Question

我正在尝试使用mxnet用3D空间数据训练3D convnet。当我运行程序时，损耗通常会随着时间的流逝而减少，但是训练精度和测试精度却保持不变。我对神经网络非常陌生，我不知道为什么会这样。我不确定这是因为我的网络参数错误，还是我的数据未正确预处理，或者用于训练网络的代码是否有问题，或者完全是其他原因。

这是我当前用于convnet的代码及其产生的输出：

train_data = mx.gluon.data.DataLoader(train_dataset, batch_size= 64,shuffle= True, num_workers = cpucount)
test_data = mx.gluon.data.DataLoader(test_dataset,batch_size= 64,shuffle= True, num_workers = cpucount)

batch_size = 64
num_inputs = 2541
num_outputs = 2
num_fc = 635
net = gluon.nn.Sequential()

with net.name_scope():
    net.add(gluon.nn.Conv3D(channels=1, kernel_size=3, activation='relu'))
    net.add(gluon.nn.MaxPool3D(pool_size=2, strides=2))
    net.add(gluon.nn.Conv3D(channels=1, kernel_size=3, activation='relu'))
    net.add(gluon.nn.MaxPool3D(pool_size=2, strides=2))

    net.add(gluon.nn.Flatten())

    net.add(gluon.nn.Dense(num_fc, activation="relu"))
    net.add(gluon.nn.Dense(num_outputs))

net.collect_params().initialize(mx.init.Xavier(magnitude=2.24), ctx=ctx)
softmax_cross_entropy = gluon.loss.SoftmaxCrossEntropyLoss()
trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': .0001})

def evaluate_accuracy(data_iterator, net):
    acc = mx.metric.Accuracy()
    for i, (data, label) in enumerate(data_iterator):
        data = data.as_in_context(ctx)
        label = label.as_in_context(ctx)
        output = net(data)
        predictions = nd.argmax(output,axis = 1)
        label = label.reshape(len(label))
        acc.update(preds=predictions, labels=label)
    return acc.get()[1]

epochs = 100
smoothing_constant = .01

for e in range(epochs):
    for i, (data, label) in enumerate(train_data):
        data = data.as_in_context(ctx)
        label = label.as_in_context(ctx)
        with autograd.record():
            output = net(data)
            loss = softmax_cross_entropy(output, label)
        loss.backward()
        trainer.step(data.shape[0])

        curr_loss = nd.mean(loss).asscalar()
        moving_loss = (curr_loss if ((i == 0) and (e == 0))
                       else (1 - smoothing_constant) * moving_loss + smoothing_constant * curr_loss)

    test_accuracy = evaluate_accuracy(test_data, net)
    train_accuracy = evaluate_accuracy(train_data, net)

输出：时代0.损失：26525280.32107588，Train_acc 0.462039045553，Test_acc 0.386554621849 时期1.损失：17045452.882872812，Train_acc 0.462039045553，Test_acc 0.386554621849 时期2.损失：10953605.785322478，Train_acc 0.462039045553，Test_acc 0.386554621849 时期3.损失：7038914.162310514，Train_acc 0.462039045553，Test_acc 0.386554621849 纪元4.损失：4523287.90677917，Train_acc 0.462039045553，Test_acc 0.386554621849 时代5：损失：2906717.2884657932，Train_acc 0.462039045553，Test_acc 0.386554621849 时期6.损失：1867890.253548351，Train_acc 0.462039045553，Test_acc 0.386554621849

（我省略了其余的时期，但即使损失在0.09左右，精度仍然相同）

损失减少，但3D convnet的精度保持不变

0 个答案: