def run_experiment(batch_size):
for lr in np.logspace(-0.5, -1.5, 50):
schedule = resnet_paper_schedule(batch_size=batch_size, base_lr=lr)
sd = 0.5
model = ResNetCIFAR(n=20, stochastic_depth={'from': sd, 'to': sd})
train_on_cifar(model,
schedule,
name=f'ResNet(lr={lr:.2e}, sd={sd})',
batch_size=batch_size)
def run_experiment(batch_size):
for e in range(10, 210, 10):
schedule = resnet_paper_schedule(batch_size=batch_size, scale=e / 200)
sd = 0.5
model = ResNetCIFAR(n=20, stochastic_depth={'from': sd, 'to': sd})
train_on_cifar(model,
schedule,
name=f'ResNet(sd={sd}, epochs={e})',
batch_size=batch_size)
def run_experiment(batch_size, force_multi_gpu=False):
schedule = resnet_paper_schedule(batch_size=batch_size)
model = ResNetCIFAR(n=20)
train_on_cifar(model,
schedule,
name=f'ResNet()',
batch_size=batch_size,
force_multi_gpu=force_multi_gpu)
def run_experiment(batch_size):
schedule = resnet_paper_schedule(batch_size=batch_size)
for prob in [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]:
print(prob)
model = ResNetCIFAR(n=20, stochastic_depth={'from': prob, 'to': prob})
train_on_cifar(model,
schedule,
name=f'ResNet(sd-p={prob})',
batch_size=batch_size)
else:
splitted = sd.split('-')
stochastic_depth = {
'from': float(splitted[0]),
'to': float(splitted[1])
}
batch_size = args.batch_size
n = args.number
pre_activated = args.pre_activated
base_lr = args.learn_rate
show_test = args.show_test
epochs = args.epochs
loss_scale = args.loss_scale
schedule = resnet_paper_schedule(batch_size=batch_size,
yellow_fin=args.yellowfin,
scale=epochs / TOTAL_EPOCHS)
net = ResNetCIFAR(n,
pre_activated=pre_activated,
stochastic_depth=stochastic_depth)
name = f'ResNet({n}, lr={base_lr}, epochs={epochs}, pa={pre_activated}, sd={args.stochastic_depth}, hp={args.half_precision}, yf={args.yellowfin}), ls={args.loss_scale}'
train_on_cifar(net,
schedule,
batch_size=batch_size,
name=name,
show_test=show_test)
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""
import argparse
from fasttrain import train_on_cifar
from fasttrain.model.resnet import ResNetCIFAR
from fasttrain.schedules import resnet_paper_schedule
parser = argparse.ArgumentParser(description='Train ResNet on CIFAR10')
parser.add_argument('-n', '--number', type=int, default=20)
parser.add_argument('-b', '--batch-size', type=int, default=128)
args = parser.parse_args()
schedule = resnet_paper_schedule()
n = args.number
train_on_cifar(ResNetCIFAR(n),
schedule,
name=f'ResNet({n})-BaseLine',
batch_size=args.batch_size)