我正在尝试使用 tensorflow 实现深度自动编码器。RBM 预训练工作得很好，但是在进行微调时，它会引发错误：'ValueError: GraphDef cannot be greater than 2GB'。我的输入是 [12396, 8192] 形状的数组，这是我的图层：[8192 16384 8192 4096 2048 1024 512 256 512 1024 2048 4096 8192 16384 8192]。我知道问题出在哪里，但我不知道如何解决它。我曾考虑过使用多个图，但是如果我的输入太大而无法存储一层怎么办？此外我不知道我应该使用多少个图表。为每一层设置图表？那将太慢且没有必要。谢谢！
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