Question

我有两个模型文件model.ckpt-10 model.ckpt-10.meta 恢复图表这两个选项有什么区别？ slim.assign_from_checkpoint_fn（）和saver.restore（sess，ckptfile）

它们都会导致类似的未初始化变量异常。

这是片段和例外。

metafile = '/tmp/tfslim_model/model.ckpt-10.meta'
ckptfile = '/tmp/tfslim_model/model.ckpt-10'

def vars_from_ckpt(metafile, ckptfile, sess):
    saver = tf.train.import_meta_graph(metafile) #
    l = tf.get_collection(tf.GraphKeys.VARIABLES)
    slim.assign_from_checkpoint_fn( ckptfile, l) #
    for v in l: print v.name ,v.eval()
    return l
    #saver.restore(sess, ckptfile) #restores but bombs later do

with tf.Session() as sess:
    l = vars_from_ckpt(metafile, ckptfile, sess)
    sess.run(l)

Traceback (most recent call last):
  File "test_wrk.py", line 26, in <module>
    l = vars_from_ckpt(metafile, ckptfile, sess)
  File "test_wrk.py", line 20, in vars_from_ckpt
    slim.assign_from_checkpoint_fn( ckptfile, l) #
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/framework/python/ops/variables.py", line 573, in assign_from_checkpoint_fn
    saver = tf_saver.Saver(var_list, reshape=reshape_variables)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 972, in __init__
    self.build()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 1001, in build
    restore_sequentially=self._restore_sequentially)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 618, in build
    assert restore_op.name.endswith("restore_all"), restore_op.name
AssertionError: save/restore_all_1
#exception happens here 
with tf.Session() as sess:
    l = vars_from_ckpt(metafile, ckptfile, sess)
    sess.run(l)

saver.restore(sess, ckptfile)

似乎打印变量和值会导致以下内容。 tensorflow.python.framework.errors.FailedPreconditionError：尝试使用未初始化的值kaggle_test / conv1_7_7 / biases_1

saver.restore（sess，ckptfile）vs slim.assign_from_checkpoint_fn

0 个答案: