Python训练和加载Tensorflow/Keras模型

嘿,我想保存并加载我训练过的模型,但这样做会遇到一些问题。

在训练python脚本中,在训练模型后,它对字符串“nobody like you”进行预测,并给出正确的结果“Spam:96.0959%”。

然而,当我在python测试代码中尝试它时,我用相同的权重加载相同的模型,对于相同的文本,它会给我一个完全不同的读数,例如“批准:49%”。

所以我不确定,但我认为我以错误的方式保存或加载了经过训练的模型。

我想要实现的是,首先训练一个模型,保存该模型,并能够使用训练后的模型进行文本预测,而无需每次都对其进行重新训练。

我的培训代码片段:

# Convert labels to numerical format
label_to_index = {
    'Approve': 0,
    'Spam': 1,
    ...
}

# Tokenization and Padding
tokenizer = Tokenizer()
tokenizer.fit_on_texts(posts)
num_unique_words = len(tokenizer.word_index)
tokenizer.num_words = num_unique_words
sequences = tokenizer.texts_to_sequences(posts)
max_sequence_length = 7500;

X = pad_sequences(sequences, maxlen=max_sequence_length)
y = np.array([label_to_index[label] for label in labels])

# Split data into training and testing sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Neural Network Architecture
model = Sequential()
vocab_size = tokenizer.num_words
model.add(Embedding(input_dim=vocab_size, output_dim=50, input_length=max_sequence_length))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dense(19, activation='softmax'))

# Model Compilation
model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

# Model Training
history = model.fit(X_train, y_train, epochs=10, batch_size=32, validation_data=(X_test, y_test))

model_architecture = model.to_json()
with open('model_architecture_old.json', 'w') as json_file:
    json_file.write(model_architecture)

model.save_weights('model_weights_old.h5')

我加载并测试节省的权重和json,这样做:

# Load the model architecture from the JSON file
with open('model_architecture_old.json', 'r') as json_file:
    model_architecture = json_file.read()

# Create a new model instance using the loaded architecture
loaded_model = model_from_json(model_architecture)

# Load the saved weights from the checkpoint file
loaded_model.load_weights('model_weights_old.h5')

# Compile the model before using it (same compilation as during training)
loaded_model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

# Get user input for testing
num_texts = int(input("Enter the number of texts you want to test: "))
texts = []
for i in range(num_texts):
    text = input(f"Enter text {i+1}: ")
    texts.append(text)

# Tokenization and Padding
tokenizer = Tokenizer()
tokenizer.fit_on_texts(texts)  # Replace 'texts' with your text data
sequences = tokenizer.texts_to_sequences(texts)
max_sequence_length = 7500;
X = pad_sequences(sequences, maxlen=max_sequence_length)

# Make predictions using the loaded model
predictions = loaded_model.predict(X)