def __init__(self,num_actions=None,dim_obs=None,memory_size=100,memory_word_size=32,name='encode_agent',**kwargs):
#super(Agent, self).__init__(name=name,**kwargs)#如果继承别人的才用
#latent dim 和 image_code_size是相同的含义，用在不同的编码器中
self.num_actions= num_actions
self.memory_size =memory_size
self.memory_word_size = memory_word_size
#self._image_code_size = image_code_size
#self._image_encoder_decoder = EncoderDecoder.ImageEncoderDecoder(image_code_size=image_code_size)
self._name = name
self._obs_size = reduce(lambda x,y: x*y, list(dim_obs))
print("self.obs_size",self._obs_size)
self._im2state = EncoderDecoder.ImEncoder(self._obs_size,self.memory_word_size,64)
self._state2im = EncoderDecoder.ImDecoder(self._obs_size,self.memory_word_size,64)
self._vae = EncoderDecoder.VAE(self._obs_size,self.memory_word_size,64)
self._optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
self._controllercore = Controller.ControllerCore(num_actions=self.num_actions,num_node=10,memory_size=self.memory_size,memory_word_size=self.memory_word_size)
self._abstract_memory= self._controllercore.AbstractG
#这几个参数要改，因为要与图相关，这个只是矩阵相关
#print("memory_size",self.memory_size)
self._external_memory = Memory.ExternalMemory(memory_size=self.memory_size)
memory_num_reads = 2
memory_top_k = 3
self._memory_reader = MemReadWrite.MemReader(
memory_word_size=self.memory_word_size,#这个与隐藏层同维度
num_read_heads=memory_num_reads,
top_k=memory_top_k,
memory_size=self.memory_size)
#
self._memory_eraser =MemReadWrite.MemErase(
memory_word_size=self.memory_word_size,
memory_size=self.memory_size)
self._memory_writer = MemReadWrite.MemWriter(
memory_word_size=self.memory_word_size,
memory_size=self.memory_size)
self._memory_reconstructor = MemReconstruction.MemReconstructor(memory_word_size=self.memory_word_size)
def __init__(self,
num_actions=4,
num_node=10,
memory_size=100,
memory_word_size=100,
name='ControllerCore'):
#self.AbstractG = self._build_graph()
#由于目前还没有构造抽象图 20200430
#所以用原图代替
self.memory_size = memory_size
self.AbstractG = Memory.ExternalMemory(memory_size=self.memory_size)
self._read_content_similarity = utils.CosineSimilarity(
1, word_size=memory_word_size, name="read_content_similarity")
self.epsilon = 0.9
self.num_actions = num_actions
self.aggregator_cls = MeanAggregator
self.memory_word_size = memory_word_size
self.aggregator = self.aggregator_cls(self.memory_word_size,
self.memory_word_size,
name="aggregator",
concat=False)
def __init__(self,
num_actions=4,
num_node=10,
memory_size=100,
memory_word_size=32,
name='ControllerCore'):
# 外部存储的结构概览
self.memory_size = memory_size
self.memory_word_size = memory_word_size
# 构造抽象图,从Memory中重新实例化一个子图，作为抽象图
self.AbstractG = Memory.ExternalMemory(memory_size=self.memory_size)
self.aggregator_cls = MeanAggregator
self.aggregator = self.aggregator_cls(self.memory_word_size,
self.memory_word_size,
name="aggregator",
concat=False)
# 根据当前状态得到读取索引，用到相似度度量
self._read_content_similarity = utils.CosineSimilarity(
1, word_size=memory_word_size, name="read_content_similarity")
# 策略输出
self.epsilon = 0.9
self.num_actions = num_actions
self._optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
def __init__(self,
num_actions=None,
dim_obs=None,
memory_size=100,
memory_word_size=32,
name="TrainableAgent"):
'''
智能体对环境的基本认知，动作空间，状态空间
'''
self.num_actions = num_actions
# reduce (lambda x,y:x+y, [1,2,3]) 输出为 6
self._obs_size = reduce(lambda x, y: x * y, list(dim_obs))
self.memory_size = memory_size
self.memory_word_size = memory_word_size
'''
1.实例化各个类
五个模块的引入
'''
# 编解码器的实例化
self._im2state = EncoderDecoder.ImEncoder(self._obs_size,
self.memory_word_size, 64)
self._state2im = EncoderDecoder.ImDecoder(self._obs_size,
self.memory_word_size, 64)
self._vae = EncoderDecoder.VAE(self._obs_size, self.memory_word_size,
64)
self._optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
self._vdecoder = EncoderDecoder.VDecoder(
1, 200) # 因为没用batch 所以是一维输出，就是一个v值
self._vaev = EncoderDecoder.VAEV(self._obs_size, self.memory_word_size,
64)
# 控制器实例化
self._controllercore = Controller.ControllerCore(
num_actions=self.num_actions,
num_node=10,
memory_size=self.memory_size,
memory_word_size=self.memory_word_size)
self._aggregator = Controller.MeanAggregator(self.memory_word_size,
self.memory_word_size,
name="aggregator",
concat=False) #输入输出维度相同
self._aggmodel = Controller.AggModel(self.memory_word_size)
self.batch_num = 0 # 用来控制训练agg的batch起点，训练聚合器参数的时候用到
#这一版的a2c写的不够清晰
# self.a2cparams = {
# 'gamma': 0.99,
# 'value': 0.5,
# 'entropy':0.0001
# }
# self.a2cmodel = Controller.a2cModel(num_actions=self.num_actions)
# self.a2cmodel.compile(
# optimizer = ko.RMSprop(lr=0.0007),
# loss = [self._logits_loss,self._value_loss]
# )
self.a2cmodel = AC(self._obs_size, self.num_actions)
# 存储器实例化
self._abstract_memory = self._controllercore.AbstractG
self._external_memory = Memory.ExternalMemory(
memory_size=self.memory_size)
# 读写器实例化
memory_num_reads = 2
memory_top_k = 3
self._memory_reader = MemReadWrite.MemReader(
memory_word_size=self.memory_word_size, #这个与隐藏层同维度
num_read_heads=memory_num_reads,
top_k=memory_top_k,
memory_size=self.memory_size)
self._memory_eraser = MemReadWrite.MemErase(
memory_word_size=self.memory_word_size,
memory_size=self.memory_size)
self._memory_writer = MemReadWrite.MemWriter(
memory_word_size=self.memory_word_size,
memory_size=self.memory_size)
#重构机制实例化
self._memory_reconstructor = MemReconstruction.MemReconstructor(
memory_word_size=self.memory_word_size)
def __init__(self,
num_actions=None,
dim_obs=None,
memory_size=100,
memory_word_size=32,
name="TrainableAgent"):
'''
智能体对环境的基本认知，动作空间，状态空间
'''
self.num_actions = num_actions
self._obs_size = reduce(lambda x, y: x * y, list(dim_obs))
self.memory_size = memory_size
self.memory_word_size = memory_word_size
'''
1.实例化各个类
五个模块的引入
'''
# 编解码器的实例化
self._im2state = EncoderDecoder.ImEncoder(self._obs_size,
self.memory_word_size, 64)
self._state2im = EncoderDecoder.ImDecoder(self._obs_size,
self.memory_word_size, 64)
self._vae = EncoderDecoder.VAE(self._obs_size, self.memory_word_size,
64)
self._optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
self._vdecoder = EncoderDecoder.VDecoder(
1, 200) # 因为没用batch 所以是一维输出，就是一个v值
self._vaev = EncoderDecoder.VAEV(self._obs_size, self.memory_word_size,
64)
# 控制器实例化
self._controllercore = Controller.ControllerCore(
num_actions=self.num_actions,
num_node=10,
memory_size=self.memory_size,
memory_word_size=self.memory_word_size)
self._aggregator = Controller.MeanAggregator(self.memory_word_size,
self.memory_word_size,
name="aggregator",
concat=False) #输入输出维度相同
self._aggmodel = Controller.AggModel(self.memory_word_size)
self.batch_num = 0 # 用来控制训练agg的batch起点，训练聚合器参数的时候用到
# 存储器实例化
self._abstract_memory = self._controllercore.AbstractG
self._external_memory = Memory.ExternalMemory(
memory_size=self.memory_size)
# 读写器实例化
memory_num_reads = 2
memory_top_k = 3
self._memory_reader = MemReadWrite.MemReader(
memory_word_size=self.memory_word_size, #这个与隐藏层同维度
num_read_heads=memory_num_reads,
top_k=memory_top_k,
memory_size=self.memory_size)
self._memory_eraser = MemReadWrite.MemErase(
memory_word_size=self.memory_word_size,
memory_size=self.memory_size)
self._memory_writer = MemReadWrite.MemWriter(
memory_word_size=self.memory_word_size,
memory_size=self.memory_size)
#重构机制实例化
self._memory_reconstructor = MemReconstruction.MemReconstructor(
memory_word_size=self.memory_word_size)
