accuracy layer not showing any accuracy.

47 views
Skip to first unread message

shom...@gmail.com

unread,
Aug 5, 2015, 1:26:25 PM8/5/15
to Caffe Users
I tried to use the NIN_cifar10 on my data. i downloaded the model from modelzoo. i changed the inut data layer as below(changes in bold).
i ran training successfully. then i tried testing. it ran successfully, but showed no accuracy value.
I0805 21:55:40.513593 30727 caffe.cpp:181] Batch 0, pool3 = 0
I0805 21:55:40.513607 30727 caffe.cpp:181] Batch 0, pool3 = 0
I0805 21:55:40.513619 30727 caffe.cpp:181] Batch 0, pool3 = 0
I0805 21:55:40.513633 30727 caffe.cpp:181] Batch 0, pool3 = 0
I0805 21:55:40.513646 30727 caffe.cpp:181] Batch 0, pool3 = 0
I0805 21:55:40.513659 30727 caffe.cpp:181] Batch 0, pool3 = 0

below is train_test.prototxt with changes in bold. the rest is unchanged except num_outputs =24 to suit my case.
my question is why the accuracy layer shows no accuracy value?
name: "NIN_CIFAR10"
layers {
  name: "data"
  type: IMAGE_DATA
  top: "data"
  top: "label"
  image_data_param {
    source: "examples/ASL/test_set_combinations/half_half/train_set_depth_new.txt"
    batch_size: 25
  }
  transform_param {
    mean_file: "examples/ASL/datasets/half_half/train_mean.binaryproto"
    mirror: true
  }
  include: { phase: TRAIN }
}
layers {
  name: "data"
  type: IMAGE_DATA
  top: "data"
  top: "label"
  image_data_param {
    source: "examples/ASL/datasets/test_set_combinations/half_half/test_set_depth_new.txt"
    batch_size: 100
  }
  transform_param {
    mean_file: "examples/ASL/datasets/half_half/test_mean.binaryproto"
    mirror: false
  }
  include: { phase: TEST }
}

Philip H

unread,
Aug 5, 2015, 5:12:13 PM8/5/15
to Caffe Users
usually, accuracy is evaluated on the test set. so you will only see the log for the accuracy when the solver does the testing step. depending on the settings of your solver you might just need to wait (or turn testing on at all)

cheers,
phil

shom...@gmail.com

unread,
Aug 6, 2015, 3:15:42 AM8/6/15
to Caffe Users
the cause was convention mismatch. i used layers for the most of the code, except the loss & accuracy layer. hence loss & accuracy were not incldued in the net. hence no accuracy output & also no error backpropa
Reply all
Reply to author
Forward
0 new messages