讲在前面，本来想通过一个简单的多层感知机实验一下不同的优化方法的，结果写着写着就先研究起评价指标来了，之前也写过一篇：www.cnblogs.com/xiximayou/p… 与上篇不同的是，这次我们新加了一些相关的实现，接下来我们慢慢来看。

利用pytorch搭建多层感知机分类的整个流程

导入相关包

from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
import numpy as np
import torch
import torch.nn as nn
import matplotlib.pyplot as plt
from torch.utils.data import Dataset, DataLoader
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, classification_report, confusion_matrix
from sklearn.preprocessing import label_binarize
from sklearn.metrics import roc_curve, auc

设置随机种子

设置随机种子总是需要的，它可以让我们的实验可以复现：即对于随机初始化的数据生成相同的结果。

np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)

加载数据

使用简单的sklearn自带的数字数据：

print("加载数据")
digits = load_digits()
data, label = digits.data, digits.target
# print(data.shape, label.shape)
train_data, test_data, train_label, test_label = train_test_split(data, label, test_size=.3, random_state=123)
print('训练数据：', train_data.shape)
print('测试数据：', test_data.shape)

定义相关参数

print("定义相关参数")
epochs = 30
batch_size = train_data.shape[0]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 
input_dim = data.shape[1]
hidden_dim = 256
output_dim = len(set(label))

构建数据集

pytorch构建数据集可以自己实现一个类，继承Dataset，然后在类中重写__len__和__getitem__方法。

print("构建数据集")
class DigitsDataset(Dataset):
  def __init__(self, input_data, input_label):
    data = []
    for i,j in zip(input_data, input_label):
      data.append((i,j))
    self.data = data

  def __len__(self):
    return len(self.data)

  def __getitem__(self, index):
    d, l = self.data[index]
    return d, l

在初始化的时候，我们将每一条数据及其标签放在一个列表中，然后在__len__中计算数据的总量，在__getitem__中根据索引取得每一条数据。接下我们我们要使用DataLoader将定义的数据集转换为数据加载器。

trainDataset = DigitsDataset(train_data, train_label)
testDataset = DigitsDataset(test_data, test_label)
# print(trainDataset[0])
# print(trainDataset[0])
trainDataLoader = DataLoader(trainDataset, batch_size=batch_size, shuffle=True, num_workers=2)
testDataLoader = DataLoader(testDataset, batch_size=batch_size, shuffle=False, num_workers=2)

定义模型

这里我们就简单的实现下多层感知机：

class Model(nn.Module):
  def __init__(self, input_dim, hidden_dim, output_dim):
    super(Model, self).__init__()
    self.fc1 = nn.Linear(input_dim, hidden_dim) 
    self.relu = nn.ReLU()
    self.fc2 = nn.Linear(hidden_dim, output_dim)

  def forward(self, x):
    x = self.fc1(x)
    x = self.relu(x)
    x = self.fc2(x)
    return x

定义损失函数、优化器和初始化相关参数

model = Model(input_dim, hidden_dim, output_dim)
print(model)
model.to(device)

print("定义损失函数、优化器")
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9, weight_decay=1e-4)

print("初始化相关参数")
for param in model.parameters():
  nn.init.normal_(param, mean=0, std=0.01)

进行训练和测试

这里我们就仅仅使用sklearn自带的评价指标函数来计算评价指标：accuracy_score：计算准确率, precision_score：计算精确率, recall_score：计算召回率, f1_score：计算f1, classification_report：分类报告, confusion_matrix：混淆矩阵。具体是怎么使用的，我们可以直接看代码。

print("开始训练主循环")
total_step = len(trainDataLoader)

model.train()
for epoch in range(epochs):
  tot_loss = 0.0
  tot_acc = 0.0
  train_preds = []
  train_trues = []
  # model.train()
  for i,(train_data_batch, train_label_batch) in enumerate(trainDataLoader):
    train_data_batch = train_data_batch.float().to(device) # 将double数据转换为float
    train_label_batch = train_label_batch.to(device)
    outputs = model(train_data_batch)
    # _, preds = torch.max(outputs.data, 1)
    loss = criterion(outputs, train_label_batch)
    # print(loss)
    #反向传播优化网络参数
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()
    #累加每个step的损失
    tot_loss += loss.data
    train_outputs = outputs.argmax(dim=1)

    train_preds.extend(train_outputs.detach().cpu().numpy())
    train_trues.extend(train_label_batch.detach().cpu().numpy())

    # tot_acc += (outputs.argmax(dim=1) == train_label_batch).sum().item()

  sklearn_accuracy = accuracy_score(train_trues, train_preds) 
  sklearn_precision = precision_score(train_trues, train_preds, average='micro')
  sklearn_recall = recall_score(train_trues, train_preds, average='micro')
  sklearn_f1 = f1_score(train_trues, train_preds, average='micro')
  print("[sklearn_metrics] Epoch:{} loss:{:.4f} accuracy:{:.4f} precision:{:.4f} recall:{:.4f} f1:{:.4f}".format(epoch, tot_loss, sklearn_accuracy, sklearn_precision, sklearn_recall, sklearn_f1))

test_preds = []
test_trues = []
model.eval()
with torch.no_grad():
  for i,(test_data_batch, test_data_label) in enumerate(testDataLoader):
    test_data_batch = test_data_batch.float().to(device) # 将double数据转换为float
    test_data_label = test_data_label.to(device)
    test_outputs = model(test_data_batch)
    test_outputs = test_outputs.argmax(dim=1)
    test_preds.extend(test_outputs.detach().cpu().numpy())
    test_trues.extend(test_data_label.detach().cpu().numpy())

  sklearn_precision = precision_score(test_trues, test_preds, average='micro')
  sklearn_recall = recall_score(test_trues, test_preds, average='micro')
  sklearn_f1 = f1_score(test_trues, test_preds, average='micro')
  print(classification_report(test_trues, test_preds))
  conf_matrix = get_confusion_matrix(test_trues, test_preds)
  print(conf_matrix)
  plot_confusion_matrix(conf_matrix)
  print("[sklearn_metrics] accuracy:{:.4f} precision:{:.4f} recall:{:.4f} f1:{:.4f}".format(sklearn_accuracy, sklearn_precision, sklearn_recall, sklearn_f1))

定义和绘制混淆矩阵

额外的，我们补充一下混淆矩阵的计算和绘制。

def get_confusion_matrix(trues, preds):
  labels = [0,1,2,3,4,5,6,7,8,9]
  conf_matrix = confusion_matrix(trues, preds, labels)
  return conf_matrix
  
def plot_confusion_matrix(conf_matrix):
  plt.imshow(conf_matrix, cmap=plt.cm.Greens)
  indices = range(conf_matrix.shape[0])
  labels = [0,1,2,3,4,5,6,7,8,9]
  plt.xticks(indices, labels)
  plt.yticks(indices, labels)
  plt.colorbar()
  plt.xlabel('y_pred')
  plt.ylabel('y_true')
  # 显示数据
  for first_index in range(conf_matrix.shape[0]):
    for second_index in range(conf_matrix.shape[1]):
      plt.text(first_index, second_index, conf_matrix[first_index, second_index])
  plt.savefig('heatmap_confusion_matrix.jpg')
  plt.show()

结果显示

加载数据
训练数据： (1257, 64)
测试数据： (540, 64)
定义相关参数
构建数据集
定义计算评价指标
定义模型
Model(
  (fc1): Linear(in_features=64, out_features=256, bias=True)
  (relu): ReLU()
  (fc2): Linear(in_features=256, out_features=10, bias=True)
)
定义损失函数、优化器
初始化相关参数
开始训练主循环
[sklearn_metrics] Epoch:0 loss:2.2986 accuracy:0.1098 precision:0.1098 recall:0.1098 f1:0.1098
[sklearn_metrics] Epoch:1 loss:2.2865 accuracy:0.1225 precision:0.1225 recall:0.1225 f1:0.1225
[sklearn_metrics] Epoch:2 loss:2.2637 accuracy:0.1702 precision:0.1702 recall:0.1702 f1:0.1702
[sklearn_metrics] Epoch:3 loss:2.2316 accuracy:0.3174 precision:0.3174 recall:0.3174 f1:0.3174
[sklearn_metrics] Epoch:4 loss:2.1915 accuracy:0.5561 precision:0.5561 recall:0.5561 f1:0.5561
[sklearn_metrics] Epoch:5 loss:2.1438 accuracy:0.6881 precision:0.6881 recall:0.6881 f1:0.6881
[sklearn_metrics] Epoch:6 loss:2.0875 accuracy:0.7669 precision:0.7669 recall:0.7669 f1:0.7669
[sklearn_metrics] Epoch:7 loss:2.0213 accuracy:0.8226 precision:0.8226 recall:0.8226 f1:0.8226
[sklearn_metrics] Epoch:8 loss:1.9428 accuracy:0.8409 precision:0.8409 recall:0.8409 f1:0.8409
[sklearn_metrics] Epoch:9 loss:1.8494 accuracy:0.8552 precision:0.8552 recall:0.8552 f1:0.8552
[sklearn_metrics] Epoch:10 loss:1.7397 accuracy:0.8568 precision:0.8568 recall:0.8568 f1:0.8568
[sklearn_metrics] Epoch:11 loss:1.6140 accuracy:0.8632 precision:0.8632 recall:0.8632 f1:0.8632
[sklearn_metrics] Epoch:12 loss:1.4748 accuracy:0.8616 precision:0.8616 recall:0.8616 f1:0.8616
[sklearn_metrics] Epoch:13 loss:1.3259 accuracy:0.8640 precision:0.8640 recall:0.8640 f1:0.8640
[sklearn_metrics] Epoch:14 loss:1.1735 accuracy:0.8703 precision:0.8703 recall:0.8703 f1:0.8703
[sklearn_metrics] Epoch:15 loss:1.0245 accuracy:0.8791 precision:0.8791 recall:0.8791 f1:0.8791
[sklearn_metrics] Epoch:16 loss:0.8858 accuracy:0.8878 precision:0.8878 recall:0.8878 f1:0.8878
[sklearn_metrics] Epoch:17 loss:0.7625 accuracy:0.9006 precision:0.9006 recall:0.9006 f1:0.9006
[sklearn_metrics] Epoch:18 loss:0.6575 accuracy:0.9045 precision:0.9045 recall:0.9045 f1:0.9045
[sklearn_metrics] Epoch:19 loss:0.5709 accuracy:0.9077 precision:0.9077 recall:0.9077 f1:0.9077
[sklearn_metrics] Epoch:20 loss:0.5004 accuracy:0.9093 precision:0.9093 recall:0.9093 f1:0.9093
[sklearn_metrics] Epoch:21 loss:0.4436 accuracy:0.9101 precision:0.9101 recall:0.9101 f1:0.9101
[sklearn_metrics] Epoch:22 loss:0.3982 accuracy:0.9109 precision:0.9109 recall:0.9109 f1:0.9109
[sklearn_metrics] Epoch:23 loss:0.3615 accuracy:0.9149 precision:0.9149 recall:0.9149 f1:0.9149
[sklearn_metrics] Epoch:24 loss:0.3314 accuracy:0.9173 precision:0.9173 recall:0.9173 f1:0.9173
[sklearn_metrics] Epoch:25 loss:0.3065 accuracy:0.9196 precision:0.9196 recall:0.9196 f1:0.9196
[sklearn_metrics] Epoch:26 loss:0.2856 accuracy:0.9228 precision:0.9228 recall:0.9228 f1:0.9228
[sklearn_metrics] Epoch:27 loss:0.2673 accuracy:0.9236 precision:0.9236 recall:0.9236 f1:0.9236
[sklearn_metrics] Epoch:28 loss:0.2512 accuracy:0.9268 precision:0.9268 recall:0.9268 f1:0.9268
[sklearn_metrics] Epoch:29 loss:0.2370 accuracy:0.9300 precision:0.9300 recall:0.9300 f1:0.9300
              precision    recall  f1-score   support

           0       0.98      0.98      0.98        59
           1       0.86      0.86      0.86        56
           2       0.98      0.91      0.94        53
           3       0.98      0.93      0.96        46
           4       0.95      0.97      0.96        61
           5       0.98      0.91      0.95        57
           6       0.96      0.96      0.96        57
           7       0.92      0.98      0.95        50
           8       0.87      0.81      0.84        48
           9       0.77      0.91      0.83        53

    accuracy                           0.92       540
   macro avg       0.93      0.92      0.92       540
weighted avg       0.93      0.92      0.92       540

[[58  0  0  0  1  0  0  0  0  0]
 [ 0 48  0  0  0  0  1  0  0  7]
 [ 0  2 48  0  0  0  0  1  2  0]
 [ 0  0  1 43  0  0  0  1  1  0]
 [ 0  0  0  0 59  0  0  1  1  0]
 [ 0  0  0  0  1 52  0  0  0  4]
 [ 1  1  0  0  0  0 55  0  0  0]
 [ 0  0  0  0  0  0  0 49  0  1]
 [ 0  4  0  0  1  1  1  0 39  2]
 [ 0  1  0  1  0  0  0  1  2 48]]
<Figure size 640x480 with 2 Axes>
[sklearn_metrics] accuracy:0.9241 precision:0.9241 recall:0.9241 f1:0.9241

评价指标相关：准确率-精确率-召回率-f1

（1）基本知识 之前我们将pytorch加载数据、建立模型、训练和测试、使用sklearn评估模型都完整的过了一遍，接下来我们要再细讲下评价指标。首先大体的讲下四个基本的评价指标（针对于多分类）： accuracy：准确率。准确率就是有多少数据被正确识别了。针对整体，比如预测标签[0,1,1,3,2,2,1]，真实标签[1,2,1,3,2,1,1]，此时正确率就是：5 / 7 = 0.7142，5是指列表中有5个对应的位置是相同的，即第0 2 3 4 6个位置，7是列表总长度。 precision：精确率，就是在预测为正的数据中，有多少是正确的。这里以标签为1的那一类说明，对于标签不为1的，我们先全部置为0（这里0表示的是负样本，不是第0类），则有：预测标签[0,1,1,0,0,0,1] 真实标签[1,0,1,0,0,1,1] 在预测标签中，有3个预测为正，也就是1，在这三个中，有2个是与真实标签相同，也就是第2 6个位置，则精确率就是：2 / 3 = 0.6666 recall：在正样本中，有多少被正确识别。还是以标签为1的进行说明：在真实标签中有4个为1，在这4个中有2个被预测出来，所以召回率就是：2 / 4 =0.5000。 f1：综合考虑精确率和召回率。其值就是2 * p * r) / (p + r)

（2）具体计算 使用到的就是TP、FP、FN、TN，分别解释一下这些是什么：第一位是True False的意思，第二位是Positive Negative。相当于第一位是对第二位的一个判断。 TP，即True Positive，预测为Positive的是True，也就是预测为正的，真实值是正。 FP，即False Positive，预测为Positive的是False，也就是预测为正的，真实值是负。 FN，即False Negative，预测为Negative的是False，也就是预测为负的，真实值是正。 TN，即True Negative，预测为Negative的是True，也就是预测为负的，真实值是是负。那么根据之前我们的定义：准确率（accuracy）不就是：(TP + FN) / (TP + FP + FN + TN) 精确率（precision）不就是：(TP) / (TP + FP) 召回率（recall）不就是：(TP) / (TP + FN) f1 不就是：2 * precision * recall / (precision + recall)

（3）micro-f1和macro-f1 简单来讲，micro-f1就是先计算每一类的TP、FP、FN、TN，再计算相关的评价指标，在数据不平衡的情况下考虑到了每一类的数量。macro-f1就是先计算每一类的评价指标，最后取平均，它易受精确率和召回率较大的类的影响。

基本实现

接下来，我们要根据理解去实现评价指标。

（1）基本实现

def get_acc_p_r_f1(trues, preds):
  labels = [0,1,2,3,4,5,6,7,8,9]
  TP,FP,FN,TN = 0,0,0,0
  for label in labels:
    preds_tmp = np.array([1 if pred == label else 0 for pred in preds])
    trues_tmp = np.array([1 if true == label else 0 for true in trues])
    # print(preds_tmp, trues_tmp)
    # print()
    # TP预测为1真实为1
    # TN预测为0真实为0
    # FN预测为0真实为1
    # FP预测为1真实为0
    TP += ((preds_tmp == 1) & (trues_tmp == 1)).sum()
    TN += ((preds_tmp == 0) & (trues_tmp == 0)).sum()
    FN += ((preds_tmp == 0) & (trues_tmp == 1)).sum()
    FP += ((preds_tmp == 1) & (trues_tmp == 0)).sum()
  # print(TP, FP, FN)
  precision = TP / (TP + FP)
  recall = TP / (TP + FN)
  f1 = 2 * precision * recall / (precision + recall)
  return precision, recall, f1

def get_acc(trues, preds):
  accuracy = (np.array(trues) == np.array(preds)).sum() / len(trues)
  return accuracy

具体的就不细讲了，代码很容易看懂。

（2）根据混淆矩阵实现

def get_p_r_f1_from_conf_matrix(conf_matrix):
  TP,FP,FN,TN = 0,0,0,0
  labels = [0,1,2,3,4,5,6,7,8,9]
  nums = len(labels)
  for i in labels:
    TP += conf_matrix[i, i]
    FP += (conf_matrix[:i, i].sum() + conf_matrix[i+1:, i].sum())
    FN += (conf_matrix[i, i+1:].sum() + conf_matrix[i, :i].sum())
  print(TP, FP, FN)
  precision = TP / (TP + FP)
  recall = TP / (TP + FN)
  f1 = 2 * precision * recall / (precision + recall)
  return precision, recall, f1

def get_acc_from_conf_matrix(conf_matrix):
  labels = [0,1,2,3,4,5,6,7,8,9]
  return sum([conf_matrix[i, i] for i in range(len(labels))]) / np.sum(np.sum(conf_matrix, axis=0))

最终结果

加载数据
训练数据： (1257, 64)
测试数据： (540, 64)
定义相关参数
构建数据集
定义计算评价指标
定义模型
Model(
  (fc1): Linear(in_features=64, out_features=256, bias=True)
  (relu): ReLU()
  (fc2): Linear(in_features=256, out_features=10, bias=True)
)
定义损失函数、优化器
初始化相关参数
开始训练主循环
[custom_metrics] Epoch:0 loss:2.2986 accuracy:0.1098 precision:0.1098 recall:0.1098 f1:0.1098
[sklearn_metrics] Epoch:0 loss:2.2986 accuracy:0.1098 precision:0.1098 recall:0.1098 f1:0.1098
[custom_metrics] Epoch:1 loss:2.2865 accuracy:0.1225 precision:0.1225 recall:0.1225 f1:0.1225
[sklearn_metrics] Epoch:1 loss:2.2865 accuracy:0.1225 precision:0.1225 recall:0.1225 f1:0.1225
[custom_metrics] Epoch:2 loss:2.2637 accuracy:0.1702 precision:0.1702 recall:0.1702 f1:0.1702
[sklearn_metrics] Epoch:2 loss:2.2637 accuracy:0.1702 precision:0.1702 recall:0.1702 f1:0.1702
[custom_metrics] Epoch:3 loss:2.2316 accuracy:0.3174 precision:0.3174 recall:0.3174 f1:0.3174
[sklearn_metrics] Epoch:3 loss:2.2316 accuracy:0.3174 precision:0.3174 recall:0.3174 f1:0.3174
[custom_metrics] Epoch:4 loss:2.1915 accuracy:0.5561 precision:0.5561 recall:0.5561 f1:0.5561
[sklearn_metrics] Epoch:4 loss:2.1915 accuracy:0.5561 precision:0.5561 recall:0.5561 f1:0.5561
[custom_metrics] Epoch:5 loss:2.1438 accuracy:0.6881 precision:0.6881 recall:0.6881 f1:0.6881
[sklearn_metrics] Epoch:5 loss:2.1438 accuracy:0.6881 precision:0.6881 recall:0.6881 f1:0.6881
[custom_metrics] Epoch:6 loss:2.0875 accuracy:0.7669 precision:0.7669 recall:0.7669 f1:0.7669
[sklearn_metrics] Epoch:6 loss:2.0875 accuracy:0.7669 precision:0.7669 recall:0.7669 f1:0.7669
[custom_metrics] Epoch:7 loss:2.0213 accuracy:0.8226 precision:0.8226 recall:0.8226 f1:0.8226
[sklearn_metrics] Epoch:7 loss:2.0213 accuracy:0.8226 precision:0.8226 recall:0.8226 f1:0.8226
[custom_metrics] Epoch:8 loss:1.9428 accuracy:0.8409 precision:0.8409 recall:0.8409 f1:0.8409
[sklearn_metrics] Epoch:8 loss:1.9428 accuracy:0.8409 precision:0.8409 recall:0.8409 f1:0.8409
[custom_metrics] Epoch:9 loss:1.8494 accuracy:0.8552 precision:0.8552 recall:0.8552 f1:0.8552
[sklearn_metrics] Epoch:9 loss:1.8494 accuracy:0.8552 precision:0.8552 recall:0.8552 f1:0.8552
[custom_metrics] Epoch:10 loss:1.7397 accuracy:0.8568 precision:0.8568 recall:0.8568 f1:0.8568
[sklearn_metrics] Epoch:10 loss:1.7397 accuracy:0.8568 precision:0.8568 recall:0.8568 f1:0.8568
[custom_metrics] Epoch:11 loss:1.6140 accuracy:0.8632 precision:0.8632 recall:0.8632 f1:0.8632
[sklearn_metrics] Epoch:11 loss:1.6140 accuracy:0.8632 precision:0.8632 recall:0.8632 f1:0.8632
[custom_metrics] Epoch:12 loss:1.4748 accuracy:0.8616 precision:0.8616 recall:0.8616 f1:0.8616
[sklearn_metrics] Epoch:12 loss:1.4748 accuracy:0.8616 precision:0.8616 recall:0.8616 f1:0.8616
[custom_metrics] Epoch:13 loss:1.3259 accuracy:0.8640 precision:0.8640 recall:0.8640 f1:0.8640
[sklearn_metrics] Epoch:13 loss:1.3259 accuracy:0.8640 precision:0.8640 recall:0.8640 f1:0.8640
[custom_metrics] Epoch:14 loss:1.1735 accuracy:0.8703 precision:0.8703 recall:0.8703 f1:0.8703
[sklearn_metrics] Epoch:14 loss:1.1735 accuracy:0.8703 precision:0.8703 recall:0.8703 f1:0.8703
[custom_metrics] Epoch:15 loss:1.0245 accuracy:0.8791 precision:0.8791 recall:0.8791 f1:0.8791
[sklearn_metrics] Epoch:15 loss:1.0245 accuracy:0.8791 precision:0.8791 recall:0.8791 f1:0.8791
[custom_metrics] Epoch:16 loss:0.8858 accuracy:0.8878 precision:0.8878 recall:0.8878 f1:0.8878
[sklearn_metrics] Epoch:16 loss:0.8858 accuracy:0.8878 precision:0.8878 recall:0.8878 f1:0.8878
[custom_metrics] Epoch:17 loss:0.7625 accuracy:0.9006 precision:0.9006 recall:0.9006 f1:0.9006
[sklearn_metrics] Epoch:17 loss:0.7625 accuracy:0.9006 precision:0.9006 recall:0.9006 f1:0.9006
[custom_metrics] Epoch:18 loss:0.6575 accuracy:0.9045 precision:0.9045 recall:0.9045 f1:0.9045
[sklearn_metrics] Epoch:18 loss:0.6575 accuracy:0.9045 precision:0.9045 recall:0.9045 f1:0.9045
[custom_metrics] Epoch:19 loss:0.5709 accuracy:0.9077 precision:0.9077 recall:0.9077 f1:0.9077
[sklearn_metrics] Epoch:19 loss:0.5709 accuracy:0.9077 precision:0.9077 recall:0.9077 f1:0.9077
[custom_metrics] Epoch:20 loss:0.5004 accuracy:0.9093 precision:0.9093 recall:0.9093 f1:0.9093
[sklearn_metrics] Epoch:20 loss:0.5004 accuracy:0.9093 precision:0.9093 recall:0.9093 f1:0.9093
[custom_metrics] Epoch:21 loss:0.4436 accuracy:0.9101 precision:0.9101 recall:0.9101 f1:0.9101
[sklearn_metrics] Epoch:21 loss:0.4436 accuracy:0.9101 precision:0.9101 recall:0.9101 f1:0.9101
[custom_metrics] Epoch:22 loss:0.3982 accuracy:0.9109 precision:0.9109 recall:0.9109 f1:0.9109
[sklearn_metrics] Epoch:22 loss:0.3982 accuracy:0.9109 precision:0.9109 recall:0.9109 f1:0.9109
[custom_metrics] Epoch:23 loss:0.3615 accuracy:0.9149 precision:0.9149 recall:0.9149 f1:0.9149
[sklearn_metrics] Epoch:23 loss:0.3615 accuracy:0.9149 precision:0.9149 recall:0.9149 f1:0.9149
[custom_metrics] Epoch:24 loss:0.3314 accuracy:0.9173 precision:0.9173 recall:0.9173 f1:0.9173
[sklearn_metrics] Epoch:24 loss:0.3314 accuracy:0.9173 precision:0.9173 recall:0.9173 f1:0.9173
[custom_metrics] Epoch:25 loss:0.3065 accuracy:0.9196 precision:0.9196 recall:0.9196 f1:0.9196
[sklearn_metrics] Epoch:25 loss:0.3065 accuracy:0.9196 precision:0.9196 recall:0.9196 f1:0.9196
[custom_metrics] Epoch:26 loss:0.2856 accuracy:0.9228 precision:0.9228 recall:0.9228 f1:0.9228
[sklearn_metrics] Epoch:26 loss:0.2856 accuracy:0.9228 precision:0.9228 recall:0.9228 f1:0.9228
[custom_metrics] Epoch:27 loss:0.2673 accuracy:0.9236 precision:0.9236 recall:0.9236 f1:0.9236
[sklearn_metrics] Epoch:27 loss:0.2673 accuracy:0.9236 precision:0.9236 recall:0.9236 f1:0.9236
[custom_metrics] Epoch:28 loss:0.2512 accuracy:0.9268 precision:0.9268 recall:0.9268 f1:0.9268
[sklearn_metrics] Epoch:28 loss:0.2512 accuracy:0.9268 precision:0.9268 recall:0.9268 f1:0.9268
[custom_metrics] Epoch:29 loss:0.2370 accuracy:0.9300 precision:0.9300 recall:0.9300 f1:0.9300
[sklearn_metrics] Epoch:29 loss:0.2370 accuracy:0.9300 precision:0.9300 recall:0.9300 f1:0.9300
              precision    recall  f1-score   support

           0       0.98      0.98      0.98        59
           1       0.86      0.86      0.86        56
           2       0.98      0.91      0.94        53
           3       0.98      0.93      0.96        46
           4       0.95      0.97      0.96        61
           5       0.98      0.91      0.95        57
           6       0.96      0.96      0.96        57
           7       0.92      0.98      0.95        50
           8       0.87      0.81      0.84        48
           9       0.77      0.91      0.83        53

    accuracy                           0.92       540
   macro avg       0.93      0.92      0.92       540
weighted avg       0.93      0.92      0.92       540

[[58  0  0  0  1  0  0  0  0  0]
 [ 0 48  0  0  0  0  1  0  0  7]
 [ 0  2 48  0  0  0  0  1  2  0]
 [ 0  0  1 43  0  0  0  1  1  0]
 [ 0  0  0  0 59  0  0  1  1  0]
 [ 0  0  0  0  1 52  0  0  0  4]
 [ 1  1  0  0  0  0 55  0  0  0]
 [ 0  0  0  0  0  0  0 49  0  1]
 [ 0  4  0  0  1  1  1  0 39  2]
 [ 0  1  0  1  0  0  0  1  2 48]]
<Figure size 640x480 with 2 Axes>
[custom_metrics] accuracy:0.9241 precision:0.9241 recall:0.9241 f1:0.9241
[sklearn_metrics] accuracy:0.9241 precision:0.9241 recall:0.9241 f1:0.9241
[cm_metrics] accuracy:0.9241 precision:0.9241 recall:0.9241 f1:0.9241

我们计算出的和sklearn自带的计算出的结果是一样的。为了确保是正确的，这里我们再打印一下测试的时候的每一类的精确率、召回率和micro-f1。

[custom_metrics] 0 precision:0.9831 recall:0.9831 f1:0.9831
[custom_metrics] 1 precision:0.8571 recall:0.8571 f1:0.8571
[custom_metrics] 2 precision:0.9796 recall:0.9057 f1:0.9412
[custom_metrics] 3 precision:0.9773 recall:0.9348 f1:0.9556
[custom_metrics] 4 precision:0.9516 recall:0.9672 f1:0.9593
[custom_metrics] 5 precision:0.9811 recall:0.9123 f1:0.9455
[custom_metrics] 6 precision:0.9649 recall:0.9649 f1:0.9649
[custom_metrics] 7 precision:0.9245 recall:0.9800 f1:0.9515
[custom_metrics] 8 precision:0.8667 recall:0.8125 f1:0.8387
[custom_metrics] 9 precision:0.7742 recall:0.9057 f1:0.8348

[cm_metrics] 0 precision:0.9831 recall:0.9831 f1:0.9831
[cm_metrics] 1 precision:0.8571 recall:0.8571 f1:0.8571
[cm_metrics] 2 precision:0.9796 recall:0.9057 f1:0.9412
[cm_metrics] 3 precision:0.9773 recall:0.9348 f1:0.9556
[cm_metrics] 4 precision:0.9516 recall:0.9672 f1:0.9593
[cm_metrics] 5 precision:0.9811 recall:0.9123 f1:0.9455
[cm_metrics] 6 precision:0.9649 recall:0.9649 f1:0.9649
[cm_metrics] 7 precision:0.9245 recall:0.9800 f1:0.9515
[cm_metrics] 8 precision:0.8667 recall:0.8125 f1:0.8387
[cm_metrics] 9 precision:0.7742 recall:0.9057 f1:0.8348

和sklearn中的classification_report是一致的。

绘制ROC和计算AUC

最后的最后，绘制ROC曲线和计算AUC，这两个评价指标就偷个懒，不介绍了，先要将标签进行二值化：

def get_roc_auc(trues, preds):
  labels = [0,1,2,3,4,5,6,7,8,9]
  nb_classes = len(labels)
  fpr = dict()
  tpr = dict()
  roc_auc = dict()
  print(trues, preds)
  for i in range(nb_classes):
    fpr[i], tpr[i], _ = roc_curve(trues[:, i], preds[:, i])
    roc_auc[i] = auc(fpr[i], tpr[i])
    # Compute micro-average ROC curve and ROC area
  fpr["micro"], tpr["micro"], _ = roc_curve(trues.ravel(), preds.ravel())
  roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
  # First aggregate all false positive rates
  all_fpr = np.unique(np.concatenate([fpr[i] for i in range(nb_classes)]))
  # Then interpolate all ROC curves at this points
  mean_tpr = np.zeros_like(all_fpr)
  for i in range(nb_classes):
    mean_tpr += interp(all_fpr, fpr[i], tpr[i])
  # Finally average it and compute AUC
  mean_tpr /= nb_classes
  fpr["macro"] = all_fpr
  tpr["macro"] = mean_tpr
  roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
  # Plot all ROC curves
  lw = 2
  plt.figure()
  plt.plot(fpr["micro"], tpr["micro"],label='micro-average ROC curve (area = {0:0.2f})'.format(roc_auc["micro"]),color='deeppink', linestyle=':', linewidth=4)
  plt.plot(fpr["macro"], tpr["macro"],label='macro-average ROC curve (area = {0:0.2f})'.format(roc_auc["macro"]),color='navy', linestyle=':', linewidth=4)
  colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
  for i, color in zip(range(nb_classes), colors):
    plt.plot(fpr[i], tpr[i], color=color, lw=lw, label='ROC curve of class {0} (area = {1:0.2f})'.format(i, roc_auc[i]))
  plt.plot([0, 1], [0, 1], 'k--', lw=lw)
  plt.xlim([0.0, 1.0])
  plt.ylim([0.0, 1.05])
  plt.xlabel('False Positive Rate')
  plt.ylabel('True Positive Rate')
  plt.title('Some extension of Receiver operating characteristic to multi-class')
  plt.legend(loc="lower right")
  plt.savefig("ROC_10分类.png")
  plt.show()
  
test_trues = label_binarize(test_trues, classes=[i for i in range(10)])
test_preds = label_binarize(test_preds, classes=[i for i in range(10)])
get_roc_auc(test_trues, test_preds)

一文弄懂pytorch搭建网络流程+多分类评价指标