多标签CNN模型预测输出数字数组，如何转换为对应标签？

我是神经网络新手，训练了一个多标签CNN模型，用model.predict(image)对新图像预测时，得到的是数字数组而非对应标签，输出如下：

1/1 [==============================] - 0s 294ms/step
array([[1.2423882e-07, 4.8644133e-10, 6.2801077e-11, ..., 5.9059038e-09,
    7.9583238e-08, 1.8839316e-07],
   [6.7338499e-07, 3.7959146e-11, 6.3526181e-12, ..., 6.6384481e-10,
    6.5803110e-09, 7.2412973e-09],
   [3.4079651e-06, 1.3292389e-12, 1.1482074e-09, ..., 2.3795290e-09,
    4.0437489e-09, 2.5665909e-09],
   ...,
   [4.5694168e-09, 5.8126518e-14, 7.4569753e-18, ..., 9.1445762e-16,
    8.6637960e-15, 1.7640524e-11],
   [7.3903612e-09, 1.1566522e-13, 6.5436875e-17, ..., 7.0952320e-15,
    8.3692771e-14, 1.2283501e-10],
   [7.1919551e-08, 1.9803466e-14, 2.5265929e-16, ..., 5.3995962e-13,
    5.4146050e-12, 3.9835553e-09]], dtype=float32)

我使用的预测代码如下：

from PIL import Image
import numpy as np
from skimage import transform

def load(filename):
  np_image = Image.open(filename)
  np_image = np.array(np_image).astype('float32')/255
  np_image = transform.resize(np_image, (256, 256, 3))
  np_image = np.expand_dims(np_image, axis=0)
  return np_image

image = load('/content/1000_IM-0003-1001.dcm.png')
loaded_2.predict(image)

希望能将预测结果转换为对应的标签，而非数字数组。

模型架构 #

Model: "sequential_1"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_4 (Conv2D)           (None, 196, 196, 16)      1216      
                                                                 
 batch_normalization_4 (BatchNormalization)  (None, 196, 196, 16)     64        
                                                                 
 max_pooling2d_4 (MaxPooling2D)              (None, 98, 98, 16)       0         
                                                                 
 dropout_6 (Dropout)         (None, 98, 98, 16)        0         
                                                                 
 conv2d_5 (Conv2D)           (None, 94, 94, 32)        12832     
                                                                 
 max_pooling2d_5 (MaxPooling2D)              (None, 47, 47, 32)       0         
                                                                 
 batch_normalization_5 (BatchNormalization)  (None, 47, 47, 32)       128       
                                                                 
 dropout_7 (Dropout)         (None, 47, 47, 32)        0         
                                                                 
 conv2d_6 (Conv2D)           (None, 43, 43, 64)        51264     
                                                                 
 max_pooling2d_6 (MaxPooling2D)              (None, 21, 21, 64)       0         
                                                                 
 batch_normalization_6 (BatchNormalization)  (None, 21, 21, 64)       256       
                                                                 
 dropout_8 (Dropout)         (None, 21, 21, 64)        0         
                                                                 
 conv2d_7 (Conv2D)           (None, 17, 17, 64)        102464    
                                                                 
 max_pooling2d_7 (MaxPooling2D)              (None, 8, 8, 64)         0         
                                                                 
 batch_normalization_7 (BatchNormalization)  (None, 8, 8, 64)         256       
                                                                 
 dropout_9 (Dropout)         (None, 8, 8, 64)          0         
                                                                 
 flatten_1 (Flatten)         (None, 4096)              0         
                                                                 
 dense_3 (Dense)             (None, 128)               524416    
                                                                 
 dropout_10 (Dropout)        (None, 128)               0         
                                                                 
 dense_4 (Dense)             (None, 64)                8256      
                                                                 
 dropout_11 (Dropout)        (None, 64)                0         
                                                                 
 dense_5 (Dense)             (None, 524)               34060     
                                                                 
=================================================================
Total params: 735,212
Trainable params: 734,860
Non-trainable params: 352

标签信息 #

标签采用one-hot编码，形状如下：

y.shape
(4312, 524)

1. 理解预测输出含义 #

模型最后一层是Dense(524)，未指定激活函数时默认用线性激活，输出的是原始logit值。多标签分类任务通常需要用sigmoid激活将输出转为0-1之间的概率值，再通过阈值判断标签是否激活。

如果训练时已用sigmoid激活，输出直接是概率值；否则先做转换：

predictions = loaded_2.predict(image)
probabilities = tf.nn.sigmoid(predictions).numpy()  # 转为概率值

2. 阈值筛选标签 #

多标签分类中每个标签独立判断，设定阈值（比如0.5），筛选出概率大于阈值的标签索引：

threshold = 0.5
predicted_indices = np.where(probabilities[0] > threshold)[0]  # 取单张图片的预测结果

3. 映射到真实标签名称 #

你需要有一个标签索引与名称的映射列表（顺序必须和one-hot编码一致），通过索引获取对应标签：

# 替换为你训练时使用的真实标签名称列表
label_names = ["标签1", "标签2", ..., "标签524"]

predicted_labels = [label_names[idx] for idx in predicted_indices]
print("预测标签：", predicted_labels)

完整示例代码 #

import tensorflow as tf
import numpy as np
from PIL import Image
from skimage import transform

def load(filename):
  np_image = Image.open(filename)
  np_image = np.array(np_image).astype('float32')/255
  np_image = transform.resize(np_image, (256, 256, 3))
  np_image = np.expand_dims(np_image, axis=0)
  return np_image

image = load('/content/1000_IM-0003-1001.dcm.png')

# 预测并转换为标签
predictions = loaded_2.predict(image)
# 若模型最后一层无sigmoid激活，启用下面一行
probabilities = tf.nn.sigmoid(predictions).numpy()

threshold = 0.5
predicted_indices = np.where(probabilities[0] > threshold)[0]

# 替换为你的真实标签名称列表
label_names = ["标签1", "标签2", ..., "标签524"]

predicted_labels = [label_names[i] for i in predicted_indices]
print("预测的标签：", predicted_labels)

注意事项 #

• 阈值可根据任务调整：想减少误判就提高阈值（如0.7），想召回更多标签就降低阈值（如0.3）。
• 必须保证label_names的顺序和训练时one-hot编码的顺序完全一致，否则会出现标签映射错误。
• 若训练时模型最后一层已用sigmoid激活，无需再调用tf.nn.sigmoid，直接用predictions作为概率值即可。

内容的提问来源于stack exchange，提问作者Saad Khattak