创建LeNet 10神经网络

Question

我正在尝试使用Gluon API在MXNet中编写一个训练循环，并编写一个验证循环来评估一个训练过的网络。

我目前正在尝试定义一个网络。我需要定义一个基于LeNet架构的网络，而不是原始MNIST数据的10个输出层，我应该有2个输出层。

我真的不知道该怎么做。我尝试在创建网络时包含Flatten和Dense，但是我得到了下面的错误。是否与Conv2D和MaxPool2D层中的某些设置有关？

编辑:在更多地摆弄它之后，我现在认为它与损失函数有关……

def get_network():
    """
    Should create the LeNet 10 network but with 2 output units instead of 10 and return a classification loss function
    
    :return: the network and the loss function
    :rtype: (gluon.Block, gluon.Block)
    """
    
    net = None
    loss_fn = None
    
    # create a network
    net = nn.Sequential()
    
    net.add(
        nn.Conv2D(channels=6, kernel_size=5, activation='relu'),
        nn.MaxPool2D(pool_size=2, strides=2),
        nn.Conv2D(channels=16, kernel_size=3, activation='relu'),
        nn.MaxPool2D(pool_size=2, strides=2)  
    )
    
    net.initialize(init=init.Xavier())
    
    # choose and set loss_fn for a classification task
    loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
    
    return net, loss_fn

n, loss_fn = get_network()

assert isinstance(n[0], nn.Conv2D)
assert isinstance(n[2], nn.Conv2D)
assert isinstance(n[1], nn.MaxPool2D)
assert isinstance(n[3], nn.MaxPool2D)

for l in n[-3:]:
    assert isinstance(l, nn.Dense)

我得到以下错误：

AssertionError                            Traceback (most recent call last)
<ipython-input-29-e1d75f4e8660> in <module>
      7 
      8 for l in n[-3:]:
----> 9     assert isinstance(l, nn.Dense)

AssertionError:

Answer 1

你应该为模型添加3个密集层，下面的代码运行时不会有任何错误：

from mxnet import init, gluon
from mxnet.gluon import nn

def get_network():
    """
    Should create the LeNet 10 network but with 2 output units instead of 10 and return a classification loss function
    
    :return: the network and the loss function
    :rtype: (gluon.Block, gluon.Block)
    """
    
    net = None
    loss_fn = None
    
    # create a network
    net = nn.Sequential()

    net.add(
      nn.Conv2D(channels=6,kernel_size=5,activation='relu'),
      nn.MaxPool2D(pool_size=2,strides=2),
      nn.Conv2D(channels=16,kernel_size=3,activation='relu'),
      nn.MaxPool2D(pool_size=2,strides=2),
      nn.Dense(120, activation='relu'),
      nn.Dense(84, activation='relu'),
      nn.Dense(2))
    
    net.initialize(init=init.Xavier())
    
    # choose and set loss_fn for a classification task
    loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
    
    return net, loss_fn

n, loss_fn = get_network()

assert isinstance(n[0], nn.Conv2D)
assert isinstance(n[2], nn.Conv2D)
assert isinstance(n[1], nn.MaxPool2D)
assert isinstance(n[3], nn.MaxPool2D)

for l in n[-3:]:
    assert isinstance(l, nn.Dense)

此外，here是使用FashionMNIST训练LeNet的示例代码