if __name__ == '__main__':
from argparse import ArgumentParser
parser = ArgumentParser()
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--initial_lr', type=float, default=0.1)
parser.add_argument('--n_layers', type=int, required=True)
parser.add_argument('--postfix', type=str, default='')
parser.add_argument('--sharing', type=bool, default=False)
args = parser.parse_args()
network = build_network(n_layers=args.n_layers)
from lr_scheduler import AtIterationScheduler
lr_table = {32000: args.initial_lr * 0.1, 48000: args.initial_lr * 0.01}
lr_scheduler = AtIterationScheduler(args.initial_lr, lr_table)
optimizer_settings = {
'args': {
'momentum': 0.9
},
'initial_lr': args.initial_lr,
'lr_scheduler': lr_scheduler,
'optimizer': 'SGD',
'weight_decay': 0.0001,
}
from mx_solver import MXSolver
from mx_initializer import PReLUInitializer
solver = MXSolver(
batch_size=args.batch_size,
from data_utilities import load_cifar10_record
from mx_layers import ReLU
from mx_solver import MXSolver
from drelu import drelu, DReLUInitializer
from nin import nin
ACTIVATE = sys.argv[1]
BATCH_SIZE = 128
if ACTIVATE == 'relu' : activate = ReLU
elif ACTIVATE == 'drelu': activate = lambda X : drelu(X, {'data' : (BATCH_SIZE, 3, 32, 32)})
network = nin(activate)
lr = 0.1
lr_table = {100000 : lr * 0.1}
lr_scheduler = AtIterationScheduler(lr, lr_table)
optimizer_settings = {
'args' : {'momentum' : 0.9},
'initial_lr' : lr,
'lr_scheduler' : lr_scheduler,
'optimizer' : 'SGD',
'weight_decay' : 0.0001,
}
solver = MXSolver(
batch_size = BATCH_SIZE,
devices = (0, 1, 2, 3),
epochs = 300,
initializer = DReLUInitializer(0, 1),
optimizer_settings = optimizer_settings,
symbol = network,
DATA_SHAPE = (BATCH_SIZE / len(devices), 3, 32, 32)
MODES = {'mode': 'hyper', 'embedding': 'feature_map', 'data_shape': DATA_SHAPE}
# MODES = {'mode' : 'hyper', 'embedding' : 'parameter'}
N = int(sys.argv[1])
network = triple_state_residual_network(N, **MODES)
data = load_cifar10_record(BATCH_SIZE)
lr = 0.1
lr_table = {32000: lr * 0.1, 48000: lr * 0.01}
optimizer_settings = {
'args': {
'momentum': 0.9
},
'initial_lr': lr,
'lr_scheduler': AtIterationScheduler(lr, lr_table),
'optimizer': 'SGD',
'weight_decay': 0.0001,
}
solver = MXSolver(
batch_size=BATCH_SIZE,
devices=devices,
epochs=150,
initializer=PReLUInitializer(),
optimizer_settings=optimizer_settings,
symbol=network,
verbose=True,
)
info = solver.train(data)