深度学习|2|softmax回归pytorch实现iris鸢尾花数据的softmax回归分类任务

iris数据集介绍

由统计学家和植物学家Ronald Fisher在1936年收集并发布。该数据集中包含了150个样本，其中每个样本代表了一朵鸢尾花（iris flower），并且包含了四个特征（sepal length（花萼长度）、sepal width（花萼宽度）、petal length（花瓣长度）和petal width（花瓣宽度））以及对应的类别标签（iris setosa、iris versicolor和iris virginica）。

• 样本数量：150条
• 类别数量：3类
• 每类样本：50条
• 特征维度：4

读取数据集

import numpy as np
import torch

def load_iris(filename):
    data = np.load(filename)
    features = data['data']
    labels = data['label']
    return torch.tensor(features, dtype=torch.float64), torch.tensor(labels, dtype=torch.int64)

train_data, train_label = load_iris(r"../../Dataset/iris/iris_train.npz")
valid_data, valid_label = load_iris(r"../../Dataset/iris/iris_valid.npz")
print(train_data.shape, train_label.shape, valid_data.shape, valid_label.shape,)
input_dim = train_data.shape[1]
output_dim = int(train_label.max().numpy()) + 1
print(input_dim, output_dim)

import random

def data_iter(feature, label, _batch_size):
    num_samples = len(label)
    index_list = list(range(num_samples))
    random.shuffle(index_list)
    
    for i in range(0, num_samples, _batch_size):
        batch_index = index_list[i: min(i + _batch_size, num_samples)]
        batch_features = torch.index_select(feature, dim=0, index=torch.LongTensor(batch_index))
        batch_labels = torch.index_select(label, dim=0, index=torch.LongTensor(batch_index))
        yield batch_features, batch_labels
        
for x, y in data_iter(train_data, train_label, 2):
    print(x, y)
    break

在Iris数据集中，标签的取值为0、1、2，因此最大值为2。为了将标签用于多分类问题，需要将其转换为one-hot编码，输出维度为3。

网络模型

采用最基础的softmax网络

def net(_input, _w, _b):
    output = torch.mm(_input, _w) + _b
    exp = torch.exp(output)
    exp_sum = torch.sum(exp, dim=1, keepdim=True)
    output = exp / exp_sum
    return output

w = torch.normal(0, .01, [input_dim, output_dim], requires_grad=True, dtype=torch.float64)
b = torch.normal(0, .01, [1, output_dim], requires_grad=True, dtype=torch.float64)

with torch.no_grad():
    random_input = torch.normal(0, .01, [10, 4], dtype=torch.float64)
    output = net(random_input, w, b)
    print(output.shape)

定义损失函数

def cross_entropy(y_pred, y):
    pred_value = torch.gather(y_pred, 1, y.view(-1, 1))
    _loss = -torch.log(pred_value)
    return _loss.sum()

一个交叉熵损失函数，用于度量模型输出与真实标签之间的差异。

定义优化器

def optimizer(params, _lr, _batch_size):
    with torch.no_grad():
        for param in params:
            param -= _lr * param.grad / _batch_size
            param.grad.zero_()

定义一个梯度下降优化器，用于更新模型的参数。

训练评估

epoch_num = 5
lr = 0.05
batch_size = 5

for epoch in range(epoch_num):
    train(_net=net,
          _params=[w, b],
          _loss=cross_entropy,
          _opt=optimizer,
          _lr=lr,
          _batch_size=batch_size,
          _epoch=epoch,
          _data_iter=data_iter(train_data, train_label, batch_size))

    acc = evaluation(_net=net, 
                     _params=[w, b], 
                     _data_iter=data_iter(valid_data, valid_label, batch_size))
    
    print('epoch %d,valid acc %.3f' % (epoch + 1, acc))
    

结果图：

鸢尾花数据集下载免费