Problem extracting features from MNIST autoencoder

[wri...@ohsu.edu]

Hi, all; I am encountering difficulty with the feature_extract.bin program of caffe RC2 in a Linux 2.6.32-504.30.3.el6.x86_64 enivronment when applied to the MNIST autoencoder. I've modified the model to use a data set of my own in an LMDB database; while I have not yet tried the autoencoder on the MNIST data itself, aside from changing the input file and the number of final output neurons the model is the same. The error that I am encountering is:

extract_features.bin _iter_5001.caffemodel mnist_autoencoder.prototxt decode1 ./features 10 lmdb

 E1019 15:19:58.720538 184586 extract_features.cpp:64] Using CPU
F1019 15:19:58.725827 184586 insert_splits.cpp:35] Unknown blob input data to layer 0
*** Check failure stack trace: ***
    @     0x2ae1a1cbad0d  google::LogMessage::Fail()
    @     0x2ae1a1cc00ba  google::LogMessage::SendToLog()
    @     0x2ae1a1cbcde0  google::LogMessage::Flush()
    @     0x2ae1a1cbd0cd  google::LogMessageFatal::~LogMessageFatal()
    @     0x2ae1a19eaf1d  caffe::InsertSplits()
    @     0x2ae1a1a206e5  caffe::Net<>::Init()
    @     0x2ae1a1a22a51  caffe::Net<>::Net()
    @           0x408834  feature_extraction_pipeline<>()
    @       0x303361ed5d  (unknown)
    @           0x404c49  (unknown)
Aborted (core dumped)

The model seems to train without issues. Looking at previous topics related to this error I have tried removing all but one of the training iterations for the model from the training file to avoid repeating the data keyword in the model definition and re-compiling after using "make superclean" since some posters have said this is a version issue with protobuf, but nothing has fixed the issue. I am currently using 2.6.1 as the 3.0 alpha and beta versions of protobuf don't seem to be compatible at the compilation stage. Any thoughts as to what could be going on?

Hollis

[wri...@ohsu.edu]

A quick update: same problem occurs if I run the MNIST autoencoder as is on the MNIST data. I only ran the model out to 5000 iterations, but aside from that and changing paths to accommodate the filesystem on the cluster I'm using the model files are stock. As before, the error is:

extract_features.bin hw_mnist_iter_5001.caffemodel mnist_autoencoder.prototxt data ./features 10 lmdb

E1020 13:34:13.107272 188340 extract_features.cpp:64] Using CPU
F1020 13:34:13.111383 188340 insert_splits.cpp:35] Unknown blob input data to layer 0

so it doesn't appear to be something that I did to the model file, unless there is something in extract_features that depends on the original lmdb files?

Hollis