InceptionResNetV2模型转移学习后验证准确性低

Question

我需要一个Tensorflow模型将图像分类为4个不同的类别，为此我将在预训练的InceptionResNetV2模型（权重='Imagenet'）上进行迁移学习。在model.fit（）期间，我的准确度为97.4％，损失为0.3，而我的验证准确度仍为84％，损失为0.4。我过拟合了，如何提高验证准确性？

version: '3.7'

x-common-python-api:
   &default-python-api
    build:
      context: /Users/AjayB/Desktop/python-api/
    networks:
      - app-tier
    environment:
      - PYTHON_API_ENV=development
    volumes:
      - .:/python_api/

services:

  python-api:
    <<: *default-python-api
    ports:
      - "8000:8000"
    depends_on:
      - python-model
    command: >
      sh -c "ls /python-api/ &&
             python_api_setup.sh development
             python manage.py migrate &&
             python manage.py runserver 0.0.0.0:8000"

  python-model: &python-model
     .
     .
     .

  python-celery:
    <<: *default-python-api
    links:
      - redis:redis
    depends_on:
      - redis
    command: >
      sh -c "celery -A server worker -l info"

  redis:
     .
     .
     .