def get_mnist_iterator():
train_set = MNIST(train=True, transform=dummy_transform)
train_iter = gluon.data.DataLoader(train_set, args.batch_size, True, num_workers=args.j, last_batch='discard')
val_set = MNIST(train=False, transform=dummy_transform)
val_iter = gluon.data.DataLoader(val_set, args.batch_size, False, num_workers=args.j)
return train_iter, val_iter, len(train_set)
def main():
ctx = mx.cpu()
batch_size = 1024
random.seed(47)
mnist_train = MNIST(train=True) # 加载训练
tr_data = mnist_train._data.reshape((-1, 28 * 28)) # 数据
tr_label = mnist_train._label # 标签
mnist_test = MNIST(train=False) # 加载测试
te_data = mnist_test._data.reshape((-1, 28 * 28)) # 数据
te_label = mnist_test._label # 标签
def transform(data_, label_):
return data_.astype(np.float32) / 255., label_.astype(np.float32)
train_data = DataLoader(
TripletDataset(rd=tr_data, rl=tr_label, transform=transform),
batch_size, shuffle=True)
test_data = DataLoader(
TripletDataset(rd=te_data, rl=te_label, transform=transform),
batch_size, shuffle=True)
base_net = Sequential()
with base_net.name_scope():
base_net.add(Dense(256, activation='relu'))
base_net.add(Dense(128, activation='relu'))
base_net.collect_params().initialize(mx.init.Uniform(scale=0.1), ctx=ctx)
triplet_loss = gluon.loss.TripletLoss() # TripletLoss损失函数
trainer_triplet = gluon.Trainer(base_net.collect_params(), 'sgd', {'learning_rate': 0.05})
for epoch in range(10):
curr_loss = 0.0
for i, (data, _) in enumerate(train_data):
data = data.as_in_context(ctx)
anc_ins, pos_ins, neg_ins = data[:, 0], data[:, 1], data[:, 2]
with autograd.record():
inter1 = base_net(anc_ins)
inter2 = base_net(pos_ins)
inter3 = base_net(neg_ins)
loss = triplet_loss(inter1, inter2, inter3) # Triplet Loss
loss.backward()
trainer_triplet.step(batch_size)
curr_loss = mx.nd.mean(loss).asscalar()
# print('Epoch: %s, Batch: %s, Triplet Loss: %s' % (epoch, i, curr_loss))
print('Epoch: %s, Triplet Loss: %s' % (epoch, curr_loss))
evaluate_net(base_net, test_data, ctx=ctx)
# 数据可视化
te_data, te_label = transform(te_data, te_label)
tb_projector(te_data.asnumpy(), te_label, os.path.join(ROOT_DIR, 'logs', 'origin'))
te_res = base_net(te_data)
tb_projector(te_res.asnumpy(), te_label, os.path.join(ROOT_DIR, 'logs', 'triplet'))
def getData():
# Download and load the training data
trainSet: MNIST = MNIST('~/.mxnet/MNIST_data/',
train=True).transform_first(preprocessor)
trainLoader: DataLoader = DataLoader(trainSet,
batch_size=64,
shuffle=False,
last_batch='keep')
return trainLoader
def load_data(self):
"""
加载MNIST数据集
"""
def transform(data, label):
return data.astype(np.float32) / 255., label.astype(np.float32)
# 训练和测试数据
train_data = DataLoader(MNIST(train=True, transform=transform),
self.batch_size, shuffle=True)
valid_data = DataLoader(MNIST(train=False, transform=transform),
self.batch_size, shuffle=False)
# mnist = get_mnist()
# train_data = NDArrayIter(mnist["train_data"], mnist["train_label"], self.batch_size)
# valid_data = NDArrayIter(mnist["test_data"], mnist["test_label"], self.batch_size)
return train_data, valid_data
def data_loader(train, batch_size, num_workers):
dataset = MNIST(train=train, transform=transform)
return DataLoader(dataset,
batch_size,
shuffle=train,
num_workers=num_workers)
def train():
epochs = 100
lr = 0.01
lr_steps = [40, 70, np.inf]
momentum = 0.9
wd = 5e-4
plot_period = 5
ctx = [mx.gpu(i) for i in range(2)]
batch_size = 64
scale = 10
margin = 0.2
train_set = MNIST(train=True, transform=transform_train)
train_data = gluon.data.DataLoader(train_set,
batch_size,
True,
num_workers=4,
last_batch='discard')
val_set = MNIST(train=False, transform=transform_val)
val_data = gluon.data.DataLoader(val_set,
batch_size,
shuffle=False,
num_workers=4)
net = LeNetPlus(embedding_size=2, feature_norm=True, weight_norm=True)
net.initialize(init=mx.init.MSRAPrelu(), ctx=ctx)
# net.load_parameters("./pretrained_mnist.params", ctx=ctx)
net.hybridize()
loss = CosLoss(classes=10, s=scale, m=margin)
train_params = net.collect_params()
trainer = gluon.Trainer(train_params, 'sgd', {
'learning_rate': lr,
'momentum': momentum,
'wd': wd
})
lr_counter = 0
metric = mtc.Accuracy()
num_batch = len(train_data)
for epoch in range(epochs + 1):
if epoch == lr_steps[lr_counter]:
trainer.set_learning_rate(trainer.learning_rate * 0.1)
lr_counter += 1
if (epoch % plot_period) == 0:
plot = True
else:
plot = False
train_loss = 0
metric.reset()
tic = time.time()
ebs = []
lbs = []
for batch in train_data:
data = gluon.utils.split_and_load(batch[0],
ctx_list=ctx,
batch_axis=0,
even_split=False)
labels = gluon.utils.split_and_load(batch[1],
ctx_list=ctx,
batch_axis=0,
even_split=False)
with ag.record():
ots = [net(X) for X in data]
embedds = [ot[0] for ot in ots]
outputs = [ot[1] for ot in ots]
losses = [loss(yhat, y) for yhat, y in zip(outputs, labels)]
for l in losses:
ag.backward(l)
if plot:
for es, ls in zip(embedds, labels):
assert len(es) == len(ls)
for idx in range(len(es)):
ebs.append(es[idx].asnumpy())
lbs.append(ls[idx].asscalar())
trainer.step(batch_size)
metric.update(labels, outputs)
train_loss += sum([l.mean().asscalar()
for l in losses]) / len(losses)
_, train_acc = metric.get()
train_loss /= num_batch
val_acc, val_loss, val_ebs, val_lbs = validate(net, val_data, ctx,
loss, plot)
if plot:
ebs = np.vstack(ebs)
lbs = np.hstack(lbs)
plot_result(
ebs, lbs,
os.path.join("../../resources",
"cosloss-train-epoch{}.png".format(epoch)))
plot_result(
val_ebs, val_lbs,
os.path.join("../../resources",
"cosloss-val-epoch{}.png".format(epoch)))
toc = time.time()
print('[epoch % 3d] train accuracy: %.6f, train loss: %.6f | '
'val accuracy: %.6f, val loss: %.6f, time: %.6f' %
(epoch, train_acc, train_loss, val_acc, val_loss, toc - tic))
def train():
epochs = 101
lr = 0.1
momentum = 0.9
wd = 5e-4
plot_period = 20
ctx = [mx.gpu(i) for i in range(2)]
batch_size = 256
train_set = MNIST(train=True, transform=transform_train)
train_data = gluon.data.DataLoader(train_set, batch_size, True, num_workers=4, last_batch='discard')
val_set = MNIST(train=False, transform=transform_val)
val_data = gluon.data.DataLoader(val_set, batch_size, shuffle=False, num_workers=4)
net = MnistNet(embedding_size=2)
net.initialize(init=mx.init.MSRAPrelu(), ctx=ctx)
net.hybridize()
loss = CenterLoss(10, 2, 1)
loss.initialize(ctx=ctx)
num_batches = len(train_set) // batch_size
train_params = net.collect_params()
train_params.update(loss.params)
lr_scheduler = LRScheduler("cosine", lr, niters=num_batches, nepochs=epochs, targetlr=1e-8,
warmup_epochs=10, warmup_lr=0.001)
trainer = gluon.Trainer(train_params, 'nag', {'lr_scheduler': lr_scheduler, 'momentum': momentum, 'wd': wd})
metric = mtc.Accuracy()
num_batch = len(train_data)
for epoch in range(epochs):
plot = True if (epoch % plot_period) == 0 else False
train_loss = 0
metric.reset()
tic = time.time()
ebs, lbs = [], []
for i, batch in enumerate(train_data):
data = gluon.utils.split_and_load(batch[0], ctx_list=ctx, batch_axis=0, even_split=False)
labels = gluon.utils.split_and_load(batch[1], ctx_list=ctx, batch_axis=0, even_split=False)
with ag.record():
ots = [net(X) for X in data]
embedds = [ot[0] for ot in ots]
outputs = [ot[1] for ot in ots]
losses = [loss(yhat, y, emb) for yhat, y, emb in zip(outputs, labels, embedds)]
for l in losses:
ag.backward(l)
if plot:
for es, ls in zip(embedds, labels):
assert len(es) == len(ls)
for idx in range(len(es)):
ebs.append(es[idx].asnumpy())
lbs.append(ls[idx].asscalar())
lr_scheduler.update(i, epoch)
trainer.step(batch_size)
metric.update(labels, outputs)
train_loss += sum([l.mean().asscalar() for l in losses]) / len(losses)
_, train_acc = metric.get()
train_loss /= num_batch
val_acc, val_loss, val_ebs, val_lbs = validate(net, val_data, ctx, loss, plot)
toc = time.time()
print('[epoch % 3d] train accuracy: %.6f, train loss: %.6f | '
'val accuracy: %.6f, val loss: %.6f, time: %.6f'
% (epoch, train_acc, train_loss, val_acc, val_loss, toc - tic))
if plot:
ebs, lbs = np.vstack(ebs), np.hstack(lbs)
plot_result(ebs, lbs, os.path.join("../../resources", "center-train-epoch{}.png".format(epoch)))
plot_result(val_ebs, val_lbs, os.path.join("../../resources", "center-val-epoch{}.png".format(epoch)))
gluon.data.vision.transforms.Resize(opt.imageSize, keep_ratio=True, interpolation=3),
transforms.CenterCrop(opt.imageSize),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]))
elif opt.dataset == 'cifar10':
dataset = CIFAR10(root=opt.dataroot, train=True).transform_first(transforms.Compose([
gluon.data.vision.transforms.Resize(opt.imageSize, keep_ratio=True, interpolation=3),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
)
elif opt.dataset == 'mnist':
dataset = MNIST(root=opt.dataroot, train=True).transform_first(transforms.Compose([
gluon.data.vision.transforms.Resize(opt.imageSize, keep_ratio=True, interpolation=3),
transforms.ToTensor(),
transforms.Normalize((0.5), (0.5)),
]))
elif opt.dataset == 'rem_face':
dataset, _ = load_rem_face()
dataset = dataset.transform_first(transforms.Compose([
gluon.data.vision.transforms.Resize(opt.imageSize, keep_ratio=True, interpolation=3),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
assert dataset
dataloader = DataLoader(dataset, batch_size=opt.batchSize,
shuffle=True, num_workers=int(opt.workers))
print('finish init dataloader')
ngpu = int(opt.ngpu)
nz = int(opt.nz)
import mxnet.gluon as gl
from mxnet import autograd, image
from mxnet.gluon import nn
from mxnet.gluon.data import DataLoader
from mxnet.gluon.data.vision import transforms, MNIST
from mxnet.gluon.nn import Sequential
from mxnet.ndarray import NDArray
transform: Sequential = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=0.5, std=0.5)
])
# Download and load the training data
trainSet: MNIST = MNIST('~/.mxnet/MNIST_data/', train=True).transform_first(transform)
trainLoader: DataLoader = DataLoader(trainSet, batch_size=64, shuffle=True)
# Build a feed-forward network
model = nn.Sequential()
# with model.name_scope():
model.add(
nn.Dense(128, activation='relu'),
# nn.Activation('relu'),
nn.Dense(64, activation='relu'),
nn.Dense(10)
)
model.initialize()
criterion = gl.loss.SoftmaxCrossEntropyLoss()
optimizer = gl.Trainer(model.collect_params(), 'sgd', {'learning_rate': 0.01})