mxnet的基本用法

本文章全为一个note, 代码来自zh.d2.ai. 目的在于自己看到这代码能想起来mxnet的基本用法.

import d2lzh as dd2l
from mxnet import autograd, gluon, init, nd
from mxnet.gluon import data as gdata, loss as gloss, nn
import numpy as np
import pandas as pd


train_data = pd.read_csv('data/kaggle_house_pred_train.csv')
test_data = pd.read_csv('data/kaggle_house_pred_test.csv')

# 获得数字特征
all_features = pd.concat((train_data.iloc[:, 1:-1], test_data.iloc[:,1:]))
numeric_features = all_features.dtypes[all_features.dtypes != 'object'].index


# 数字特征进行规则化
all_features[numeric_features] = all_features[numeric_features].apply(lambda x : (x - x.mean()) / x.std())
all_features[numeric_features] = all_features[numeric_features].fillna(0)
# 非数字特征进行数字化 get_dummies() 
all_features = pd.get_dummies(all_features, dummy_na=True)



m_train = train_data.shape[0]
train_features = nd.array(all_features.iloc[0:m_train, :])
test_features = nd.array(all_features.iloc[m_train:, :])
train_labels = nd.array(train_data.iloc[:, -1]).reshape((-1, 1))


loss = gloss.L2Loss()

def get_net():
    net = nn.Sequential()
    net.add(nn.Dense(1))
    net.initialize()
    return net

def log_error(net, features, labels):
    # 将小于1的log值设成1
    preds = nd.clip(net(features), 1, float('inf'))
    error = nd.sqrt(2 * loss(preds.log(), labels.log()).mean())
    return error.asscalar()


def train(net, train_features, train_labels, num_epochs, learning_rate, weight_decay, batch_size):
    train_loss = []
    train_iter = gdata.DataLoader(gdata.ArrayDataset(train_features, train_labels), batch_size=batch_size, shuffle=True)
    #ADM优化算法 - 梯度下降改进
    trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate' : learning_rate, 'wd' : weight_decay})
    for epoch in range(num_epochs):
        for data, label in train_iter:
            # 自动求梯度
            with autograd.record():
                l = loss(net(data), label)
            l.backward()
            # 更新网络权重
            trainer.update(batch_size) #or trainer.step(batch_size)
        train_loss.append(log_error(net, train_features, train_labels))
    return train_loss



def train_and_pred(train_features, train_labels, test_features, test_labels, num_epochs, learning_rate, weight_decay, batch_size):
    net = get_net()
    train_loss = train(net, train_features, train_labels, num_epochs, learning_rate, weight_decay, batch_size)
    # d2lzh包画图
    dd2l.semilogy(range(1, num_epochs + 1), train_loss, 'num_epoch', 'loss')
    # 测试数据预测
    preds = net(test_features).asnumpy()
    test_data['SalePrice'] = preds.reshape((-1, 1))
    submission = pd.concat([test_data['Id'], test_data['SalePrice']], axis=1)
    # 写入预测结果
    submission.to_csv('submission.csv', index=False)



# 超参数设置
num_epochs, learning_rate, weight_decay, batch_size = 100, 10, 0, 64


train_and_pred(train_features, train_labels, test_features, None, num_epochs, learning_rate, weight_decay, batch_size)