def train_models(configs, project_dir):
"""Run all the training experiments specified in configs."""
# download dataset
data_dir = os.path.join(project_dir, "data")
datasets.CIFAR10(data_dir, download=True, train=True)
# Run all experiments in serial
if len(configs) == 0:
print("No experiments to run!")
for exp in configs:
config = configs[exp]
if "name" not in config:
config["name"] = exp
# Make sure local directories are relative to the project location
path = config.get("path", "results")
if not os.path.isabs(path):
config["path"] = os.path.join(project_dir, path)
data_dir = config.get("data_dir", "data")
if not os.path.isabs(data_dir):
config["data_dir"] = os.path.join(project_dir, data_dir)
model = TinyCIFAR()
model.model_setup(config)
for epoch in range(config["iterations"]):
ret = model.train_epoch(epoch)
print("epoch=", epoch, ":", ret)
# if ret['stop'] == 1:
# print("Stopping early!")
# break
if config.get("checkpoint_at_end", False):
model.model_save(path)
def run_noise_tests(config, options, project_dir):
"""
Test a pretrained network, specified in this config against noisy vectors.
"""
# download dataset
data_dir = os.path.join(project_dir, "data")
_ = datasets.CIFAR10(data_dir, download=True, train=False)
# Make sure local directories are relative to the project location
path = config.get("path", "results")
if not os.path.isabs(path):
config["path"] = os.path.join(project_dir, path)
data_dir = config.get("data_dir", "data")
if not os.path.isabs(data_dir):
config["data_dir"] = os.path.join(project_dir, data_dir)
# Load the pretrained model
tiny_cifar = TinyCIFAR()
tiny_cifar.model_setup(config)
tiny_cifar.model_restore(options.checkpoint_path
or os.path.join(path, tiny_cifar.model_filename))
# Create new loaders with batch size 1
noise_values = [0.0, 0.025, 0.05, 0.075, 0.1, 0.125, 0.15, 0.175]
loaders = create_test_loaders(noise_values,
batch_size=64,
data_dir=tiny_cifar.data_dir)
print("Running full noise tests using noise values", noise_values)
ret = tiny_cifar.run_noise_tests(noise_values, loaders, 300)
print(ret)
def layer_noise_sensitivities(config, options, project_dir):
"""
Test a pretrained network, specified in this config against noisy vectors.
"""
# download dataset
data_dir = os.path.join(project_dir, "data")
_ = datasets.CIFAR10(data_dir, download=True, train=False)
# Make sure local directories are relative to the project location
path = config.get("path", "results")
if not os.path.isabs(path):
config["path"] = os.path.join(project_dir, path)
data_dir = config.get("data_dir", "data")
if not os.path.isabs(data_dir):
config["data_dir"] = os.path.join(project_dir, data_dir)
# Load the pretrained model
tiny_cifar = TinyCIFAR()
tiny_cifar.model_setup(config)
tiny_cifar.model_restore(options.checkpoint_path
or os.path.join(path, tiny_cifar.model_filename))
# Create new loaders with batch size 1
noise_values = [0.0, 0.05, 0.1, 0.15, 0.2]
loaders = create_test_loaders(noise_values,
batch_size=1,
data_dir=tiny_cifar.data_dir)
noise_datasets = [l.dataset for l in loaders]
# Test noise sensitivity for different layers
model = tiny_cifar.model
one_layer_noise_sensitivity(
model=tiny_cifar.model,
noise_datasets=noise_datasets,
layer_name="image",
num_images=options.num_images,
)
modules_to_check = [
"cnn_0_0",
"avgpool_0_0",
"kwinners_2d_0_0",
"cnn_1_0",
"avgpool_1_0",
"kwinners_2d_1_0",
"flatten",
"linear_0",
"kwinners_linear",
]
for m in modules_to_check:
if m in model._modules.keys():
one_layer_noise_sensitivity(
model=tiny_cifar.model,
noise_datasets=noise_datasets,
layer_name=m,
num_images=options.num_images,
)
def train_models(configs, project_dir):
"""Run all the training experiments specified in configs."""
# Default data dir if not specified
data_dir = os.path.join(os.environ["HOME"], "nta", "data")
# Run all experiments in serial
if len(configs) == 0:
print("No experiments to run!")
sys.exit()
for exp in configs:
config = configs[exp]
if "name" not in config:
config["name"] = exp
# Make sure local directories are relative to the project location
path = config.get("path", "results")
path = Path(path).expanduser().resolve()
config["path"] = path
data_dir = config.get("data_dir", data_dir)
data_dir = Path(data_dir).expanduser().resolve()
config["data_dir"] = data_dir
datasets.CIFAR10(data_dir, download=True, train=True)
model = TinyCIFAR()
model.model_setup(config)
for epoch in range(config["iterations"]):
ret = model.train_epoch(epoch)
print("epoch=", epoch, ":", ret)
# if ret['stop'] == 1:
# print("Stopping early!")
# break
if config.get("checkpoint_at_end", False):
model.model_save(path)
def __init__(self, config=None, logger_creator=None):
TinyCIFAR.__init__(self)
tune.Trainable.__init__(self,
config=config,
logger_creator=logger_creator)
def __init__(self, config=None, logger_creator=None): TinyCIFAR.__init__(self) tune.Trainable.__init__(self, config=config, logger_creator=logger_creator)