嗨，我刚刚通过udacity课程学会了在pytorch中实现NN模型，从而创建了一个带有几个CNN和FC层的简单模型。经过多次努力，我使模型发挥了作用。但即使在反复执行死刑之后，它似乎也陷入了同样的损失。我不知道哪里出了问题。一定是逻辑错误，我看不出来。这是代码。
型号
class cifar_clasify(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(3,16,3)
self.BNorm1 = nn.BatchNorm2d(16)
self.conv2 = nn.Conv2d(16,32,3)
self.BNorm2 = nn.BatchNorm2d(32)
self.fc1 = nn.Linear(32*6*6,256)
self.fc2 = nn.Linear(256,512)
self.fc3 = nn.Linear(512,10)
self.drop = nn.Dropout(p =0.2)
def forward(self,x):
out = self.conv1(x)
out = F.relu(out)
#print(out.shape)
out = F.max_pool2d(out,2)
out = self.BNorm1(out)
#print(out.shape)
out = self.conv2(out)
out = F.relu(out)
#print(out.shape)
out = F.max_pool2d(out,2)
out = self.BNorm2(out)
#print(out.shape)
out = out.view(out.shape[0],-1)
out = self.fc1(out)
out = self.drop(F.relu(out))
out = self.fc2(out)
out = self.drop(F.relu(out))
final = F.log_softmax(F.relu(self.fc3(out)) , dim = 1)
return final
训练代码
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
model = cifar_clasify()
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.parameters(), lr =0.03)
epoch =2
step = 2
running_loss = 0
accuracy = 0
print_every = 5
model.to(device)
for e in range(epoch):
for inputs,label_ in zip(train_X,train_labels):
step +=1
inputs = inputs.view((-1,3,32,32))
inputs,label_ = inputs.to(device) , label_.to(device)
#inputs.cuda()
#label.cuda()
optimizer.zero_grad()
logps = model.forward(inputs)
loss = criterion(logps , label_.reshape(1))
loss.backward()
optimizer.step()
running_loss += loss.item()
if step % print_every == 0:
test_loss = 0
accuracy = 0
model.eval()
with torch.no_grad():
for testx, lab in zip(test_X , test_labels):
testx = testx.view((-1,3,32,32))
testx,lab = testx.to(device) , lab.to(device)
lab = lab.reshape(1)
logps = model.forward(testx)
batch_loss = criterion(logps , lab)
#print(batch_loss.item())
test_loss += batch_loss.item()
ps = torch.exp(logps)
top_p , topclass = ps.topk(1,dim = 1)
equals = topclass == lab.view(*topclass.shape)
accuracy += torch.mean(torch.mean(equals.type(torch.FloatTensor))).item()
print(f"Epoch {e+1}/{epoch}.. "
f"Train loss: {running_loss/print_every:.3f}.. "
f"Test loss: {test_loss/len(test_X):.3f}.. "
f"Test accuracy: {accuracy/len(test_X):.3f}")
running_loss = 0
model.train()
以下是我不得不停止的结果，因为它没有改善：
Epoch 1/2.. Train loss: 1.396.. Test loss: 5.288.. Test accuracy: 0.104
step = 5
Epoch 1/2.. Train loss: 3.038.. Test loss: 2.303.. Test accuracy: 0.104
step = 10
Epoch 1/2.. Train loss: 2.303.. Test loss: 2.303.. Test accuracy: 0.104
step = 15
Epoch 1/2.. Train loss: 2.669.. Test loss: 2.318.. Test accuracy: 0.105
step = 20
Epoch 1/2.. Train loss: 3.652.. Test loss: 2.303.. Test accuracy: 0.104
step = 25
Epoch 1/2.. Train loss: 2.303.. Test loss: 2.303.. Test accuracy: 0.104
step = 30
Epoch 1/2.. Train loss: 2.303.. Test loss: 2.303.. Test accuracy: 0.104
step = 35
Epoch 1/2.. Train loss: 2.303.. Test loss: 2.303.. Test accuracy: 0.104
step = 40
Epoch 1/2.. Train loss: 2.303.. Test loss: 2.303.. Test accuracy: 0.104
step = 45
Epoch 1/2.. Train loss: 2.303.. Test loss: 2.303.. Test accuracy: 0.104
step = 50
Epoch 1/2.. Train loss: 2.303.. Test loss: 2.303.. Test accuracy: 0.104
step = 55
如果你想要任何其他信息，这是代码：
谷歌colab 中CIFAR 10分类的简单CNN
由于批量大小为1，请使用较低的学习率，如1e-4或增加批量大小。
不过，我建议批量生产16号或更大的。
编辑：要创建一批数据，可以执行以下操作。
N = input.shape[0] #know the total size/samples in input
for i in range(n_epochs):
# this is to shuffle data
indices = torch.randperm(N)
for idx in range(0, N, batch_size):
batch_input = input[idx:idx+batch_size] # this will get you input of size batch_size
# do whatever you want with the batch_input
# ....