def _main(port):
base_model = ShuffleNetV2(32)
base_predictor = 'cortexA76cpu_tflite21'
transf = [
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip()
]
normalize = [
transforms.ToTensor(),
transforms.Normalize([0.49139968, 0.48215827, 0.44653124], [0.24703233, 0.24348505, 0.26158768])
]
train_dataset = serialize(CIFAR10, 'data', train=True, download=True, transform=transforms.Compose(transf + normalize))
test_dataset = serialize(CIFAR10, 'data', train=False, transform=transforms.Compose(normalize))
trainer = pl.Classification(train_dataloader=pl.DataLoader(train_dataset, batch_size=64),
val_dataloaders=pl.DataLoader(test_dataset, batch_size=64),
max_epochs=2, gpus=1)
simple_strategy = strategy.Random(model_filter=LatencyFilter(threshold=100, predictor=base_predictor))
exp = RetiariiExperiment(base_model, trainer, strategy=simple_strategy)
exp_config = RetiariiExeConfig('local')
exp_config.trial_concurrency = 2
exp_config.max_trial_number = 2
exp_config.trial_gpu_number = 1
exp_config.training_service.use_active_gpu = False
exp_config.execution_engine = 'base'
exp_config.dummy_input = [1, 3, 32, 32]
exp.run(exp_config, port)
print('Exported models:')
for model in exp.export_top_models(formatter='dict'):
print(model)
def _multi_trial_test(epochs, batch_size, port, benchmark):
# initalize dataset. Note that 50k+10k is used. It's a little different from paper
transf = [
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip()
]
normalize = [
transforms.ToTensor(),
transforms.Normalize([0.49139968, 0.48215827, 0.44653124],
[0.24703233, 0.24348505, 0.26158768])
]
train_dataset = serialize(CIFAR10,
'data',
train=True,
download=True,
transform=transforms.Compose(transf + normalize))
test_dataset = serialize(CIFAR10,
'data',
train=False,
transform=transforms.Compose(normalize))
# specify training hyper-parameters
training_module = NasBench101TrainingModule(max_epochs=epochs)
# FIXME: need to fix a bug in serializer for this to work
# lr_monitor = serialize(LearningRateMonitor, logging_interval='step')
trainer = pl.Trainer(max_epochs=epochs, gpus=1)
lightning = pl.Lightning(
lightning_module=training_module,
trainer=trainer,
train_dataloader=pl.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True),
val_dataloaders=pl.DataLoader(test_dataset, batch_size=batch_size),
)
strategy = Random()
model = NasBench101()
exp = RetiariiExperiment(model, lightning, [], strategy)
exp_config = RetiariiExeConfig('local')
exp_config.trial_concurrency = 2
exp_config.max_trial_number = 20
exp_config.trial_gpu_number = 1
exp_config.training_service.use_active_gpu = False
if benchmark:
exp_config.benchmark = 'nasbench101'
exp_config.execution_engine = 'benchmark'
exp.run(exp_config, port)
def test_multitrial_experiment(pytestconfig):
base_model = Net()
evaluator = get_mnist_evaluator()
search_strategy = strategy.Random()
exp = RetiariiExperiment(base_model, evaluator, strategy=search_strategy)
exp_config = RetiariiExeConfig('local')
exp_config.trial_concurrency = 1
exp_config.max_trial_number = 1
exp_config._trial_command_params = nas_experiment_trial_params(pytestconfig.rootpath)
exp.run(exp_config)
ensure_success(exp)
assert isinstance(exp.export_top_models()[0], dict)
exp.stop()
def _multi_trial_test(epochs, batch_size, port):
# initalize dataset. Note that 50k+10k is used. It's a little different from paper
transf = [
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip()
]
normalize = [
transforms.ToTensor(),
transforms.Normalize([x / 255 for x in [129.3, 124.1, 112.4]],
[x / 255 for x in [68.2, 65.4, 70.4]])
]
train_dataset = serialize(CIFAR100,
'data',
train=True,
download=True,
transform=transforms.Compose(transf + normalize))
test_dataset = serialize(CIFAR100,
'data',
train=False,
transform=transforms.Compose(normalize))
# specify training hyper-parameters
training_module = NasBench201TrainingModule(max_epochs=epochs)
# FIXME: need to fix a bug in serializer for this to work
# lr_monitor = serialize(LearningRateMonitor, logging_interval='step')
trainer = pl.Trainer(max_epochs=epochs, gpus=1)
lightning = pl.Lightning(
lightning_module=training_module,
trainer=trainer,
train_dataloader=pl.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True),
val_dataloaders=pl.DataLoader(test_dataset, batch_size=batch_size),
)
strategy = Random()
model = NasBench201()
exp = RetiariiExperiment(model, lightning, [], strategy)
exp_config = RetiariiExeConfig('local')
exp_config.trial_concurrency = 2
exp_config.max_trial_number = 20
exp_config.trial_gpu_number = 1
exp_config.training_service.use_active_gpu = False
exp.run(exp_config, port)
def _test_experiment_in_separate_process(rootpath):
try:
base_model, evaluator = _mnist_net('simple', {'max_epochs': 1})
search_strategy = strategy.Random()
exp = RetiariiExperiment(base_model, evaluator, strategy=search_strategy)
exp_config = RetiariiExeConfig('local')
exp_config.experiment_name = 'mnist_unittest'
exp_config.trial_concurrency = 1
exp_config.max_trial_number = 1
exp_config._trial_command_params = nas_experiment_trial_params(rootpath)
exp.run(exp_config)
ensure_success(exp)
assert isinstance(exp.export_top_models()[0], dict)
finally:
# https://stackoverflow.com/questions/34506638/how-to-register-atexit-function-in-pythons-multiprocessing-subprocess
import atexit
atexit._run_exitfuncs()
def test_multi_trial(model, pytestconfig):
evaluator_kwargs = {
'max_epochs': 1
}
base_model, evaluator = _mnist_net(model, evaluator_kwargs)
search_strategy = strategy.Random()
exp = RetiariiExperiment(base_model, evaluator, strategy=search_strategy)
exp_config = RetiariiExeConfig('local')
exp_config.experiment_name = 'mnist_unittest'
exp_config.trial_concurrency = 1
exp_config.max_trial_number = 1
exp_config._trial_command_params = nas_experiment_trial_params(pytestconfig.rootpath)
exp.run(exp_config)
ensure_success(exp)
assert isinstance(exp.export_top_models()[0], dict)
exp.stop()
val_dataloaders=pl.DataLoader(test_dataset,
batch_size=100),
max_epochs=1,
limit_train_batches=0.2)
applied_mutators = [BlockMutator('mutable_0'), BlockMutator('mutable_1')]
simple_strategy = TPEStrategy()
exp = RetiariiExperiment(base_model, trainer, applied_mutators,
simple_strategy)
exp_config = RetiariiExeConfig('remote')
exp_config.experiment_name = 'darts_search'
exp_config.trial_concurrency = 3
exp_config.max_trial_number = 10
exp_config.trial_gpu_number = 1
exp_config.training_service.reuse_mode = True
rm_conf = RemoteMachineConfig()
rm_conf.host = '127.0.0.1'
rm_conf.user = '******'
rm_conf.password = '******'
rm_conf.port = 22
rm_conf.python_path = '/home/xxx/py38/bin'
rm_conf.gpu_indices = [0, 1, 2]
rm_conf.use_active_gpu = True
rm_conf.max_trial_number_per_gpu = 3
exp_config.training_service.machine_list = [rm_conf]
exp_config.execution_engine = CgoEngineConfig(max_concurrency_cgo=3,
transform=transform)
trainer = pl.Classification(train_dataloader=pl.DataLoader(train_dataset,
batch_size=100),
val_dataloaders=pl.DataLoader(test_dataset,
batch_size=100),
max_epochs=2,
gpus=1,
limit_train_batches=0.1,
limit_val_batches=0.1)
simple_strategy = strategy.Random()
exp = RetiariiExperiment(base_model, trainer, [], simple_strategy)
exp_config = RetiariiExeConfig('local')
exp_config.experiment_name = 'mnist_search'
exp_config.trial_concurrency = 2
exp_config.max_trial_number = args.budget
exp_config.trial_gpu_number = 1
exp_config.training_service.use_active_gpu = True # Integration test GPU has a Xorg running
export_formatter = 'dict'
if args.exec == 'graph':
exp_config.execution_engine = 'base'
export_formatter = 'code'
exp.run(exp_config, args.port)
print('Final model:')
for model_code in exp.export_top_models(formatter=export_formatter):
print(model_code)
# In future, we will support mutation on the arguments of evaluators, which is commonly called "Hyper-parmeter tuning".
#
# Launch an Experiment
# --------------------
#
# After all the above are prepared, it is time to start an experiment to do the model search. An example is shown below.
from nni.retiarii.experiment.pytorch import RetiariiExperiment, RetiariiExeConfig
exp = RetiariiExperiment(model_space, evaluator, [], search_strategy)
exp_config = RetiariiExeConfig('local')
exp_config.experiment_name = 'mnist_search'
# %%
# The following configurations are useful to control how many trials to run at most / at the same time.
exp_config.max_trial_number = 4 # spawn 4 trials at most
exp_config.trial_concurrency = 2 # will run two trials concurrently
# %%
# Remember to set the following config if you want to GPU.
# ``use_active_gpu`` should be set true if you wish to use an occupied GPU (possibly running a GUI).
exp_config.trial_gpu_number = 1
exp_config.training_service.use_active_gpu = True
# %%
# Launch the experiment. The experiment should take several minutes to finish on a workstation with 2 GPUs.
exp.run(exp_config, 8081)
# %%
def _main():
parser = argparse.ArgumentParser("SPOS Evolutional Search")
parser.add_argument("--port", type=int, default=8084)
parser.add_argument("--imagenet-dir", type=str, default="./data/imagenet")
parser.add_argument("--checkpoint",
type=str,
default="./data/checkpoint-150000.pth.tar")
parser.add_argument(
"--spos-preprocessing",
action="store_true",
default=False,
help="When true, image values will range from 0 to 255 and use BGR "
"(as in original repo).")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--workers", type=int, default=6)
parser.add_argument("--train-batch-size", type=int, default=128)
parser.add_argument("--train-iters", type=int, default=200)
parser.add_argument("--test-batch-size", type=int, default=512)
parser.add_argument("--log-frequency", type=int, default=10)
parser.add_argument("--label-smoothing", type=float, default=0.1)
parser.add_argument("--evolution-sample-size", type=int, default=10)
parser.add_argument("--evolution-population-size", type=int, default=50)
parser.add_argument("--evolution-cycles", type=int, default=10)
parser.add_argument(
"--latency-filter",
type=str,
default=None,
help="Apply latency filter by calling the name of the applied hardware."
)
parser.add_argument("--latency-threshold", type=float, default=100)
args = parser.parse_args()
# use a fixed set of image will improve the performance
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.deterministic = True
assert torch.cuda.is_available()
base_model = ShuffleNetV2OneShot()
criterion = CrossEntropyLabelSmooth(1000, args.label_smoothing)
if args.latency_filter:
latency_filter = LatencyFilter(threshold=args.latency_threshold,
predictor=args.latency_filter)
else:
latency_filter = None
evaluator = FunctionalEvaluator(evaluate_acc,
criterion=criterion,
args=args)
evolution_strategy = strategy.RegularizedEvolution(
model_filter=latency_filter,
sample_size=args.evolution_sample_size,
population_size=args.evolution_population_size,
cycles=args.evolution_cycles)
exp = RetiariiExperiment(base_model,
evaluator,
strategy=evolution_strategy)
exp_config = RetiariiExeConfig('local')
exp_config.trial_concurrency = 2
exp_config.trial_gpu_number = 1
exp_config.max_trial_number = args.evolution_cycles
exp_config.training_service.use_active_gpu = False
exp_config.execution_engine = 'base'
exp_config.dummy_input = [1, 3, 224, 224]
exp.run(exp_config, args.port)
print('Exported models:')
for i, model in enumerate(exp.export_top_models(formatter='dict')):
print(model)
with open(f'architecture_final_{i}.json', 'w') as f:
json.dump(get_archchoice_by_model(model), f, indent=4)