以下代码清单：
class Dense:
def __init__(self, hidden_dim, output_dim):
self.hidden_dim = hidden_dim
self.output_dim = output_dim
with tf.variable_scope('Dense', reuse=tf.AUTO_REUSE):
self.w = tf.get_variable('Dense_weight_matrix', shape=[hidden_dim, output_dim], dtype=tf.float32, initializer=tf.truncated_normal_initializer())
self.b = tf.get_variable('Dense_bias_vector', shape =[output_dim], dtype=tf.float32, initializer=tf.constant_initializer(0))
def predict(self, x):
return tf.matmul(x, self.w) + self.b
class Encoder:
def __init__(self, num_features, seq_len, hidden_dim, num_layers):
self.num_features = num_features
self.hidden_dim = hidden_dim
self.num_layers = num_layers
self.input = [tf.placeholder(tf.float32, shape=(None, num_features), name="Input_{}".format(t)) for t in range(seq_len)]
with tf.variable_scope('Encoder_LSTMCell'):
cells = []
for i in range(num_layers):
with tf.variable_scope('RNN_{}'.format(i)):
cells.append(tf.contrib.rnn.LSTMCell(hidden_dim))
self.cell = tf.contrib.rnn.MultiRNNCell(cells)
_, self.state = tf.contrib.rnn.static_rnn(self.cell, self.input, dtype=tf.float32)
class Decoder:
def __init__(self, seq_len, hidden_dim, output_dim, num_layers, scope):
self.hidden_dim = hidden_dim
self.output_dim = output_dim
self.num_layers = num_layers
self.target = [tf.placeholder(tf.float32, shape=(None, output_dim), name="Target_".format(t)) for t in range(seq_len)]
self.input = [tf.zeros_like(self.target[0], dtype=tf.float32, name='GO')] + self.target[:-1]
self.dense = self.dense = Dense(hidden_dim, output_dim)
self.scope = scope
with tf.variable_scope('Decoder_LSTMCell'):
cells = []
for i in range(num_layers):
with tf.variable_scope('RNN_{}'.format(i)):
cells.append(tf.contrib.rnn.LSTMCell(hidden_dim))
self.cell = tf.contrib.rnn.MultiRNNCell(cells)
def predict(self, init_state, feed_prev=False):
with self.scope:
state = init_state
outputs = []
prev = None
for i, inp in enumerate(self.input):
if feed_prev and prev is not None:
with variable_scope.variable_scope("predict_func", reuse=True):
inp = self.dense.predict(prev)
if i > 0:
variable_scope.get_variable_scope().reuse_variables()
output, state = self.cell(inp, state)
outputs.append(output)
if feed_prev:
prev = output
outputs = [self.dense.predict(o) for o in outputs]
return outputs
class Seq2seq:
def __init__(self, input_seq_len, output_seq_len, num_features, hidden_dim, output_dim, num_layers):
tf.reset_default_graph()
self.scope = tf.variable_scope("Seq2seq")
with self.scope:
self.encoder = Encoder(num_features, input_seq_len, hidden_dim, num_layers)
self.decoder = Decoder(output_seq_len, hidden_dim, output_dim, num_layers, self.scope)
def loss(self, alpha=0.03):
with tf.variable_scope('Loss'):
output_loss = 0
for y, t in zip(self.predict(), self.decoder.target):
output_loss += tf.reduce_mean(tf.pow(y-t, 2))
reg_loss = 0
for var in tf.trainable_variables():
if 'weight' in var.name:
reg_loss += tf.reduce_mean(tf.nn.l2_loss(var))
return output_loss + alpha * reg_loss
def optimize(self, learning_rate = 0.01, alpha = 0.03, gradient_clip=2.5):
global_step = tf.Variable(initial_value=0,
name="global_step",
trainable=False,
collections = [tf.GraphKeys.GLOBAL_STEP, tf.GraphKeys.GLOBAL_VARIABLES])
with tf.variable_scope('Optimizer'):
train_op = tf.contrib.layers.optimize_loss(loss=self.loss(alpha), learning_rate=learning_rate, global_step=global_step, optimizer='Adam', clip_gradients=gradient_clip)
return train_op
def predict(self, feed_prev=False):
enc_state = self.encoder.state
preds = self.decoder.predict(enc_state, feed_prev)
return preds
试图从以下仓库中重构代码： https://github.com/aaxwaz/Multivariate-Time-Series-forecast-using-seq2seq-in-TensorFlow/blob/master/build_model_basic.py
创建和训练模型效果很好：
model = Seq2seq(input_seq_len=input_seq_len, output_seq_len=output_seq_len, num_features=num_features, hidden_dim=64, output_dim=output_dim, num_layers=2)
loss = model.loss()
train_op = model.optimize()
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
for i in range(total_iterations):
batch_input, batch_output = generate_train_samples(batch_size=batch_size)
feed_dict = {model.encoder.input[t]:batch_input[:,t].reshape(-1, num_features) for t in range(input_seq_len)}
feed_dict.update({model.decoder.target[t]:batch_output[:,t].reshape(-1, output_dim) for t in range(output_seq_len)})
_, loss_t = sess.run([train_op, loss], feed_dict)
print(loss_t)
saver = tf.train.Saver()
save_path = saver.save(sess, 'refactored.ckpt'))
尝试还原保存的模型时：
model = Seq2seq(input_seq_len=input_seq_len, output_seq_len=output_seq_len, num_features=num_features, hidden_dim=64, output_dim=output_dim, num_layers=2)
pred_op = model.predict(feed_prev=True)
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
saver = tf.train.Saver()
saver.restore(sess, save_path)
我收到以下错误：
NotFoundError：从检查点还原失败。这很可能是由于检查点缺少变量名或其他图形键。请确保您没有更改基于检查点的预期图形。原始错误：
在检查点中找不到键Seq2seq / multi_rnn_cell / cell_0 / lstm_cell / bias [{ ：0 /设备：CPU：0“]（_ arg_save / Const_0_0，保存/还原V2 / tensor_names，保存/还原V2 / shape_and_slices）]]