2019년 3월 12일
하이퍼파라미터를 변경하며 모델을 반복해서 만드는 일은 단순하지만, 귀찮은 일입니다. 파라미터를 전달하여 파라미터 구성을 달리 할 수 있는 다음은 ConvNet 모델입니다.
class Conv2dClassifier(nn.Module):
def __init__(self, features, fcdims, conv_kernel_size=3, conv_stride=1, pool_kernel_size=2, pool_stride=2):
super(Conv2dClassifier, self).__init__()
self.pool_kernel_size = pool_kernel_size
self.pool_stride = pool_stride
self.convs = []
for i in range(len(features)-1):
name = 'conv'+str(i+1)
setattr(self, name, nn.Conv2d(features[i], features[i+1], conv_kernel_size, conv_stride))
self.convs.extend([
getattr(self, name)
])
self.fcdim0 = fcdims[0]
self.fcs = []
for i in range(len(fcdims)-1):
name = 'fc'+str(i+1)
setattr(self, name, nn.Linear(fcdims[i], fcdims[i+1]))
self.fcs.extend([
getattr(self, name)
])
def forward(self, x):
for i in range(len(self.convs)):
x = self.convs[i](x)
x = F.relu(x)
x = F.max_pool2d(x, self.pool_kernel_size, self.pool_stride)
x = x.view(-1, self.fcdim0)
for i in range(len(self.fcs)-1):
x = self.fcs[i](x)
x = F.relu(x)
x = self.fcs[-1](x)
return F.log_softmax(x, dim=1)
def train_model(self, device, train_loader, optimizer, step_func):
self.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = self(data)
loss = F.nll_loss(output, target)
loss.backward()
optimizer.step()
step_func(batch_idx, data, loss)
def test_model(self, device, test_loader, step_func):
self.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = self(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
step_func(test_loss, correct)MNIST를 사용하면 다음과 같이 호출하여 모델을 구성합니다.
model = tmodel.Conv2dClassifier(features=[1, 20, 50], fcdims=[4*4*50, 500, 10], kernel_size=5, stride=1).to(device)
...
def train_step(batch_idx, data, loss):
...
def test_step(test_loss, correct):
...
...
for epoch in range(1, opt.epochs + 1):
model.train_model(device, train_loader, optimizer, train_step)
model.test_model(device, test_loader, test_step)