machine-learning - 如何保存经过训练的强化学习代理以避免每次都对其进行训练？

Question

我尝试使用pickle来保存受污染的代理

   try:
      agent1 = pickle.load(open(model_file_path, 'rb'))
    except:
      print("An exception occurred")
      train_agent(True)

    if agent1 == None:
        train_agent(True)

    human = Human()
    human.set_sym(env1.o)

    agent1.set_verbose(True)
    start_session(agent1, human, Environment(), draw=2)

    pickle.dump(agent1, open(model_file_path, 'wb')) 

    return agent1.prediction

但是保存代理的文件变得非常重，大约 1GB，因此我无法恢复代理

Answer 1

HDF5 Format是一种网格格式，非常适合存储多维数字数组。例如:使用 Keras/Tensorflow 您可以非常轻松地保存/加载模型和权重:

# Save the model
model.save('path_to_my_model.h5')

# Recreate the exact same model purely from the file
new_model = keras.models.load_model('path_to_my_model.h5')

# Save weights
model.save_weights('path_to_my_weights.h5')

# Load weights
new_model.load_weights('path_to_my_weights.h5')