
or stderr))
return results