def __init__(self, n_in, n_units, n_out, gpu0, gpu1):
super(ParallelMLP, self).__init__(
first0=train_mnist.MLP(n_in, n_units // 2, n_units).to_gpu(gpu0),
first1=train_mnist.MLP(n_in, n_units // 2, n_units).to_gpu(gpu1),
second0=train_mnist.MLP(n_units, n_units // 2, n_out).to_gpu(gpu0),
second1=train_mnist.MLP(n_units, n_units // 2, n_out).to_gpu(gpu1),
)
self.gpu0 = gpu0
self.gpu1 = gpu1
def __init__(self, n_units, n_out, gpu0, gpu1):
super(ParallelMLP, self).__init__(
# the input size, 784, is inferred
first0=train_mnist.MLP(n_units // 2, n_units).to_gpu(gpu0),
first1=train_mnist.MLP(n_units // 2, n_units).to_gpu(gpu1),
# the input size, n_units, is inferred
second0=train_mnist.MLP(n_units // 2, n_out).to_gpu(gpu0),
second1=train_mnist.MLP(n_units // 2, n_out).to_gpu(gpu1),
)
self.gpu0 = gpu0
self.gpu1 = gpu1
def __init__(self, n_units, n_out, gpu0, gpu1):
super(ParallelMLP, self).__init__()
self.gpu0 = gpu0
self.gpu1 = gpu1
with self.init_scope():
# the input size, 784, is inferred
self.first0 = train_mnist.MLP(n_units // 2, n_units).to_gpu(gpu0)
self.first1 = train_mnist.MLP(n_units // 2, n_units).to_gpu(gpu1)
# the input size, n_units, is inferred
self.second0 = train_mnist.MLP(n_units // 2, n_out).to_gpu(gpu0)
self.second1 = train_mnist.MLP(n_units // 2, n_out).to_gpu(gpu1)
def __init__(self, n_units, n_out, device0, device1):
super(ParallelMLP, self).__init__()
self.device0 = device0
self.device1 = device1
with self.init_scope():
# the input size, 784, is inferred
self.first0 = train_mnist.MLP(n_units // 2, n_units)
self.first1 = train_mnist.MLP(n_units // 2, n_units)
self.first0.to_device(device0)
self.first1.to_device(device1)
# the input size, n_units, is inferred
self.second0 = train_mnist.MLP(n_units // 2, n_out)
self.second1 = train_mnist.MLP(n_units // 2, n_out)
self.second0.to_device(device0)
self.second1.to_device(device1)
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize',
'-b',
type=int,
default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=20,
help='Number of sweeps over the dataset to train')
parser.add_argument('--device',
'-d',
type=str,
default='-1',
help='Device specifier. Either ChainerX device '
'specifier or an integer. If non-negative integer, '
'CuPy arrays with specified device id are used. If '
'negative integer, NumPy arrays are used')
parser.add_argument('--out',
'-o',
default='result',
help='Directory to output the result')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot using model '
'and state files in the specified directory')
parser.add_argument('--unit',
'-u',
type=int,
default=1000,
help='Number of units')
group = parser.add_argument_group('deprecated arguments')
group.add_argument('--gpu',
'-g',
type=int,
nargs='?',
const=0,
help='GPU ID (negative value indicates CPU)')
args = parser.parse_args()
device = parse_device(args)
print('Device: {}'.format(device))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
# Set up a neural network to train
model = L.Classifier(train_mnist.MLP(args.unit, 10))
model.to_device(device)
device.use()
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
if args.resume:
# Resume from a snapshot
serializers.load_npz('{}/mlp.model'.format(args.resume), model)
serializers.load_npz('{}/mlp.state'.format(args.resume), optimizer)
# Load the MNIST dataset
train, test = chainer.datasets.get_mnist()
train_count = len(train)
test_count = len(test)
with SerialIterator(train, args.batchsize) as train_iter, \
SerialIterator(
test, args.batchsize, repeat=False, shuffle=False) as test_iter:
sum_accuracy = 0
sum_loss = 0
while train_iter.epoch < args.epoch:
batch = train_iter.next()
x, t = convert.concat_examples(batch, device)
loss = model(x, t) # type: chainer.Variable
model.cleargrads()
loss.backward()
optimizer.update()
del loss
sum_loss += float(model.loss.array) * len(t)
sum_accuracy += float(model.accuracy.array) * len(t)
if train_iter.is_new_epoch:
print('epoch: {}'.format(train_iter.epoch))
print('train mean loss: {}, accuracy: {}'.format(
sum_loss / train_count, sum_accuracy / train_count))
# evaluation
sum_accuracy = 0
sum_loss = 0
# Enable evaluation mode.
with configuration.using_config('train', False):
# This is optional but can reduce computational overhead.
with chainer.using_config('enable_backprop', False):
for batch in test_iter:
x, t = convert.concat_examples(batch, device)
loss = model(x, t)
sum_loss += float(loss.array) * len(t)
sum_accuracy += float(
model.accuracy.array) * len(t)
test_iter.reset()
print('test mean loss: {}, accuracy: {}'.format(
sum_loss / test_count, sum_accuracy / test_count))
sum_accuracy = 0
sum_loss = 0
# Save the model and the optimizer
print('save the model')
serializers.save_npz('{}/mlp.model'.format(args.out), model)
print('save the optimizer')
serializers.save_npz('{}/mlp.state'.format(args.out), optimizer)
def main():
# This script is almost identical to train_mnist.py. The only difference is
# that this script uses data-parallel computation on two GPUs.
# See train_mnist.py for more details.
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize',
'-b',
type=int,
default=400,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=20,
help='Number of sweeps over the dataset to train')
parser.add_argument('--out',
'-o',
default='result_data_parallel',
help='Directory to output the result')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot')
parser.add_argument('--unit',
'-u',
type=int,
default=1000,
help='Number of units')
parser.add_argument('--device0',
'-d',
type=str,
default='0',
help='Device specifier of the first device. '
'Either ChainerX device '
'specifier or an integer. If non-negative integer, '
'CuPy arrays with specified device id are used. If '
'negative integer, NumPy arrays are used')
parser.add_argument('--device1',
'-D',
type=str,
default='1',
help='Device specifier of the second device. '
'Either ChainerX device '
'specifier or an integer. If non-negative integer, '
'CuPy arrays with specified device id are used. If '
'negative integer, NumPy arrays are used')
group = parser.add_argument_group('deprecated arguments')
group.add_argument('--gpu0',
'-g',
dest='device0',
type=int,
nargs='?',
const=0,
help='First GPU ID')
group.add_argument('--gpu1',
'-G',
dest='device1',
type=int,
nargs='?',
const=1,
help='Second GPU ID')
args = parser.parse_args()
device0 = chainer.get_device(args.device0)
device1 = chainer.get_device(args.device1)
print('Devices: {}, {}'.format(device0, device1))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
device0.use()
model = L.Classifier(train_mnist.MLP(args.unit, 10))
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
train, test = chainer.datasets.get_mnist()
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
# ParallelUpdater implements the data-parallel gradient computation on
# multiple devices. It accepts "devices" argument that specifies which
# device to use.
updater = training.updaters.ParallelUpdater(
train_iter,
optimizer,
# The device of the name 'main' is used as a "master", while others are
# used as slaves. Names other than 'main' are arbitrary.
devices={
'main': device0,
'second': device1
},
)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), out=args.out)
trainer.extend(extensions.Evaluator(test_iter, model, device=device0))
# TODO(niboshi): Temporarily disabled for chainerx. Fix it.
if device0.xp is not chainerx:
trainer.extend(extensions.DumpGraph('main/loss'))
trainer.extend(extensions.snapshot(), trigger=(args.epoch, 'epoch'))
trainer.extend(extensions.LogReport())
trainer.extend(
extensions.PrintReport([
'epoch', 'main/loss', 'validation/main/loss', 'main/accuracy',
'validation/main/accuracy', 'elapsed_time'
]))
trainer.extend(extensions.ProgressBar())
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
trainer.run()
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize', '-b', type=int, default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch', '-e', type=int, default=20,
help='Number of sweeps over the dataset to train')
parser.add_argument('--device', '-d', type=str, default='-1',
help='Device specifier. Either ChainerX device '
'specifier or an integer. If non-negative integer, '
'CuPy arrays with specified device id are used. If '
'negative integer, NumPy arrays are used')
parser.add_argument('--out', '-o', default='result',
help='Directory to output the result')
parser.add_argument('--model', '-m', default='MLP',
help='Choose the model: MLP or MLPSideEffect')
parser.add_argument('--resume', '-r', default='',
help='Resume the training from snapshot')
parser.add_argument('--unit', '-u', type=int, default=1000,
help='Number of units')
group = parser.add_argument_group('deprecated arguments')
group.add_argument('--gpu', '-g', dest='device',
type=int, nargs='?', const=0,
help='GPU ID (negative value indicates CPU)')
args = parser.parse_args()
if chainer.get_dtype() == numpy.float16:
warnings.warn(
'This example may cause NaN in FP16 mode.', RuntimeWarning)
device = chainer.get_device(args.device)
print('Device: {}'.format(device))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
device.use()
# Set up a neural network to train
if args.model == 'MLP':
model = L.Classifier(train_mnist.MLP(args.unit, 10))
elif args.model == 'MLPSideEffect':
model = L.Classifier(train_mnist.MLPSideEffect(args.unit, 10))
model.to_device(device)
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
# Load the MNIST dataset
train, test = chainer.datasets.get_mnist()
train_count = len(train)
test_count = len(test)
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(
test, args.batchsize, repeat=False, shuffle=False)
if device.xp is not chainerx:
run_train_loop(
optimizer, train_iter, test_iter, train_count, test_count,
args.epoch, device)
else:
warnings.warn(
'Static subgraph optimization does not support ChainerX and will'
' be disabled.', UserWarning)
with chainer.using_config('use_static_graph', False):
run_train_loop(
optimizer, train_iter, test_iter, train_count, test_count,
args.epoch, device)
# Save the model and the optimizer
print('save the model')
serializers.save_npz('mlp.model', model)
print('save the optimizer')
serializers.save_npz('mlp.state', optimizer)
def main():
# This script is almost identical to train_mnist.py. The only difference is
# that this script uses data-parallel computation on two GPUs.
# See train_mnist.py for more details.
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize',
'-b',
type=int,
default=400,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=20,
help='Number of sweeps over the dataset to train')
parser.add_argument('--gpu0',
'-g',
type=int,
default=0,
help='First GPU ID')
parser.add_argument('--gpu1',
'-G',
type=int,
default=1,
help='Second GPU ID')
parser.add_argument('--out',
'-o',
default='result_parallel',
help='Directory to output the result')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot')
parser.add_argument('--unit',
'-u',
type=int,
default=1000,
help='Number of units')
args = parser.parse_args()
print('GPU: {}, {}'.format(args.gpu0, args.gpu1))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
chainer.backends.cuda.get_device_from_id(args.gpu0).use()
model = L.Classifier(train_mnist.MLP(args.unit, 10))
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
train, test = chainer.datasets.get_mnist()
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
# ParallelUpdater implements the data-parallel gradient computation on
# multiple GPUs. It accepts "devices" argument that specifies which GPU to
# use.
updater = training.updaters.ParallelUpdater(
train_iter,
optimizer,
# The device of the name 'main' is used as a "master", while others are
# used as slaves. Names other than 'main' are arbitrary.
devices={
'main': args.gpu0,
'second': args.gpu1
},
)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), out=args.out)
trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu0))
trainer.extend(extensions.dump_graph('main/loss'))
trainer.extend(extensions.snapshot(), trigger=(args.epoch, 'epoch'))
trainer.extend(extensions.LogReport())
trainer.extend(
extensions.PrintReport([
'epoch', 'main/loss', 'validation/main/loss', 'main/accuracy',
'validation/main/accuracy', 'elapsed_time'
]))
trainer.extend(extensions.ProgressBar())
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
trainer.run()
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize',
'-b',
type=int,
default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=20,
help='Number of sweeps over the dataset to train')
parser.add_argument('--gpu',
'-g',
type=int,
default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--out',
'-o',
default='result',
help='Directory to output the result')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot using model '
'and state files in the specified directory')
parser.add_argument('--unit',
'-u',
type=int,
default=1000,
help='Number of units')
args = parser.parse_args()
print('GPU: {}'.format(args.gpu))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
# Set up a neural network to train
model = L.Classifier(train_mnist.MLP(args.unit, 10))
if args.gpu >= 0:
# Make a speciied GPU current
chainer.backends.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
if args.resume:
# Resume from a snapshot
serializers.load_npz('{}/mlp.model'.format(args.resume), model)
serializers.load_npz('{}/mlp.state'.format(args.resume), optimizer)
# Load the MNIST dataset
train, test = chainer.datasets.get_mnist()
train_count = len(train)
test_count = len(test)
with MultiprocessIterator(train, args.batchsize) as train_iter, \
MultiprocessIterator(test, args.batchsize,
repeat=False, shuffle=False) as test_iter:
sum_accuracy = 0
sum_loss = 0
while train_iter.epoch < args.epoch:
batch = train_iter.next()
x_array, t_array = convert.concat_examples(batch, args.gpu)
x = chainer.Variable(x_array)
t = chainer.Variable(t_array)
optimizer.update(model, x, t)
sum_loss += float(model.loss.data) * len(t.data)
sum_accuracy += float(model.accuracy.data) * len(t.data)
if train_iter.is_new_epoch:
print('epoch: {}'.format(train_iter.epoch))
print('train mean loss: {}, accuracy: {}'.format(
sum_loss / train_count, sum_accuracy / train_count))
# evaluation
sum_accuracy = 0
sum_loss = 0
for batch in test_iter:
x_array, t_array = convert.concat_examples(batch, args.gpu)
x = chainer.Variable(x_array)
t = chainer.Variable(t_array)
loss = model(x, t)
sum_loss += float(loss.data) * len(t.data)
sum_accuracy += float(model.accuracy.data) * len(t.data)
test_iter.reset()
print('test mean loss: {}, accuracy: {}'.format(
sum_loss / test_count, sum_accuracy / test_count))
sum_accuracy = 0
sum_loss = 0
# Save the model and the optimizer
print('save the model')
serializers.save_npz('{}/mlp.model'.format(args.out), model)
print('save the optimizer')
serializers.save_npz('{}/mlp.state'.format(args.out), optimizer)
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize', '-b', type=int, default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch', '-e', type=int, default=20,
help='Number of sweeps over the dataset to train')
parser.add_argument('--device', '-d', type=str, default='-1',
help='Device specifier. Either ChainerX device '
'specifier or an integer. If non-negative integer, '
'CuPy arrays with specified device id are used. If '
'negative integer, NumPy arrays are used')
parser.add_argument('--out', '-o', default='result',
help='Directory to output the result')
parser.add_argument('--model', '-m', default='MLP',
help='Choose the model: MLP or MLPSideEffect')
parser.add_argument('--resume', '-r', default='',
help='Resume the training from snapshot')
parser.add_argument('--unit', '-u', type=int, default=1000,
help='Number of units')
group = parser.add_argument_group('deprecated arguments')
group.add_argument('--gpu', '-g', dest='device',
type=int, nargs='?', const=0,
help='GPU ID (negative value indicates CPU)')
args = parser.parse_args()
if chainer.get_dtype() == numpy.float16:
warnings.warn(
'This example may cause NaN in FP16 mode.', RuntimeWarning)
device = chainer.get_device(args.device)
if device.xp is chainerx:
sys.stderr.write('This example does not support ChainerX devices.\n')
sys.exit(1)
print('Device: {}'.format(device))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
device.use()
# Set up a neural network to train
if args.model == 'MLP':
model = L.Classifier(train_mnist.MLP(args.unit, 10))
elif args.model == 'MLPSideEffect':
model = L.Classifier(train_mnist.MLPSideEffect(args.unit, 10))
model.to_device(device)
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
# Load the MNIST dataset
train, test = chainer.datasets.get_mnist()
train_count = len(train)
test_count = len(test)
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test, args.batchsize,
repeat=False, shuffle=False)
sum_accuracy = 0
sum_loss = 0
while train_iter.epoch < args.epoch:
batch = train_iter.next()
x_array, t_array = convert.concat_examples(batch, device)
x = chainer.Variable(x_array)
t = chainer.Variable(t_array, requires_grad=False)
optimizer.update(model, x, t)
sum_loss += float(model.loss.array) * len(t)
sum_accuracy += float(model.accuracy.array) * len(t)
if train_iter.is_new_epoch:
print('epoch: ', train_iter.epoch)
print('train mean loss: {}, accuracy: {}'.format(
sum_loss / train_count, sum_accuracy / train_count))
# evaluation
sum_accuracy = 0
sum_loss = 0
# It is good practice to turn off train mode during evaluation.
with configuration.using_config('train', False):
for batch in test_iter:
x_array, t_array = convert.concat_examples(batch, device)
x = chainer.Variable(x_array)
t = chainer.Variable(t_array)
loss = model(x, t)
sum_loss += float(loss.array) * len(t)
sum_accuracy += float(model.accuracy.array) * len(t)
test_iter.reset()
print('test mean loss: {}, accuracy: {}'.format(
sum_loss / test_count, sum_accuracy / test_count))
sum_accuracy = 0
sum_loss = 0
# Save the model and the optimizer
print('save the model')
serializers.save_npz('mlp.model', model)
print('save the optimizer')
serializers.save_npz('mlp.state', optimizer)
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize',
'-b',
type=int,
default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=20,
help='Number of sweeps over the dataset to train')
parser.add_argument('--gpu',
'-g',
type=int,
default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--out',
'-o',
default='result',
help='Directory to output the result')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot')
parser.add_argument('--unit',
'-u',
type=int,
default=1000,
help='Number of units')
args = parser.parse_args()
print('GPU: {}'.format(args.gpu))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
# Set up a neural network to train
model = L.Classifier(train_mnist.MLP(args.unit, 10))
if args.gpu >= 0:
# Make a speciied GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
# Load the MNIST dataset
train, test = chainer.datasets.get_mnist()
train_count = len(train)
test_count = len(test)
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
sum_accuracy = 0
sum_loss = 0
# [ChainerUI] setup log reporter to show on ChainerUI
ui_report = LogReport(args.out, conditions=args)
while train_iter.epoch < args.epoch:
batch = train_iter.next()
x_array, t_array = convert.concat_examples(batch, args.gpu)
x = chainer.Variable(x_array)
t = chainer.Variable(t_array)
optimizer.update(model, x, t)
sum_loss += float(model.loss.data) * len(t.data)
sum_accuracy += float(model.accuracy.data) * len(t.data)
if train_iter.is_new_epoch:
print('epoch: ', train_iter.epoch)
train_loss = sum_loss / train_count
train_accuracy = sum_accuracy / train_count
print('train mean loss: {}, accuracy: {}'.format(
train_loss, train_accuracy))
# evaluation
sum_accuracy = 0
sum_loss = 0
for batch in test_iter:
x_array, t_array = convert.concat_examples(batch, args.gpu)
x = chainer.Variable(x_array)
t = chainer.Variable(t_array)
loss = model(x, t)
sum_loss += float(loss.data) * len(t.data)
sum_accuracy += float(model.accuracy.data) * len(t.data)
test_iter.reset()
test_loss = sum_loss / test_count
test_accuracy = sum_accuracy / test_count
print('test mean loss: {}, accuracy: {}'.format(
test_loss, test_accuracy))
# [ChainerUI] write values to 'log' file
stats = {
'epoch': train_iter.epoch,
'iteration': train_iter.epoch * args.batchsize,
'train/loss': train_loss,
'train/accuracy': train_accuracy,
'test/loss': test_loss,
'test/accuracy': test_accuracy
}
ui_report(stats)
sum_accuracy = 0
sum_loss = 0
# Save the model and the optimizer
print('save the model')
serializers.save_npz('mlp.model', model)
print('save the optimizer')
serializers.save_npz('mlp.state', optimizer)
