Bug? Caffe doesn't read the HDF5 files for TEST phase (instead reads the path given in TRAIN phase twice)

[duygusar]

Although I declare different training and validation txt files (that contains a list of hdf5 files), the
Net initializes the TEST phase with the path I give for the TRAIN phase.

For both TRAIN and TEST phase initialization I get:

Creating training net from net file: /home/caffe/models/det_alexnet/train_val.prototxt
... hdf5_data_layer.cpp:80] Loading list of HDF5 filenames from: /home/caffe/examples/det/train_h5_list.txt
... hdf5_data_layer.cpp:94] Number of HDF5 files: 17

Creating test net (#0) specified by net file: /home/caffe/models/det_alexnet/train_val.prototxt
... hdf5_data_layer.cpp:80] Loading list of HDF5 filenames from: /home/caffe/examples/det_/train_h5_list.txt_
... hdf5_data_layer.cpp:94] Number of HDF5 files: 17

Although my prototxt (train_val.prototxt) clearly provides different txt files for TRAIN and TEST phases:

layer {
name: "data"
type: "HDF5Data"
top: "imgdata"
top: "ph"
top: "sm"
include {
phase: TRAIN
}
hdf5_data_param {
source: "/home/caffe/examples/det/train_h5_list.txt"
batch_size: 256
}
}

layer {
name: "data"
type: "HDF5Data"
top: "imgdata"
top: "ph"
top: "sm"
include {
phase: TEST
}
hdf5_data_param {
source: "/home/caffe/examples/det/eval_h5_list.txt"
batch_size: 256
}
}

And I get this error:

Check failed: data_
*** Check failure stack trace: ***
@ 0x7fa264a335cd google::LogMessage::Fail()
@ 0x7fa264a35433 google::LogMessage::SendToLog()
@ 0x7fa264a3315b google::LogMessage::Flush()
@ 0x7fa264a35e1e google::LogMessageFatal::~LogMessageFatal()
@ 0x7fa2651d018b caffe::Blob<>::mutable_cpu_data()
@ 0x7fa2651ed4a1 caffe::hdf5_load_nd_dataset<>()
@ 0x7fa2650c3f73 caffe::HDF5DataLayer<>::LoadHDF5FileData()
@ 0x7fa2650c21fd caffe::HDF5DataLayer<>::Next()
@ 0x7fa265214a93 caffe::HDF5DataLayer<>::Forward_gpu()
@ 0x7fa265048a21 caffe::Net<>::ForwardFromTo()
@ 0x7fa265048b27 caffe::Net<>::Forward()
@ 0x7fa2651c7332 caffe::Solver<>::Test()
@ 0x7fa2651c7d4e caffe::Solver<>::TestAll()
@ 0x7fa2651cb2c7 caffe::Solver<>::Step()
@ 0x7fa2651cb58a caffe::Solver<>::Solve()
@ 0x40aba4 train()
@ 0x407390 main
@ 0x7fa2639a3830 __libc_start_main
@ 0x407bb9 _start
@ (nil) (unknown)
Aborted (core dumped)

PS: I work on only one GPU (Titan X)

[duygusar]

This seems to be a bug? I really don't understand why it would read the training file instead of the validation file provided at the TEST phase. I would appreciate if someone can look into this.