framework caffe does not work when testing

Question

I am trying to use caffe test in order to check out the accuracy of caffennet. Pretty simple, isn't it?

Really NOT. It doesn't work at all.

Below I describe the steps I am following (To my mind everythink is OK):

I have downloaded imagenet training and validation set.

I have used the app "create_imagenet.sh" to create a LMDB database. Here I depict my "create_imagenet.sh" (I have modified some things since I only need validation database):

    RESIZE=true
    if $RESIZE; then
      RESIZE_HEIGHT=256
      RESIZE_WIDTH=256
    else
      RESIZE_HEIGHT=0
      RESIZE_WIDTH=0
    fi

if [ ! -d "$VAL_DATA_ROOT" ]; then
  echo "Error: VAL_DATA_ROOT is not a path to a directory: $VAL_DATA_ROOT"
  echo "Set the VAL_DATA_ROOT variable in create_imagenet.sh to the path" \
       "where the ImageNet validation data is stored."
  exit 1
fi

echo "Creating train lmdb..."

GLOG_logtostderr=1 $TOOLS/convert_imageset \
    -resize_height=$RESIZE_HEIGHT \
    -resize_width=$RESIZE_WIDTH \
    --shuffle \
    $VAL_DATA_ROOT \
    $DATA/val.txt \
    $EXAMPLE/ilsvrc12_val_leveldb

echo "Done."

in $VAL_DATA_ROOT I storage the path to the validation images, in $DATA the path tu val.txt file and finally example the path when I want to save my database. I can MAKE SURE that the paths are totally correct.

Once I have my database I have downloaded the weights file bvlc_reference_caffenet.caffemodel from official website and it's OK.

I have modified train_val.prototxt to link the database. Next I depict the first lines of the file. The rest does not change at all (respect to the original).

# mean pixel / channel-wise mean instead of mean image
#  transform_param {
#    crop_size: 227
#    mean_value: 104
#    mean_value: 117
#    mean_value: 123
#    mirror: true
#  }
  data_param {
    source: "/home/paco/Caffe_Original/caffe/examples/imagenet/ilsvrc12_train_lmdb"
    batch_size: 256
    backend: LMDB
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
#  transform_param {
#    mirror: false
#    crop_size: 227
#    mean_file: "/home/paco/Caffe/caffe/data/ilsvrc12/imagenet_mean.binaryproto"
#  }
# mean pixel / channel-wise mean instead of mean image
  transform_param {
    crop_size: 227
    mean_value: 104
    mean_value: 117
    mean_value: 123
    mirror: false
  }
  data_param {
    source: "/home/paco/imagenet_noformat/ilsvrc12_val_lmdb"
    batch_size: 50
    backend: LMDB
  }
}

I just execute caffe tool as follows:

caffe test -model=/home/paco/Caffe/caffe/models/bvlc_reference_caffenet/train_val.prototxt -weights=/home/paco/Caffe/caffe/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel -iterations=2000

As you all can see everything looks PERFECT and I don't obtain any error at all!!

However, it does not work since I obtain this on my shell!!:

**

I1016 16:43:32.526727 22178 net.cpp:255] Network initialization done.
I1016 16:43:32.675029 22178 upgrade_proto.cpp:46] Attempting to upgrade input file specified using deprecated transformation parameters: /home/paco/Caffe/caffe/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
I1016 16:43:32.675051 22178 upgrade_proto.cpp:49] Successfully upgraded file specified using deprecated data transformation parameters.
W1016 16:43:32.675071 22178 upgrade_proto.cpp:51] Note that future Caffe releases will only support transform_param messages for transformation fields.
I1016 16:43:32.675122 22178 upgrade_proto.cpp:55] Attempting to upgrade input file specified using deprecated V1LayerParameter: /home/paco/Caffe/caffe/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
I1016 16:43:32.882395 22178 upgrade_proto.cpp:63] Successfully upgraded file specified using deprecated V1LayerParameter
I1016 16:43:32.943102 22178 caffe.cpp:281] Running for 2000 iterations.
I1016 16:43:34.434036 22178 caffe.cpp:304] Batch 0, accuracy = 1
I1016 16:43:34.434161 22178 caffe.cpp:304] Batch 0, loss = -nan
I1016 16:43:35.862941 22178 caffe.cpp:304] Batch 1, accuracy = 1
I1016 16:43:35.862982 22178 caffe.cpp:304] Batch 1, loss = -nan
I1016 16:43:37.310729 22178 caffe.cpp:304] Batch 2, accuracy = 1
I1016 16:43:37.310768 22178 caffe.cpp:304] Batch 2, loss = -nan
I1016 16:43:38.719712 22178 caffe.cpp:304] Batch 3, accuracy = 1
I1016 16:43:38.719833 22178 caffe.cpp:304] Batch 3, loss = -nan
I1016 16:43:40.171242 22178 caffe.cpp:304] Batch 4, accuracy = 1
I1016 16:43:40.171273 22178 caffe.cpp:304] Batch 4, loss = -nan
I1016 16:43:41.674332 22178 caffe.cpp:304] Batch 5, accuracy = 1
I1016 16:43:41.674373 22178 caffe.cpp:304] Batch 5, loss = -nan
I1016 16:43:43.160789 22178 caffe.cpp:304] Batch 6, accuracy = 1
I1016 16:43:43.160826 22178 caffe.cpp:304] Batch 6, loss = -nan

**

This is even worse with alexnet and Squeezenet where following the same steps I ALWAYS obtain 0 of accuracy (really times to times I get 0.001 or something like that).

WHAT IS HAPPENENING? Maybe I have to preprocess the images before dump it into the database? ANYBODY CAN HELP ME? I WOULD REALLY APPRECIATE IT.

Thank you so much for your attention and help.

Paco Muñoz Martinez Paco Muñoz Martinez · Accepted Answer · 2018-10-18T08:53:30

The problem was MKL library. It was not working yet I was not given any information at all. I have fixed the problem by compiling caffe all over again using OpenBlas library.

Thank you to myself.

framework caffe does not work when testing

1 Answers