Caffe：从经过训练的快照恢复失败并出现错误

Question

我用一些配置训练了我的网络，然后保存了它的快照。
现在我正在尝试从最后一个快照恢复，但失败并显示以下错误消息：

I0328 13:44:30.756110 24238 net.cpp:283] Network initialization done.
I0328 13:44:30.756206 24238 solver.cpp:60] Solver scaffolding done.
I0328 13:44:30.757062 24238 caffe.cpp:209] Resuming from /media/hossein/tmpstore/caffe_new/examples/cifar10/cifar10_full_relu_bn_iter_60000.caffemodel.h5
HDF5-DIAG: Error detected in HDF5 (1.8.15-patch1) thread 0:
  #000: H5D.c line 358 in H5Dopen2(): not found
    major: Dataset
    minor: Object not found
  #001: H5Gloc.c line 430 in H5G_loc_find(): can't find object
    major: Symbol table
    minor: Object not found
  #002: H5Gtraverse.c line 861 in H5G_traverse(): internal path traversal failed
    major: Symbol table
    minor: Object not found
  #003: H5Gtraverse.c line 641 in H5G_traverse_real(): traversal operator failed
    major: Symbol table
    minor: Callback failed
  #004: H5Gloc.c line 385 in H5G_loc_find_cb(): object 'iter' doesn't exist
    major: Symbol table
    minor: Object not found
F0328 13:44:30.786376 24238 hdf5.cpp:153] Check failed: status >= 0 (-1 vs. 0) Failed to load int dataset with name iter
*** Check failure stack trace: ***
    @     0x7f2d6e635daa  (unknown)
    @     0x7f2d6e635ce4  (unknown)
    @     0x7f2d6e6356e6  (unknown)
    @     0x7f2d6e638687  (unknown)
    @     0x7f2d6ed74acd  caffe::hdf5_load_int()
    @     0x7f2d6ed678d0  caffe::SGDSolver<>::RestoreSolverStateFromHDF5()
    @     0x7f2d6ed4bf19  caffe::Solver<>::Restore()
    @           0x408038  train()
    @           0x405a0c  main
    @     0x7f2d6d943ec5  (unknown)
    @           0x406141  (unknown)
    @              (nil)  (unknown)
Aborted (core dumped)

这就是我试图恢复它的方式：

#!/usr/bin/env sh

TOOLS=./build/tools

$TOOLS/caffe train \
    --solver=examples/cifar10/cifar10_full_solver_bn_lr2.prototxt \
    --snapshot=/media/hossein/tmpstore/caffe_new/examples/cifar10/cifar10_full_relu_bn_iter_60000.caffemodel.h5

然后我放弃了，我尝试使用BINARYPROTO 而不是HDF5，但我得到了这个错误：

I0328 16:35:34.721277 27243 net.cpp:283] Network initialization done.
I0328 16:35:34.721369 27243 solver.cpp:60] Solver scaffolding done.
I0328 16:35:34.722338 27243 caffe.cpp:209] Resuming from /media/hossein/tmpstore/caffe_new/examples/cifar10_full_relu_bn_iter_60000.caffemodel
F0328 16:35:39.143900 27243 sgd_solver.cpp:316] Check failed: state.history_size() == history_.size() (0 vs. 28) Incorrect length of history blobs.

*** Check failure stack trace: ***

    @     0x7fd1c2cbbdaa  (unknown)
    @     0x7fd1c2cbbce4  (unknown)
    @     0x7fd1c2cbb6e6  (unknown)
    @     0x7fd1c2cbe687  (unknown)
    @     0x7fd1c33ef097  caffe::SGDSolver<>::RestoreSolverStateFromBinaryProto()
    @     0x7fd1c33d1ed3  caffe::Solver<>::Restore()

    @           0x408038  train()
    @           0x405a0c
 main
    @     0x7fd1c1fc9ec5  (unknown)

    @           0x406141  (unknown)
    @              (nil)  (unknown)
Aborted (core dumped)

当我用不同的模型尝试不同的时间时，历史部分会发生变化（例如 58 vs 28、32 vs 28 和这样，总体错误是相同的，但数字不同！）

我该怎么办？这让我发疯了！

Answer 1

作为 --snapshot 参数的值，您必须传递 .solverstate.h5 文件，而不是 .caffemodel.h5 文件。