kevinlin311tw / caffe-cvprw15 Goto Github PK

I don't understand how to initialize weights by LSH. I have two points, but i don't know which is true.
The one is using LSH on the dataset before the training and we can get the mapped data. Then we can use the data pair (data, mapped data) to fine tuning our Net.
The other one is that we can get the mapping matrix by LSH and use this as the weights directly.

matlab is too expensive..
or maybe it can be completely implemented by caffe?

when i train my datas i face the problem.

After we execute the command >> run_cifar10, we get the >> MAP = 0.897373
Is this the result after reranking ? I mean the Fine-level Search.
@kevinlin311tw

When I test the feature after cropped image,the binary code change strongely,and result change when do the image retrieval,how to solve the problem?appreciate your help

#when i execute make all -j16, this is error codes
i tried everything i can do, still can not figure it out.
thanks very much if anyone can help me.

`In file included from ./include/caffe/loss_layers.hpp:11:0,
from src/caffe/layers/contrastive_loss_layer.cpp:5:
./include/caffe/neuron_layers.hpp:383:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:384:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:467:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:468:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:553:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:554:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
Makefile:501: recipe for target '.build_release/src/caffe/layers/contrastive_loss_layer.o' failed
make: *** [.build_release/src/caffe/layers/contrastive_loss_layer.o] Error 1
make: *** Waiting for unfinished jobs....

In file included from ./include/caffe/loss_layers.hpp:11:0,
from ./include/caffe/common_layers.hpp:12,
from ./include/caffe/vision_layers.hpp:10,
from src/caffe/layers/relu_layer.cpp:5:
./include/caffe/neuron_layers.hpp:383:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:384:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:467:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:468:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:553:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:554:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
In file included from ./include/caffe/vision_layers.hpp:10:0,
from src/caffe/layers/relu_layer.cpp:5:
./include/caffe/common_layers.hpp:410:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/common_layers.hpp:411:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
In file included from src/caffe/layers/relu_layer.cpp:5:0:
./include/caffe/vision_layers.hpp:148:10: error: ‘cudnnTensor4dDescriptor_t’ was not declared in this scope
vector<cudnnTensor4dDescriptor_t> bottom_descs_, top_descs_;
^
./include/caffe/vision_layers.hpp:148:35: error: template argument 1 is invalid
vector<cudnnTensor4dDescriptor_t> bottom_descs_, top_descs_;
^
./include/caffe/vision_layers.hpp:148:35: error: template argument 2 is invalid
./include/caffe/vision_layers.hpp:149:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bias_desc_;
^
./include/caffe/vision_layers.hpp:347:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_, top_desc_;
^
Makefile:501: recipe for target '.build_release/src/caffe/layers/relu_layer.o' failed
make: *** [.build_release/src/caffe/layers/relu_layer.o] Error 1

In file included from ./include/caffe/loss_layers.hpp:11:0,
from ./include/caffe/common_layers.hpp:12,
from ./include/caffe/vision_layers.hpp:10,
from src/caffe/layers/euclidean_loss_layer.cpp:6:
./include/caffe/neuron_layers.hpp:383:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:384:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:467:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:468:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:553:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:554:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
In file included from ./include/caffe/vision_layers.hpp:10:0,
from src/caffe/layers/euclidean_loss_layer.cpp:6:
./include/caffe/common_layers.hpp:410:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/common_layers.hpp:411:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
In file included from src/caffe/layers/euclidean_loss_layer.cpp:6:0:
./include/caffe/vision_layers.hpp:148:10: error: ‘cudnnTensor4dDescriptor_t’ was not declared in this scope
vector<cudnnTensor4dDescriptor_t> bottom_descs_, top_descs_;
^
./include/caffe/vision_layers.hpp:148:35: error: template argument 1 is invalid
vector<cudnnTensor4dDescriptor_t> bottom_descs_, top_descs_;
^
./include/caffe/vision_layers.hpp:148:35: error: template argument 2 is invalid
./include/caffe/vision_layers.hpp:149:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bias_desc_;
^
./include/caffe/vision_layers.hpp:347:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_, top_desc_;
^
Makefile:501: recipe for target '.build_release/src/caffe/layers/euclidean_loss_layer.o' failed
make: *** [.build_release/src/caffe/layers/euclidean_loss_layer.o] Error 1

In file included from ./include/caffe/loss_layers.hpp:11:0,
from ./include/caffe/common_layers.hpp:12,
from ./include/caffe/vision_layers.hpp:10,
from src/caffe/layers/tanh_layer.cpp:8:
./include/caffe/neuron_layers.hpp:383:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:384:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:467:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:468:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
./include/caffe/neuron_layers.hpp:553:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/neuron_layers.hpp:554:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
In file included from ./include/caffe/vision_layers.hpp:10:0,
from src/caffe/layers/tanh_layer.cpp:8:
./include/caffe/common_layers.hpp:410:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_;
^
./include/caffe/common_layers.hpp:411:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t top_desc_;
^
In file included from src/caffe/layers/tanh_layer.cpp:8:0:
./include/caffe/vision_layers.hpp:148:10: error: ‘cudnnTensor4dDescriptor_t’ was not declared in this scope
vector<cudnnTensor4dDescriptor_t> bottom_descs_, top_descs_;
^
./include/caffe/vision_layers.hpp:148:35: error: template argument 1 is invalid
vector<cudnnTensor4dDescriptor_t> bottom_descs_, top_descs_;
^
./include/caffe/vision_layers.hpp:148:35: error: template argument 2 is invalid
./include/caffe/vision_layers.hpp:149:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bias_desc_;
^
./include/caffe/vision_layers.hpp:347:3: error: ‘cudnnTensor4dDescriptor_t’ does not name a type
cudnnTensor4dDescriptor_t bottom_desc_, top_desc_;
^
Makefile:501: recipe for target '.build_release/src/caffe/layers/tanh_layer.o' failed
make: *** [.build_release/src/caffe/layers/tanh_layer.o] Error 1
`

I found that after running demo.m, I got binary_codes of images.

However, how can I achieve Image Retrieval after getting binary_codes? It seems that your code doesn't contain image retrieval part. @kevinlin311tw

As to this error, is my GPU memory too small? my GPU: GTX1050Ti 4G

At the beginning, i change both batch_size to 256 in train_prototxt file and test_prototxt file because the number of my training image is 71460 and test image is 18465, i only change this one parameter, that error occurs. i don't know if the proportion of training images and test images is correct. do you have some suggestions about this? oh, the origin size of image datasets is 250250 while cifar-10 is 3232. but i think it does't matter because when generating leveldb, the shell will resize all images to 256*256, right?

Later, i use the same configuration as your train_prototxt, test_prototxt and solver prototxt, but the highest accuracy is only about 43% and later it reduce to 0.08 gradually... do you know what the problem is?

At last i considered if my image dataset is reasonable, i craw them from taobao, i think taobao's image quality is not high because the backgroundn of image is too complex, and it may interferer training result.
Are my guessing right?

These problems have bothered me for a long time. i wish your help ,thanks!!

Dear,

Good day.
Thank you for your contribution!
But I can't the model for the BaiduYun or other link,could you have handle it?

When I run $ cd analysis $ gnuplot plot-p-at-k.gnuplot no response did not draw the image, what is the reason?

do you mean that given a query image and a training dataset with 1 million images, the code needs to compare its hash codes with the training images 1 million times? How much time does this take???!!!

Hi~ kevin, I get a problem shows like below after make all

[ 77%] Building CXX object tools/CMakeFiles/dump_network.bin.dir/dump_network.cpp.o
In file included from /home/aem/Downloads/caffe-cvprw15-master/tools/dump_network.cpp:23:0:
/home/aem/Downloads/caffe-cvprw15-master/include/caffe/util/io.hpp:8:18: fatal error: hdf5.h: No such file or directory
compilation terminated.
tools/CMakeFiles/dump_network.bin.dir/build.make:62: recipe for target 'tools/CMakeFiles/dump_network.bin.dir/dump_network.cpp.o' failed
make[2]: *** [tools/CMakeFiles/dump_network.bin.dir/dump_network.cpp.o] Error 1
CMakeFiles/Makefile2:4014: recipe for target 'tools/CMakeFiles/dump_network.bin.dir/all' failed
make[1]: *** [tools/CMakeFiles/dump_network.bin.dir/all] Error 2
Makefile:127: recipe for target 'all' failed
make: *** [all] Error 2

(Ubuntu16.04 opencv3 anaconda2 Matlab2014a python2.7)
I comment USE_CUDNN := 1 and uncomment CPU_ONLY=1 like #USE_CUDNN :=1 CPU_ONLY=1
INCLUDE_DIRS := $(PYTHON_INCLUDE) $(MATLAB_INCLUDE)/usr/local/include /usr/include /usr/include/hdf5/serial/ /usr/local/cuda/include
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /usr/lib/x86_64-linux-gnu /usr/lib/x86_64-linux-gnu/hdf5/serial

And try this way
$ sudo ln -s /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.10 /usr/lib/x86_64-linux-gnu/libhdf5.so
$ sudo ln -s /usr/lib/x86_64-linux-gnu/libhdf5_serial_hl.so.10 /usr/lib/x86_64-linux-gnu/libhdf5_hl.so

But still get the same, how could I solve this problem? Any reply will be appreciate. Thank you!

list_im = read_cell(filename); Do not have read_cell function

0804 21:10:01.285056 24109 solver.cpp:58] Creating training net from train_net file: train_CIFAR10_48.prototxt
E0804 21:10:01.285750 24109 upgrade_proto.cpp:594] Attempting to upgrade input file specified using deprecated V0LayerParameter: train_CIFAR10_48.prototxt
E0804 21:10:01.286242 24109 upgrade_proto.cpp:395] Unknown parameter det_fg_threshold for layer type data
E0804 21:10:01.286269 24109 upgrade_proto.cpp:405] Unknown parameter det_bg_threshold for layer type data
E0804 21:10:01.286285 24109 upgrade_proto.cpp:415] Unknown parameter det_fg_fraction for layer type data
E0804 21:10:01.286298 24109 upgrade_proto.cpp:425] Unknown parameter det_context_pad for layer type data
E0804 21:10:01.286310 24109 upgrade_proto.cpp:435] Unknown parameter det_crop_mode for layer type data
E0804 21:10:01.286380 24109 upgrade_proto.cpp:598] Warning: had one or more problems upgrading V0NetParameter to NetParameter (see above); continuing anyway.
E0804 21:10:01.286397 24109 upgrade_proto.cpp:604] Note that future Caffe releases will not support V0NetParameter; use ./build/tools/upgrade_net_proto_text for prototxt and ./build/tools/upgrade_net_proto_binary for model weights upgrade this and any other net protos to the new format.

.........................
I0804 21:04:28.309278 24038 data_layer.cpp:45] Opening leveldb cifar10_train_leveldb
And the code just stop here, no training!!!

Thanks a lot!

There are some errors when run the training script ./train_48.sh:

I0831 11:56:43.312741 24302 caffe.cpp:115] Finetuning from /home/dl/zy/DLBHC/caffe-cvprw15-master/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
F0831 11:56:43.668262 24302 upgrade_proto.cpp:630] Check failed: ReadProtoFromBinaryFile(param_file, param) Failed to parse NetParameter file: /home/dl/zy/DLBHC/caffe-cvprw15-master/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
*** Check failure stack trace: ***
@ 0x7f25713c0daa (unknown)
@ 0x7f25713c0ce4 (unknown)
@ 0x7f25713c06e6 (unknown)
@ 0x7f25713c3687 (unknown)
@ 0x4cec87 caffe::ReadNetParamsFromBinaryFileOrDie()
@ 0x48318a caffe::Net<>::CopyTrainedLayersFrom()
@ 0x415be6 train()
@ 0x4101ff main
@ 0x7f256d9cfec5 (unknown)
@ 0x4143b7 (unknown)
@ (nil) (unknown)

How to solve this problem？

How to train a new model using my own dataset? Thank you !

hello, i learn your paper and code recently, can you share the diffrents ways experiments data likes cifar10.thx

Could someone kindly please share the train_val.prototxt with VGG network to learn the binary hashing codes? Thanks a lot

They are both new layers, why the lrs are different?

Hi,
according to this project README.md description. The "Correction of computational cost" sections says you computes hamming distance between two 64-bit binary codes takes 23 ps (bitwise XOR operation). when i use C/C++ to achieve this process, and find compute hamming distance between two 48-bit binary codes takes 0.02us. so could you open this codes?
Thanks

I have a question about the extracted feature.
I read the code of matcaffe_batch_feat and found that you did not give a certain layer to extract the features just like the example in caffe. So my question is did you directly extract features from the last sigmoid layer? is the feature a 1 by 48 vector?

I try to download the related source ,but I find the address of all are invalid @kevinlin311tw

train_CIFAR10_48.prototxt:
name: "KevinNet_CIFAR10"
layers {
layer {
name: "data"
type: "data"
source: "cifar10_train_leveldb"
meanfile: "../../data/ilsvrc12/imagenet_mean.binaryproto"
batchsize: 32
cropsize: 227
mirror: true
det_context_pad: 16
det_crop_mode: "warp"
det_fg_threshold: 0.5
det_bg_threshold: 0.5
det_fg_fraction: 0.25
}
top: "data"
top: "label"
}

test_CIFAR10_48.prototxt:
name: "KevinNet_CIFAR10"
layers {
layer {
name: "data"
type: "data"
source: "cifar10_val_leveldb"
meanfile: "../../data/ilsvrc12/imagenet_mean.binaryproto"
batchsize: 32
cropsize: 227
mirror: true
det_context_pad: 16
det_crop_mode: "warp"
det_fg_threshold: 0.5
det_bg_threshold: 0.5
det_fg_fraction: 0.25
}
top: "data"
top: "label"
}

Hi! I was wondering if there were any pretrained models to compute image hashes that were trained on ImageNet or similar (coco, NUS-WIDE, etc).

I saw on this repo that there were models trained on Cifar10, but I was hoping to use it for some image hashes as is, ie, not re-training. Would be great if someone was willing to share their model parameters, Pytorch, caffe, whatever!

Thanks so much!

So I am wondering have you run the KMH，LSH，PCA-ITQ and other methods yourself or you cite the image retrieval results from other researcher' papers?

My data set contains 10,200 images, and the next four images are classified as a single category .
The number of my category is 2500.
With the deephash network training, loss dropped from 7.1 to 3.79. Then iterate 50000 times, loss keeps stable.
I want to know if this is a normal phenomenon. How can we let loss continue to decline?

Please help with the error. Thank You. :)

>> startup
caffe-cvprw15 startup
>> demo
    '/examples/cvprw15-cifar10/imgs/horse2.jpg'
    '/examples/cvprw15-cifar10/imgs/horse1.jpg'
    '/examples/cvprw15-cifar10/imgs/car2.jpg'
    '/examples/cvprw15-cifar10/imgs/bird2.jpg'
    '/examples/cvprw15-cifar10/imgs/bus1.jpg'
    '/examples/cvprw15-cifar10/imgs/car1.jpg'
    '/examples/cvprw15-cifar10/imgs/air1.jpg'
    '/examples/cvprw15-cifar10/imgs/bird1.jpg'
    '/examples/cvprw15-cifar10/imgs/air2.jpg'
    '/examples/cvprw15-cifar10/imgs/bus2.jpg'
    '/examples/cvprw15-cifar10/imgs/deer1.jpg'
    '/examples/cvprw15-cifar10/imgs/deer2.jpg'

Warning: Assuming batches of 10 images rest will be filled with zeros 
> In matcaffe_batch_feat at 35
  In demo at 29 
Undefined function 'caffe' for input arguments of type 'char'.

Error in matcaffe_init_feat (line 27)
  caffe('init', model_def_file, model_file)

Error in matcaffe_batch_feat (line 39)
matcaffe_init_feat(use_gpu, model_def_file, model_file);

Error in demo (line 29)
[feat_test , list_im] = matcaffe_batch_feat(test_file_list, use_gpu, feat_len, model_def_file,
model_file);

Hi, Kevinlin, if I have a dataset and the dataset has only about 10 classes and totally aorund 3000 images(before data augmentation), will the retrieval performance be influenced? If yes, how would it be influenced? Thank you !

I‘m very curious about how you do it. Which codes are modified or written ??? you add a new layer the model is change, how to fine tune the alexnet.caffemodel???
Thanks a lot !!!

I1127 10:48:39.340561 18156 net.cpp:220] Memory required for data: 219539460
I1127 10:48:39.340629 18156 solver.cpp:41] Solver scaffolding done.
I1127 10:48:39.340633 18156 caffe.cpp:115] Finetuning from ../../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
E1127 10:48:39.598387 18156 upgrade_proto.cpp:611] Attempting to upgrade input file specified using deprecated transformation parameters: ../../models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel
I1127 10:48:39.598426 18156 upgrade_proto.cpp:614] Successfully upgraded file specified using deprecated data transformation parameters.
E1127 10:48:39.598429 18156 upgrade_proto.cpp:616] Note that future Caffe releases will only support transform_param messages for transformation fields.
I1127 10:48:39.637908 18156 solver.cpp:160] Solving KevinNet_CIFAR10
I1127 10:48:39.637960 18156 solver.cpp:247] Iteration 0, Testing net (#0)
F1127 10:48:39.644999 18156 im2col.cu:59] Check failed: error == cudaSuccess (8 vs. 0) invalid device function
*** Check failure stack trace: ***
@ 0x7f71592fadaa (unknown)
@ 0x7f71592face4 (unknown)
@ 0x7f71592fa6e6 (unknown)
@ 0x7f71592fd687 (unknown)
@ 0x4d0a88 caffe::im2col_gpu<>()
@ 0x4bd4d7 caffe::ConvolutionLayer<>::Forward_gpu()
@ 0x457d5b caffe::Net<>::ForwardFromTo()
@ 0x458187 caffe::Net<>::ForwardPrefilled()
@ 0x48ec93 caffe::Solver<>::Test()
@ 0x48f566 caffe::Solver<>::TestAll()
@ 0x49605e caffe::Solver<>::Solve()
@ 0x4156a2 train()
@ 0x40fe31 main
@ 0x7f7156588f45 (unknown)
@ 0x4140d7 (unknown)
@ (nil) (unknown)

These problems have bothered me for a long time. i wish your help ,thanks!!

I get some error in running demo.m

caffe-cvprw15 startup
>> demo
Error using textscan
'BufSize' is no longer required and has been removed.

Error in read_cell (line 11)
fileLines = textscan(fid,'%s','delimiter',linesep,'BufSize',100000);

Error in matcaffe_batch_feat (line 28)
    list_im = read_cell(filename);

Error in demo (line 29)
[feat_test , list_im] = matcaffe_batch_feat(test_file_list, use_gpu, feat_len,
model_def_file, model_file);

And I tried to run run_cafair10.m,the same error about 'BufSize'.
Did I have the wrong dataset or something incorrect in the code?

I have read the README.md file and see your words:"You need to convert the dataset into leveldb format using "create_imagenet.sh". We will show you how to do this. To be continued.". Now I want to train myself Datasets with ImageNe pretrained model ,if there are any instruction about how to convert myself Datasets format？ I'm forward to your reply!

I have read the paper''Deep Learning of Binary Hash Codes for Fast Image Retrieval'', but I have some questions about optimization problem . I want to apply this method to the classification problem using tensorflow framwork.
1、What is the activation function of the F8 layer? Softmax ?
2、What is the loss function?
3、How to optimize ?

I use ubuntu 16.04 and your git code to compile matlab interface，matlab has already been rightly installed

and I also fix the Makefile.config by setting:
MATLAB_DIR := /usr/local/MATLAB/R2014a

but by running: make matcaffe -j40, it outputs errors like:

make: /usr/local/MATLAB/R2014a/bin/mexext: Command not found
/usr/local/MATLAB/R2014a/bin/mex matlab/caffe/matcaffe.cpp
CXX="g++"
CXXFLAGS="$CXXFLAGS -pthread -fPIC -DNDEBUG -O2 -DCPU_ONLY -I/usr/include/python2.7 -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/local/include -I/usr/include/hdf5/serial -I.build_release/src -I./src -I./include -Wall -Wno-sign-compare -Wno-uninitialized"
CXXLIBS="$CXXLIBS .build_release/lib/libcaffe.a -L/usr/lib -L/usr/local/lib -L/usr/lib -L/usr/lib/x86_64-linux-gnu -L/usr/lib/x86_64-linux-gnu/hdf5/serial -lpthread -lglog -lgflags -lprotobuf -lleveldb -lsnappy -llmdb -lboost_system -lhdf5_hl -lhdf5 -lopencv_core -lopencv_highgui -lopencv_imgproc -lboost_thread -lcblas -latlas" -output matlab/caffe/caffe.
/bin/sh: 1: /usr/local/MATLAB/R2014a/bin/mex: not found
Makefile:400: recipe for target 'matlab/caffe/caffe.' failed
make: *** [matlab/caffe/caffe.] Error 127

in the matlab setting path, there exists the mexext in /usr/local/MATLAB/R2014a/bin/mexext, i am newer in matlab, so can you help me figure out the problem?

I has install caffe, and successfully run examples .
erros:
head lines:
make: /usr/local/MATLAB/R2012a/bin/mexext: Command not found
g++ src/caffe/net.cpp -pthread -fPIC -DNDEBUG -O2 -DUSE_MKL -I/usr/include/python2.7 -I/usr/lib/python2.7/dist-packages/numpy/core/include -I/usr/local/include -I.build_release/src -I./src -I./include -I/usr/local/cuda/include -I/opt/intel/mkl/include -Wall -Wno-sign-compare -c -o .build_release/src/caffe/net.o 2> .build_release/src/caffe/net.o.warnings.txt
|| (cat .build_release/src/caffe/net.o.warnings.txt; exit 1)

the last few lines:
compilation terminated.
make: *** [.build_release/src/caffe/blob.o] Error 1
In file included from ./include/caffe/util/math_functions.hpp:11:0,
from ./include/caffe/syncedmem.hpp:7,
from ./include/caffe/blob.hpp:6,
from ./include/caffe/layer.hpp:8,
from src/caffe/layers/absval_layer.cpp:3:
./include/caffe/util/mkl_alternate.hpp:6:17: fatal error: mkl.h: No such file or directory
#include <mkl.h>
^
compilation terminated.
make: *** [.build_release/src/caffe/layers/absval_layer.o] Error 1

I run the scripts and get the caffemodel ,KevinNet_CIFAR10_48.caffemodel
Undefined function or variable 'matcaffe_batch_feat'.

Error in demo (line 29)
[feat_test , list_im] = matcaffe_batch_feat(test_file_list, use_gpu, feat_len,
model_def_file, model_file);