CNN恶意URL分类模型训练正常，Flask部署后预测失效求助

我近期开展了一个机器学习项目，训练了一款用于恶意URL分类的CNN模型。该模型在训练集和测试集上表现均良好，但将模型保存后，在EC2实例的Flask环境中重新加载并进行预测时，表现极差，仿佛从未经过训练。项目思路参考了一个GitHub仓库，恳请告知可能的问题原因。

模型构建代码 #

class CNN(object):
    def __init__(self) -> None:
        super(CNN, self).__init__()
        self.max_len = 75
        self.emb_dim = 32
        self.max_vocab_len = 100
        self.W_reg = regularizers.l2(1e-4)

    def sum_1d(self, X):
        return K.sum(X, axis=1)

    def get_conv_layer(self, emb, kernel_size=5, filters=256):
        conv = Convolution1D(kernel_size=kernel_size, filters=filters,
                             padding='same')(emb)
        conv = ELU()(conv)

        conv = Lambda(self.sum_1d, output_shape=(filters,))(conv)

        # conv = BatchNormalization()(conv)

        conv = Dropout(0.5)(conv)
        return conv

    def build_model(self):
        main_input = Input(shape=(self.max_len,),
                           dtype='int32', name='main_input')
        emb = Embedding(input_dim=self.max_vocab_len, output_dim=self.emb_dim, input_length=self.max_len,
                        embeddings_regularizer=self.W_reg)(main_input)
        emb = Dropout(0.25)(emb)

        conv1 = self.get_conv_layer(emb, kernel_size=2, filters=256)
        conv2 = self.get_conv_layer(emb, kernel_size=3, filters=256)
        conv3 = self.get_conv_layer(emb, kernel_size=4, filters=256)
        conv4 = self.get_conv_layer(emb, kernel_size=5, filters=256)

        merged = concatenate([conv1, conv2, conv3, conv4], axis=1)

        hidden1 = Dense(1024)(merged)
        hidden1 = ELU()(hidden1)
        hidden1 = BatchNormalization()(hidden1)
        hidden1 = Dropout(0.5)(hidden1)

        hidden2 = Dense(1024)(hidden1)
        hidden2 = ELU()(hidden2)
        hidden2 = BatchNormalization()(hidden2)
        hidden2 = Dropout(0.5)(hidden2)

        output = Dense(1, activation='sigmoid', name='output')(hidden2)

        model = Model(inputs=[main_input], outputs=[output])

        adam = Adam(lr=1e-4, beta_1=0.9, beta_2=0.999,
                    epsilon=1e-08, decay=0.0)
        model.compile(optimizer=adam, loss='binary_crossentropy',
                      metrics=['accuracy'])
        return 

模型保存代码 #

# Function to Save Model
def save_model(model, json_dir, weights_dir):
    # have h5py installed
    if Path(json_dir).is_file():
        os.remove(json_dir)
    json_string = model.to_json()
    with open(json_dir, 'w') as f:
        json.dump(json_string, f)

    if Path(weights_dir).is_file():
        os.remove(weights_dir)
    model.save_weights(weights_dir)

模型加载代码 #

def load_model():
    json_dir = os.path.join('models\cnn_lstm\conv_lstm.json')
    weights_dir = os.path.join('models\cnn_lstm\conv_lstm.h5')
    with open(json_dir, 'r') as f:
        model_json = json.load(f)
        model = model_from_json(model_json)
        print("here")

    model.load_weights(weights_dir)
    return model

Flask服务代码 #

import argparse
import json
import os
import pandas as pd
from string import printable
import pickle as p

import flask 
from flask import Flask, jsonify, request

import tensorflow as tf
from tensorflow import keras
from keras_preprocessing.sequence import pad_sequences
from keras.models import model_from_json


def load_model():
    json_dir = os.path.join('models\cnn_lstm\conv_lstm.json')
    weights_dir = os.path.join('models\cnn_lstm\conv_lstm.h5')
    with open(json_dir, 'r') as f:
        model_json = json.load(f)
        model = model_from_json(model_json)
        print("here")

    model.load_weights(weights_dir)
    return model


def prepare_urls(urls):
    url_int_tokens = [
        [printable.index(x) + 1 for x in url if x in printable] for url in urls]

    # Cut URL string at max_len or pad with zeros if shorter, return result
    return pad_sequences(url_int_tokens, maxlen=75)

def predict_type(urls):
    model = load_model()
    predictions = model.predict(prepare_urls(urls))
    print(predictions)
    url_type = []

    for prediction in predictions:
        if prediction > 0.5:
            url_type.append('Malicious')
        else:
            url_type.append('Safe')
    
    return url_type


app = Flask(__name__)

@app.route('/')
def welcome():
    return 'Malicious URL Identification'

 
@app.route('/predict', methods=['GET', 'POST'])
def predict():
    if flask.request.method == 'POST':
        try:
            print(request.json)
            urls = pd.DataFrame(request.json, index=[0])            
            return jsonify(predict_type(urls))
            
        except Exception as e:
            return jsonify({
               "Exception": e
               })

 
if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.addargument('--model_name', type=str, default='cnn_lstm_model', metavar='N',
                        help='Provide the name of the model (default value is cnn_lstm_model)')
    args = parser.parse_args()

    app.run(port=8080, host='0.0.0.0')

异常发现 #

打印预测值时，所有输入的预测结果均相同，如下图所示：

内容的提问来源于stack exchange，提问作者YKM