详解数据分析中的方差，标准差和异常值的使用

In previous post, we have trained a machine learning model to classify sentiments of movie reviews. In this post, we will evaluate the performance of the model and save the model for future use.

Evaluate Model

To evaluate the performance of the model, we will use the test dataset. Using the evaluate method of the model to calculate the loss and accuracy of the model on the test dataset.

# evaluate model on test dataset
loss, accuracy = trained_model.evaluate(test_dataset)
print("Test accuracy:", accuracy)
print("Test loss:", loss)

The evaluate method returns two values, the loss and the accuracy of the model on the test dataset. We can use these values to evaluate the performance of the model. below is result of the evaluation on the test dataset:

Test accuracy: 0.8694400191307068
Test loss: 0.3562678098678589

Save Model

To save the model for future use, we will use the save method of the model. The trained model will be saved in the local machine as a .keras file. The model file format extension can be .h5, .tf or .keras.

def save_model(vectorize_layer, model, model_saved_path):
    export_model = tf.keras.Sequential([
        vectorize_layer, # text vectorization layer
        model # the model trained before
    ])

    export_model.compile(
        loss=losses.BinaryCrossentropy(from_logits=False), 
        optimizer="adam", 
        metrics=['accuracy']
    )

    # save model
    export_model.save(model_saved_path)
    return export_model

In above code, we first create a new model by combining the vectorize_layer and the trained model. We then compile the model with the same loss function, optimizer and metrics used during training. Finally, we save the model using the save method and return the saved model.

Here is a thing need to be noticed, we need to put the vectorize_layer as the first layer of the model so that it can be used to preprocess the text data before passing it to the trained model.

In the next blog, we will introduce how to load the model and use it to make predictions on new data.