多类逻辑回归--gluon

下面使用多类逻辑回归进行多类分类。这个模型跟线性回归的主要区别在于输出节点从一个变成了多个。

演示这个模型的常见数据集是手写数字识别MNIST，如下图所示。

这里用稍复杂点的数据集FashionMNIST（https://github.com/zalandoresearch/fashion-mnist，http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/)，它跟MNIST非常像，但是内容不再是分类数字，而是服饰。我们通过gluon的data.vision模块自动下载这个数据。

获取数据：

from mxnet import gluon
from mxnet import ndarray as nd

def transform(data, label):
    return data.astype('float32')/255, label.astype('float32')
mnist_train = gluon.data.vision.FashionMNIST(train=True, transform=transform)
mnist_test = gluon.data.vision.FashionMNIST(train=False, transform=transform)

D:\Users\Administrator\Anaconda3\lib\site-packages\h5py\__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
D:\Users\Administrator\Anaconda3\lib\site-packages\mxnet\gluon\data\vision\datasets.py:84: DeprecationWarning: The binary mode of fromstring is deprecated, as it behaves surprisingly on unicode inputs. Use frombuffer instead
  label = np.fromstring(fin.read(), dtype=np.uint8).astype(np.int32)
D:\Users\Administrator\Anaconda3\lib\site-packages\mxnet\gluon\data\vision\datasets.py:88: DeprecationWarning: The binary mode of fromstring is deprecated, as it behaves surprisingly on unicode inputs. Use frombuffer instead
  data = np.fromstring(fin.read(), dtype=np.uint8)

data, label = mnist_train[0]
('example shape: ', data.shape, 'label:', label)

('example shape: ', (28, 28, 1), 'label:', 2.0)

画出前十个样本：

import matplotlib.pyplot as plt

def show_images(images):
    n = images.shape[0]
    _, figs = plt.subplots(1, n, figsize=(15, 15))
    for i in range(n):
        figs[i].imshow(images[i].reshape((28, 28)).asnumpy())
        figs[i].axes.get_xaxis().set_visible(False)
        figs[i].axes.get_yaxis().set_visible(False)
    plt.show()

def get_text_labels(label):
    text_labels = [
        't-shirt', 'trouser', 'pullover', 'dress,', 'coat',
        'sandal', 'shirt', 'sneaker', 'bag', 'ankle boot'
    ]
    return [text_labels[int(i)] for i in label]

data, label = mnist_train[0:9]
show_images(data)
print(get_text_labels(label))

['pullover', 'ankle boot', 'shirt', 't-shirt', 'dress,', 'coat', 'coat', 'sandal', 'coat']

读取数据并初始化模型参数：

batch_size = 256
train_data = gluon.data.DataLoader(mnist_train, batch_size, shuffle=True)
test_data = gluon.data.DataLoader(mnist_test, batch_size, shuffle=False)
num_inputs = 784
num_outputs = 10

W = nd.random_normal(shape=(num_inputs, num_outputs))
b = nd.random_normal(shape=num_outputs)

params = [W, b]
for param in params:
    param.attach_grad()

定义模型和交叉熵损失函数、计算精度：

from mxnet import nd
def softmax(X):
    exp = nd.exp(X)
    # 假设exp是矩阵，这里对行进行求和，并要求保留axis 1，
    # 就是返回 (nrows, 1) 形状的矩阵
    partition = exp.sum(axis=1, keepdims=True)
    return exp / partition

X = nd.random_normal(shape=(2,5))
X_prob = softmax(X)
print(X_prob)
print(X_prob.sum(axis=1))

def net(X):
    return softmax(nd.dot(X.reshape((-1,num_inputs)), W) + b)
def cross_entropy(yhat, y):
    return - nd.pick(nd.log(yhat), y)

def accuracy(output, label):
    return nd.mean(output.argmax(axis=1)==label).asscalar()
def evaluate_accuracy(data_iterator, net):
    acc = 0.
    for data, label in data_iterator:
        output = net(data)
        acc += accuracy(output, label)
    return acc / len(data_iterator)
evaluate_accuracy(test_data, net)

[[0.01606132 0.4672664  0.16146936 0.33617657 0.01902632]
 [0.27215305 0.16692853 0.04439886 0.39427412 0.12224543]]
<NDArray 2x5 @cpu(0)>

[1. 1.]
<NDArray 2 @cpu(0)>

0.08974609375 

训练：

import sys
sys.path.append('..')
from utils import SGD
from mxnet import autograd

learning_rate = .1

for epoch in range(5):
    train_loss = 0.
    train_acc = 0.
    for data, label in train_data:
        with autograd.record():
            output = net(data)
            loss = cross_entropy(output, label)
        loss.backward()
        # 将梯度做平均，这样学习率会对batch size不那么敏感
        SGD(params, learning_rate/batch_size)

        train_loss += nd.mean(loss).asscalar()
        train_acc += accuracy(output, label)

    test_acc = evaluate_accuracy(test_data, net)
    print("Epoch %d. Loss: %f, Train acc %f, Test acc %f" % (
        epoch, train_loss/len(train_data), train_acc/len(train_data), test_acc))

Epoch 0. Loss: 3.890012, Train acc 0.441949, Test acc 0.592187
Epoch 1. Loss: 1.900754, Train acc 0.628596, Test acc 0.659961
Epoch 2. Loss: 1.559882, Train acc 0.678042, Test acc 0.696680
Epoch 3. Loss: 1.382150, Train acc 0.704854, Test acc 0.711816
Epoch 4. Loss: 1.270546, Train acc 0.721853, Test acc 0.732324 

预测：

data, label = mnist_test[0:9]
show_images(data)
print('true labels')
print(get_text_labels(label))

predicted_labels = net(data).argmax(axis=1)
print('predicted labels')
print(get_text_labels(predicted_labels.asnumpy()))

true labels
['t-shirt', 'trouser', 'pullover', 'pullover', 'dress,', 'pullover', 'bag', 'shirt', 'sandal']
predicted labels
['t-shirt', 'trouser', 'pullover', 'shirt', 't-shirt', 'shirt', 'bag', 'shirt', 'sandal']

****************************************************************************************************************************

****************************************************************************************************************************

使用gluon的多类逻辑回归

获取和读取数据，定义和初始化模型：

import sys
sys.path.append('..')
import utils

batch_size = 256
train_data, test_data = utils.load_data_fashion_mnist(batch_size)

from mxnet import gluon

net = gluon.nn.Sequential()
with net.name_scope():
    net.add(gluon.nn.Flatten())
    net.add(gluon.nn.Dense(10))
net.initialize()

D:\Users\Administrator\Anaconda3\lib\site-packages\mxnet\gluon\data\vision\datasets.py:84: DeprecationWarning: The binary mode of fromstring is deprecated, as it behaves surprisingly on unicode inputs. Use frombuffer instead
  label = np.fromstring(fin.read(), dtype=np.uint8).astype(np.int32)
D:\Users\Administrator\Anaconda3\lib\site-packages\mxnet\gluon\data\vision\datasets.py:88: DeprecationWarning: The binary mode of fromstring is deprecated, as it behaves surprisingly on unicode inputs. Use frombuffer instead
  data = np.fromstring(fin.read(), dtype=np.uint8)

Softmax和交叉熵损失函数，优化：

softmax_cross_entropy = gluon.loss.SoftmaxCrossEntropyLoss()

trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.1})

训练：

from mxnet import ndarray as nd
from mxnet import autograd

for epoch in range(5):
    train_loss = 0.
    train_acc = 0.
    for data, label in train_data:
        with autograd.record():
            output = net(data)
            loss = softmax_cross_entropy(output, label)
        loss.backward()
        trainer.step(batch_size)

        train_loss += nd.mean(loss).asscalar()
        train_acc += utils.accuracy(output, label)

    test_acc = utils.evaluate_accuracy(test_data, net)
    print("Epoch %d. Loss: %f, Train acc %f, Test acc %f" % (
        epoch, train_loss/len(train_data), train_acc/len(train_data), test_acc))

Epoch 0. Loss: 0.800685, Train acc 0.738849, Test acc 0.803085
Epoch 1. Loss: 0.577781, Train acc 0.809412, Test acc 0.819912
Epoch 2. Loss: 0.532679, Train acc 0.821548, Test acc 0.830629
Epoch 3. Loss: 0.508033, Train acc 0.828676, Test acc 0.829828
Epoch 4. Loss: 0.491286, Train acc 0.833934, Test acc 0.840745

data, label = mnist_test[0:9]
show_images(data)
print('true labels')
print(get_text_labels(label))

predicted_labels = net(data).argmax(axis=1)
print('predicted labels')
print(get_text_labels(predicted_labels.asnumpy()))

true labels
['t-shirt', 'trouser', 'pullover', 'pullover', 'dress,', 'pullover', 'bag', 'shirt', 'sandal']
predicted labels
['t-shirt', 'trouser', 'pullover', 't-shirt', 'coat', 'shirt', 'bag', 'shirt', 'sandal']