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 train(num_inputs, num_hiddens, num_outputs, train_iter, test_iter, eps,
drop_prob):
# epoch 25, loss 0.272, train acc 0.898, test acc 0.860, 467.3 examples/sec
# if eps = 1e-3, learning rate = 0.5
cnt = len(num_hiddens) + 1
odict = OrderedDict()
odict['flatten'] = custom.FlattenLayer()
odict['linear_0'] = nn.Linear(num_inputs, num_hiddens[0])
for i in range(1, cnt):
odict['relu_%d' % i] = nn.ReLU()
odict['dropout_%d' % i] = nn.Dropout(drop_prob[i - 1])
if i == cnt - 1:
odict['linear_%d' % i] = nn.Linear(num_hiddens[i - 1], num_outputs)
else:
odict['linear_%d' % i] = nn.Linear(num_hiddens[i - 1],
num_hiddens[i])
net = nn.Sequential(odict)
for params in net.parameters():
init.normal_(params, mean=0, std=.01)
loss = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(net.parameters(), lr=.5)
base.train(net,
train_iter,
test_iter,
loss,
eps=eps,
num_epochs=50,
optimizer=optimizer)
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)
def train(num_inputs, num_outputs, train_iter, test_iter, eps):
# epoch 22, loss 0.417, train acc 0.858, test acc 0.840, 2905.9 examples/sec
# if eps = 1e-3, learning rate = 0.1
net = nn.Sequential(
OrderedDict([
('flatten', custom.FlattenLayer()),
('linear', nn.Linear(num_inputs, num_outputs))
])
)
init.normal_(net.linear.weight, mean=0, std=.01)
init.constant_(net.linear.bias, val=0)
loss = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(net.parameters(), lr=.1)
base.train(net, train_iter, test_iter, loss, eps=eps, num_epochs=50, optimizer=optimizer)
return net
def train(num_inputs, num_hiddens, num_outputs, train_iter, test_iter, eps):
# epoch 46, loss 0.155, train acc 0.943, test acc 0.883, 546.2 examples/sec
# if eps = 1e-3, learning rate = 0.5
net = nn.Sequential(custom.FlattenLayer(),
nn.Linear(num_inputs, num_hiddens), nn.ReLU(),
nn.Linear(num_hiddens, num_outputs))
for params in net.parameters():
init.normal_(params, mean=0, std=.01)
loss = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(net.parameters(), lr=.5)
base.train(net,
train_iter,
test_iter,
loss,
eps=eps,
num_epochs=50,
optimizer=optimizer)