我已经完成了一个针对MNIST数据集的PyTorch MLP模型，但在训练时得到了两种不同的结果：使用PyTorch的MNIST数据集时准确率为0.90+，而使用Keras的MNIST数据集时准确率约为0.10。以下是我的代码，依赖如下：PyTorch 0.3.0.post4, keras 2.1.3, tensorflow后端1.4.1 gpu版本。
from __future__ import absolute_importfrom __future__ import divisionfrom __future__ import print_functionimport numpy as npimport torch as ptimport torchvision as ptvfrom keras.datasets import mnistfrom torch.nn import functional as Ffrom torch.utils.data import Dataset, DataLoader# training data from PyTorchtrain_set = ptv.datasets.MNIST("./data/mnist/train", train=True, transform=ptv.transforms.ToTensor(), download=True)test_set = ptv.datasets.MNIST("./data/mnist/test", train=False, transform=ptv.transforms.ToTensor(), download=True)train_dataset = DataLoader(train_set, batch_size=100, shuffle=True)test_dataset = DataLoader(test_set, batch_size=10000, shuffle=True)class MLP(pt.nn.Module): """The Multi-layer perceptron""" def __init__(self): super(MLP, self).__init__() self.fc1 = pt.nn.Linear(784, 512) self.fc2 = pt.nn.Linear(512, 128) self.fc3 = pt.nn.Linear(128, 10) self.use_gpu = True def forward(self, din): din = din.view(-1, 28 * 28) dout = F.relu(self.fc1(din)) dout = F.relu(self.fc2(dout)) # return F.softmax(self.fc3(dout)) return self.fc3(dout)model = MLP().cuda()print(model)# loss func and optimoptimizer = pt.optim.SGD(model.parameters(), lr=1)criterion = pt.nn.CrossEntropyLoss().cuda()def evaluate_acc(pred, label): pred = pred.cpu().data.numpy() label = label.cpu().data.numpy() test_np = (np.argmax(pred, 1) == label) test_np = np.float32(test_np) return np.mean(test_np)def evaluate_loader(loader): print("evaluating ...") accurarcy_list = [] for i, (inputs, labels) in enumerate(loader): inputs = pt.autograd.Variable(inputs).cuda() labels = pt.autograd.Variable(labels).cuda() outputs = model(inputs) accurarcy_list.append(evaluate_acc(outputs, labels)) print(sum(accurarcy_list) / len(accurarcy_list))def training(d, epochs): for x in range(epochs): for i, data in enumerate(d): optimizer.zero_grad() (inputs, labels) = data inputs = pt.autograd.Variable(inputs).cuda() labels = pt.autograd.Variable(labels).cuda() outputs = model(inputs) loss = criterion(outputs, labels) loss.backward() optimizer.step() if i % 200 == 0: print(i, ":", evaluate_acc(outputs, labels))# Training MLP for 4 epochs with MNIST dataset from PyTorchtraining(train_dataset, 4)# The accuracy is ~0.96.evaluate_loader(test_dataset)print("###########################################################")def load_mnist(): (x, y), (x_test, y_test) = mnist.load_data() x = x.reshape((-1, 1, 28, 28)).astype(np.float32) x_test = x_test.reshape((-1, 1, 28, 28)).astype(np.float32) y = y.astype(np.int64) y_test = y_test.astype(np.int64) print("x.shape", x.shape, "y.shape", y.shape, "\nx_test.shape", x_test.shape, "y_test.shape", y_test.shape, ) return x, y, x_test, y_testclass TMPDataset(Dataset): """Dateset for loading Keras MNIST dataset.""" def __init__(self, a, b): self.x = a self.y = b def __getitem__(self, item): return self.x[item], self.y[item] def __len__(self): return len(self.y)x_train, y_train, x_test, y_test = load_mnist()# Create dataloader for MNIST dataset from Keras.test_loader = DataLoader(TMPDataset(x_test, y_test), num_workers=1, batch_size=10000)train_loader = DataLoader(TMPDataset(x_train, y_train), shuffle=True, batch_size=100)# Evaluate the performance of MLP trained on PyTorch dataset and the accurach is ~0.96.evaluate_loader(test_loader)evaluate_loader(train_loader)model = MLP().cuda()print(model)optimizer = pt.optim.SGD(model.parameters(), lr=1)criterion = pt.nn.CrossEntropyLoss().cuda()# Train now on MNIST dataset from Keras.training(train_loader, 4)# Evaluate the trianed model on MNIST dataset from Keras and result in performance ~0.10...evaluate_loader(test_loader)evaluate_loader(train_loader)我已经检查了来自Keras的MNIST数据集的一些样本，并没有发现错误。我想知道数据集有什么问题？代码可以无错误运行，运行它来查看结果。
回答：
来自Keras的MNIST数据未被归一化；按照Keras MNIST MLP示例，你应该手动进行归一化处理，即在你的load_data()函数中应包括以下内容：
x /= 255x_test /= 255关于PyTorch我不确定，但看起来他们自己的工具函数提供的MNIST数据已经是归一化过的（就像Tensorflow的情况一样 – 参见我在这里的回答中的第三点）。
在未归一化的输入数据情况下，10%的准确率（即相当于随机猜测）是完全合理的。
