def main():
batch_size = 128
lr, eps = .0001, 1e-3
# lr, eps = .0001, 1e-1
root = '../../Datasets'
train_iter, test_iter = fmnist.load_data(batch_size, 224, root)
net = custom.VGG11(4)
optimizer = torch.optim.Adam(net.parameters(), lr)
ckpt_path = '../../checkpoint/vgg.pt'
ckpt = None
if os.path.exists(ckpt_path):
ckpt = torch.load(ckpt_path)
net.load_state_dict(ckpt['net'])
optimizer.load_state_dict(ckpt['optimizer'])
loss = nn.CrossEntropyLoss()
base.train(net,
train_iter,
test_iter,
loss,
eps=eps,
optimizer=optimizer,
checkpoint_path=ckpt_path,
checkpoint=ckpt)
def main():
batch_size = 256
num_inputs, num_outputs = 784, 10
eps = 1e-3
# eps = 1e-1
root = '../../Datasets'
train_iter, test_iter = fmnist.load_data(batch_size, root=root)
X, y = next(iter(test_iter))
true_labels = fmnist.get_labels(y.numpy())
net = train(num_inputs, num_outputs, train_iter, test_iter, eps)
pred_labels = fmnist.get_labels(net(X).argmax(dim=1).numpy())
titles = []
for true, pred in zip(true_labels, pred_labels):
title = true + '\n'
if true != pred:
title += 'X: ' + pred
titles.append(title)
fmnist.show(X, nrows=5, ncols=10, titles=titles)
def main():
batch_size = 256
num_inputs, num_outputs, num_hiddens = 784, 10, 256
eps = 1e-3
# eps = 1e-1
root = '../../Datasets'
train_iter, test_iter = fmnist.load_data(batch_size, root=root)
train(num_inputs, num_hiddens, num_outputs, train_iter, test_iter, eps)
def main():
batch_size = 256
lr, eps = .001, 1e-3
# lr, eps = .001, 1e-1
root = '../../Datasets'
train_iter, test_iter = fmnist.load_data(batch_size, root=root)
net = custom.LeNet()
optimizer = torch.optim.Adam(net.parameters(), lr)
loss = nn.CrossEntropyLoss()
base.train(net,
train_iter,
test_iter,
loss,
eps=eps,
num_epochs=50,
optimizer=optimizer)
