这个模型在极少的轮次内就学会了极快地区分1和0，这让我觉得可能有什么问题。

下面的代码下载MNIST数据集，提取只包含1或0的MNIST图像。从这个MNIST图像子集中随机选择了200个样本。这个随机样本是模型训练的数据集。仅用2个轮次，模型就达到了90%以上的测试集准确率，这是预期的行为吗？我原本预期需要更多的轮次才能训练模型达到这种测试集准确率水平。

模型代码：

%reset -fimport torchimport torch.nn as nnimport torchvisionimport torchvision.transforms as transformsimport torchimport torch.nn as nnimport torchvisionimport torchvision.transforms as transformsimport torch.utils.data as data_utilsimport numpy as npimport matplotlib.pyplot as pltfrom sklearn.datasets import make_moonsfrom matplotlib import pyplotfrom pandas import DataFrameimport torchvision.datasets as dsetimport osimport torch.nn.functional as Fimport timeimport randomimport picklefrom sklearn.metrics import confusion_matriximport pandas as pdimport sklearntrans = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])root = './data'if not os.path.exists(root):    os.mkdir(root)train_set = dset.MNIST(root=root, train=True, transform=trans, download=True)test_set = dset.MNIST(root=root, train=False, transform=trans, download=True)batch_size = 64train_loader = torch.utils.data.DataLoader(                 dataset=train_set,                 batch_size=batch_size,                 shuffle=True)test_loader = torch.utils.data.DataLoader(                dataset=test_set,                batch_size=batch_size,shuffle=True)class NeuralNet(nn.Module):    def __init__(self):        super(NeuralNet, self).__init__()        self.fc1 = nn.Linear(28*28, 500)        self.fc2 = nn.Linear(500, 256)        self.fc3 = nn.Linear(256, 2)    def forward(self, x):        x = x.view(-1, 28*28)        x = F.relu(self.fc1(x))        x = F.relu(self.fc2(x))        x = self.fc3(x)        return xnum_epochs = 2random_sample_size = 200values_0_or_1 = [t for t in train_set if (int(t[1]) == 0 or int(t[1]) == 1)]values_0_or_1_testset = [t for t in test_set if (int(t[1]) == 0 or int(t[1]) == 1)]print(len(values_0_or_1))print(len(values_0_or_1_testset))train_loader_subset = torch.utils.data.DataLoader(                 dataset=values_0_or_1,                 batch_size=batch_size,                 shuffle=True)test_loader_subset = torch.utils.data.DataLoader(                 dataset=values_0_or_1_testset,                 batch_size=batch_size,                 shuffle=False)train_loader = train_loader_subset# Hyper-parameters input_size = 100hidden_size = 100num_classes = 2# learning_rate = 0.00001learning_rate = .0001# Device configurationdevice = 'cpu'print_progress_every_n_epochs = 1model = NeuralNet().to(device)# Loss and optimizercriterion = nn.CrossEntropyLoss()optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)  N = len(train_loader)# Train the modeltotal_step = len(train_loader)most_recent_prediction = []test_actual_predicted_dict = {}rm = random.sample(list(values_0_or_1), random_sample_size)train_loader_subset = data_utils.DataLoader(rm, batch_size=4)for epoch in range(num_epochs):    for i, (images, labels) in enumerate(train_loader_subset):          # Move tensors to the configured device        images = images.reshape(-1, 2).to(device)        labels = labels.to(device)        # Forward pass        outputs = model(images)        loss = criterion(outputs, labels)        # Backward and optimize        optimizer.zero_grad()        loss.backward()        optimizer.step()    if (epoch) % print_progress_every_n_epochs == 0:        print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'.format(epoch+1, num_epochs, i+1, total_step, loss.item()))predicted_test = []model.eval()  # eval mode (batchnorm uses moving mean/variance instead of mini-batch mean/variance)probs_l = []predicted_values = []actual_values = []labels_l = []with torch.no_grad():    for images, labels in test_loader_subset:        images = images.to(device)        labels = labels.to(device)        outputs = model(images)        _, predicted = torch.max(outputs.data, 1)        predicted_test.append(predicted.cpu().numpy())        sm = torch.nn.Softmax()        probabilities = sm(outputs)         probs_l.append(probabilities)          labels_l.append(labels.cpu().numpy())    predicted_values.append(np.concatenate(predicted_test).ravel())    actual_values.append(np.concatenate(labels_l).ravel())if (epoch) % 1 == 0:    print('test accuracy : ', 100 * len((np.where(np.array(predicted_values[0])==(np.array(actual_values[0])))[0])) / len(actual_values[0]))

模型输出（12665和2115分别代表训练集和测试集的大小）：

126652115Epoch [1/2], Step [50/198], Loss: 0.1256Epoch [2/2], Step [50/198], Loss: 0.0151test accuracy :  99.76359338061465/anaconda3/envs/pytorch/lib/python3.7/site-packages/ipykernel_launcher.py:143: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.

回答：

这是我对你的二元实验的一点看法。

看起来你大大降低了数据集的复杂性，而中间层的神经元数量较多，因此模型预计会很快收敛。

请注意，MNIST数据集的通道数为1，这使得任务变得非常简单。

你可以尝试使用CIFAR10，看看是否仍然能在仅2个轮次内获得高准确率。