如何用TensorFlow模型对编码后的句子进行正确预测

你的问题主要出在三个地方：预测时的输入格式错误、未保留标签编码器实例、未将模型输出的概率数组转换为原始标签。以下是具体修改步骤和完整代码：

关键修改点 #

• 保存LabelEncoder实例：原label_encoding函数仅返回编码后的标签，现在需同时返回编码器实例，才能将预测得到的索引转回原始标签。
• 修正预测输入格式：encode_sentences接收的是句子列表，预测单个句子时必须将其放入列表传入，否则函数会把字符串拆成单个字符处理，导致维度错误。
• 解析模型输出：模型用softmax输出的是每个类别的概率，需先找到概率最高的类别索引，再通过编码器映射回原始标签。

修改后的完整代码 #

import numpy as np
import tensorflow as tf
from tensorflow import keras
import pandas as pd
import sqlite3
from sklearn.preprocessing import LabelEncoder
import spacy
from sklearn.model_selection import train_test_split


nlp = spacy.load('en_core_web_lg')
embedding_dim = nlp.vocab.vectors_length


def read_database(path):
    # Loading Data from database
    connection = sqlite3.connect(path)
    db_rows = pd.read_sql('''select intents.intent, patterns.pattern 
    from intents, patterns where intents.id = patterns.intentid''', connection)
    labels = []
    sentences = []
    intents = []
    for i in range(len(db_rows)):
        labels.append(db_rows["intent"][i])
        if db_rows["intent"][i] not in intents:
            intents.append(db_rows["intent"][i])
        sentences.append(db_rows["pattern"][i])
    return sentences, labels, intents


def label_encoding(labels):
    # Calculate the length of labels
    n_labels = len(labels)
    print('Number of labels :-', n_labels)
    le = LabelEncoder()
    y = le.fit_transform(labels)
    print('Length of y :- ', y.shape)
    # 返回编码器实例和编码后的标签
    return y, le


def encode_sentences(sentences):
    # Calculate number of sentences
    n_sentences = len(sentences)

    X = np.zeros((n_sentences, embedding_dim))

    # Iterate over the sentences
    for idx, sentence in enumerate(sentences):
        doc = nlp(sentence)
        # Save the document's .vector attribute to the corresponding row in
        X[idx, :] = doc.vector
    return X


sentences_train, labels_train, all_intents = read_database('./database_x.db')
sentences_train = encode_sentences(sentences_train)
# 接收编码后的标签和编码器
labels_train, label_encoder = label_encoding(labels_train)
x_train, x_test, y_train, y_test = train_test_split(sentences_train, labels_train, test_size=0.2)

model = keras.Sequential([keras.layers.Dense(16, activation='relu'),
                          keras.layers.Dense(16, activation='relu'),
                          keras.layers.Dense(len(all_intents), activation='softmax')])

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.01),
              loss=keras.losses.SparseCategoricalCrossentropy(),
              metrics=['accuracy'])

model.fit(x_train, y_train, batch_size=16, epochs=100)

# 预测时传入包含单个句子的列表
test_sentence = ["how can i test rf heating"]
encoded_sentence = encode_sentences(test_sentence)
prediction = model.predict(encoded_sentence)

# 取概率最大的索引，再转成原始标签
predicted_idx = np.argmax(prediction, axis=1)[0]
predicted_label = label_encoder.inverse_transform([predicted_idx])[0]

print("\n预测概率数组：")
print(prediction)
print("\n预测的标签：")
print(predicted_label)

代码说明 #

1. label_encoding函数修改：新增返回LabelEncoder实例，后续通过le.inverse_transform将索引转换为原始标签。
2. 预测输入修正：将单个句子放入列表中传入encode_sentences，确保生成维度正确的向量（对应en_core_web_lg的300维词向量）。
3. 预测结果解析：
   • np.argmax(prediction, axis=1)找出概率最高的类别索引；
   • label_encoder.inverse_transform([predicted_idx])将索引转回原始的intent标签。

内容的提问来源于stack exchange，提问作者kaan46