def train_with_data_aug(train_augs, test_augs, lr=0.001):
batch_size, ctx, net = 50, try_all_gpus(), d2l.resnet18(10) # 256
net.initialize(ctx=ctx, init=init.Xavier())
trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate':lr})
loss = gloss.SoftmaxCrossEntropyLoss()
train_iter = load_cifar10(True, train_augs, batch_size)
test_iter = load_cifar10(False, test_augs, batch_size)
train(train_iter, test_iter, net, loss, trainer, ctx, num_epochs=10)
def get_net(ctx):
num_classes = 10
net = resnet18(num_classes)
net.initialize(ctx=ctx, init=init.Xavier())
return net
def evaluate_accuracy(net, data_iter, ctx=[mx.cpu()]):
if isinstance(ctx, mx.Context):
ctx = [ctx]
acc_sum, n = 0, 0
for batch in data_iter:
Xs, ys, _ = _get_batch(batch, ctx)
for X, y in zip(Xs, ys):
acc_sum += (net(X).argmax(axis=1) == y).sum().copyto(mx.cpu())
n += y.size
acc_sum.wait_to_read()
return acc_sum.asscalar() / n
net = d2l.resnet18(10)
loss = gloss.SoftmaxCrossEntropyLoss()
ctx = try_all_gpus()
net.initialize(ctx=ctx, init=init.Xavier())
batch_size = 1000
train_iter = getCifar10(True, batch_size, aug_method)
test_iter = getCifar10(False, batch_size, no_aug)
lr = 0.01
trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': lr})
def train(net, train_iter, test_iter, num_epoches, ctxes, loss, trainer):
train_acc_sum, train_l_sum, acc_count, l_count = 0, 0, 0, 0
for epoch in range(num_epoches):
for i, batch in enumerate(train_iter):
Xs, ys, batch_size = _get_batch(batch, ctxes)