Tensorflow GraphDef不能大于2GB

我正在尝试用tensorflow实现一个深度自动编码器。RBM预训练工作正常，但在进行微调时，会出现错误：“ValueError:GraphDef不能大于2GB”。我的输入是一个[123968192]形状的数组，下面是我的层：[8192 16384 8192 4096 2048 1024 512 256 512 1024 2048 4096 8192 16384 8192]。 我知道问题在哪里，但我不知道如何解决它。我曾考虑过使用多个图形，但如果我的输入太大，甚至无法存储一层，该怎么办？此外，我不知道我应该使用多少图表。为每一层设置一个图表？这将是太慢和不必要的。 谢谢大家!

def __init__(self, input_num, layers, rbm_learning_rate, deepnn_learning_rate, rbm_num_epoch, 
             deepnn_num_epoch, momentum=0, batch_size=128, data_type='float32'):
    self.input_num = input_num
    self.layers = layers
    self.n_layers = len(self.layers)
    self.rbm_learning_rate = rbm_learning_rate
    self.deepnn_learning_rate = deepnn_learning_rate
    if momentum == 0:
        self.momentum = []            
        for _ in range(self.n_layers):            
            self.momentum.append(1)
    self.rbm_num_epoch = rbm_num_epoch
    self.deepnn_num_epoch = deepnn_num_epoch
    self.batch_size = batch_size
    self.data_type = data_type
    self.rbm_list = []
    self.rbm_list.append(RBM(self.input_num, self.layers[0], self.rbm_num_epoch, 
                            self.momentum[0], self.rbm_learning_rate[0], self.batch_size, self.data_type))            
    for i in range(self.n_layers-1):
        self.rbm_list.append(RBM(self.layers[i], self.layers[i+1], self.rbm_num_epoch,
                            self.momentum[i], self.rbm_learning_rate[i], self.batch_size, self.data_type))

def pretrain(self, train_set):
    self.W_list = []
    self.b_list = []
    self.a_list = []   
    if not cmp(train_set.dtype, self.data_type):
        train_set.dtype = self.data_type        
    next_train = train_set        
    for i, rboltz in enumerate(self.rbm_list):
        next_train = self._pretrain_and_get_para(rboltz, next_train)

def _pretrain_and_get_para(self, rboltz, next_train):
    output, W_out, a_out, b_out = rboltz.fit(next_train)
    self.W_list.append(W_out)
    self.a_list.append(a_out)
    self.b_list.append(b_out)
    return output

def fine_tune(self, train_set):
    m, _ = train_set.shape
    self.num_per_epoch = m / self.batch_size         
    train_batch = tf.placeholder(self.data_type, [None, self.input_num])        
    logits = self._build_model(train_batch)
    loss = self._loss(logits, train_batch)
    train_op = self._training(loss)
    init = tf.initialize_all_variables()
    with tf.Session() as sess:
        sess.run(init)
        for _ in range(self.deepnn_num_epoch):
            for i in range(self.num_per_epoch):
                _, cost = sess.run([train_op, loss], feed_dict = self._feed_build(train_batch, train_set, i))
            print cost    

def _feed_build(self, train_batch, train_set, i):
    batch = prepare_data.next_batch(train_set, i, self.batch_size)
    feed_dict = {train_batch: batch}
    return feed_dict    

def _build_model(self, train_batch):      
    middle_layer = self._make_encoder(train_batch)
    last_layer = self._make_decoder(middle_layer)
    return last_layer

def _make_encoder(self, train_batch):
    encoder = []
    encoder.append(train_batch)        
    for i, layer in enumerate(self.layers):
        with tf.name_scope('encoder'+str(i)):
            W = tf.Variable(self.W_list[i], name = 'weights')
            b = tf.Variable(self.b_list[i], name = 'biases')
            encoder.append(tf.sigmoid(b + tf.matmul(encoder[i], W)))
    return encoder[self.n_layers]


def _make_decoder(self, middle_layer):
    decoder = []
    decoder.append(middle_layer)        
    for i, layer in enumerate(self.layers):
        with tf.name_scope('decoder'+str(i)):
            W = tf.Variable(self.W_list[self.n_layers-i-1], name = 'weights')
            a = tf.Variable(self.a_list[self.n_layers-i-1], name = 'biases')
            decoder.append(tf.sigmoid(a + tf.matmul(decoder[i], W, transpose_b = True)))
    return decoder[self.n_layers]        

def _loss(self, logits, labels):
    loss = tf.nn.l2_loss(logits-labels)
    return loss

def _training(self, loss):
    optimizer = tf.train.GradientDescentOptimizer(self.deepnn_learning_rate)
    train_op = optimizer.minimize(loss)
    return train_op 

---

这个问题还在发生吗？解决这个问题的一种方法是用变量替换大常量，而变量不会占用graphdef中的空间。