kevinlin311tw / cvpr16-deepbit Goto Github PK

Hi, I have a quick and basic question to solve a doubt.

For evaluating DeepBit on the (Photo Tourism) Brown dataset, did you up-sample each 64x64 patch to the 224x224 network input size?

Hi, I have a question,how the accuracy rate be calculated without a label? Thanks very much.

When I fine tune VGG16 with the code provided in stage1, loss1 increases until 0.25 and loss2 decreases to a small value < 1e-5. I use output bits: 64, batch size:32. Can someone give me any advice about this strange situation?

Hi, When I run the ./train.sh common, the following error ,how to solve the problem , thanks
error: check failed : fd != -1 (-1 vs. -1) File not found : DeepBit32_stage2_iter_5000.caffemodel

I find this line is different to your paper.
According to your paper, I think it should be

loss = loss + (binary[i]-0.5-bottom_data[i])*(binary[i]-0.5-bottom_data[i]);

Can you tell me where is I am wrong?

Execute "./prepare.sh" failed:

I can access dropbox.com but cannot access the model and dataset mentioned in prepare.sh.

Any alternative solution?

After installing the requirements, when I run the code make all -j8, I got this error.

LD -o .build_release/lib/libcaffe.so
ld: framework not found vecLib
clang: error: linker command failed with exit code 1 (use -v to see invocation)
make: *** [.build_release/lib/libcaffe.so] Error 1

So I think it can not find vecLib.
I use macOs Mojave 10.14.5.

So what should I do?
Thanks in advance

I had followed your instruction step by step, why that I can't get the ideal results presented in your paper. Is there anything that I should pay extra attention to?

Hello, can help me explain this loss code, I do not understand, thank you very much！For example, why add (i*channels) in dot product operation? why have two weight?

for (int i = 0; i < bottom[0]->num(); ++i) {
dist_sq_.mutable_cpu_data()[i] = caffe_cpu_dot(channels,
diff_.cpu_data() + (ichannels), diff_.cpu_data() + (ichannels));

loss += (rotation_weight2[i]*(rotation_weight1[i]*dist_sq_.cpu_data()[i]));

}

We get the following error when we do "make runtest". We use CUDA 7.5 on Ubuntu 14.04. Could you please suggest a solution for it?

https://github.com/kevinlin311tw/cvpr16-deepbit/blob/master/README.md

[----------] 4 tests from ImageDataLayerTest/0, where TypeParam = caffe::CPUDevice

[ RUN ] ImageDataLayerTest/0.TestShuffle

*** Aborted at 1470802253 (unix time) try "date -d @1470802253" if you are using GNU date ***

PC: @ 0x2b17d11def20 (unknown)

*** SIGSEGV (@0x0) received by PID 43804 (TID 0x2b17cbfc91c0) from PID 0; stack trace: ***

@     0x2b17d2301d40 (unknown)

@     0x2b17d11def20 (unknown)

@           0x569cac std::operator+<>()

@     0x2b17d1645117 caffe::ImageDataLayer<>::DataLayerSetUp()

@     0x2b17d1626ce3 caffe::BasePrefetchingDataLayer<>::LayerSetUp()

@           0x5519ea caffe::ImageDataLayerTest_TestShuffle_Test<>::TestBody()

@           0x7b4803 testing::internal::HandleExceptionsInMethodIfSupported<>()

@           0x7ab4e7 testing::Test::Run()

@           0x7ab58e testing::TestInfo::Run()

@           0x7ab695 testing::TestCase::Run()

@           0x7ae9d8 testing::internal::UnitTestImpl::RunAllTests()

@           0x7aec67 testing::UnitTest::Run()

@           0x44fd4a main

@     0x2b17d22ecec5 (unknown)

@           0x456169 (unknown)

@                0x0 (unknown)

make: *** [runtest] Segmentation fault (core dumped)

Section 4.2. "Results in Image Matching" What was the processing time of the descriptor ORB and DeepBit?
Because I would use the DeepBit in time real for image matching.

I got the following error while executing train.sh script.How can I handle this error? Thanks very much.

Check failed: fd != -1 (-1 vs. -1) File not found: solver_stage1.prototxt
*** Check failure stack trace: ***
@ 0x7ffbddd355cd google::LogMessage::Fail()
@ 0x7ffbddd37433 google::LogMessage::SendToLog()
@ 0x7ffbddd3515b google::LogMessage::Flush()
@ 0x7ffbddd37e1e google::LogMessageFatal::~LogMessageFatal()
@ 0x7ffbde3fa178 caffe::ReadProtoFromTextFile()
@ 0x7ffbde400ff6 caffe::ReadSolverParamsFromTextFileOrDie()
@ 0x40ad3a train()
@ 0x407fc0 main
@ 0x7ffbdcc2f830 __libc_start_main
@ 0x4087e9 _start
@ (nil) (unknown)

Hi guys,
Although I read the paper and have a look at the source code, I could not understand how it works with the unsupervised approach.

This code snippet is from K2_min_quantization_loss_layer.
How is bottom_data calculated?

  for ( int i = 0; i < count; i++)
  {
	if(bottom_data[i] > 0.5){
		binary[i] = 1;
	}
	else{
		binary[i] = 0; 
	}
  }
  for (int i = 0; i < count; ++i) {
	loss = loss + (binary[i]-bottom_data[i])*(binary[i]-bottom_data[i]);
        diff_.mutable_cpu_data()[0] = diff_.mutable_cpu_data()[0] + (binary[i]-bottom_data[i])*(-bottom_data[i]);
  } 

run_cifar10
cvpr16-deepbit startup
Cleared 0 solvers and 0 stand-alone nets
WARNING: Logging before InitGoogleLogging() is written to STDERR
I0617 21:46:16.693743 67684 upgrade_proto.cpp:52] Attempting to upgrade input file specified using deprecated V1LayerParameter: ./models/deepbit/deploy32.prototxt
F0617 21:46:16.715786 67684 upgrade_proto.cpp:660] Refusing to upgrade inconsistent NetParameter input; the definition includes both 'layer' and 'layers' fields. The current format defines 'layer' fields with string type like layer { type: 'Layer' ... } and not layers { type: LAYER ... }. Manually switch the definition to 'layer' format to continue.

Hi, thank you for your working.

I just wonder how I can change max_iter parameter (in Optimization step) as your paper presented. In your example with Cifar I could not find where to change max_iter.

Thank you in advance.

Hi, I have noticed that reported performance has far beyond baselines but which is mAP@1K. Consider your baselines are all mAP which should be lower than mAP@1K in your implementation. As claimed in section 4.3,

Following the settings in [24](Deep hashing for compact binary codes learning, CVPR 15) ...

however, DH reported mAP rather than mAP@1K (if you double check the setting in [24]). Could you please give us more details about the real mAP on Cifar10 to fully compare with others.
In our evaluation, even ITQ (well-known and still competitive baseline) has reached 20+ mAP@1K, which makes us wonder whether Deepbit outperforms baselines that much?

Anyway, good work and thanks for the open source codes.

I am trying to deploy the network with python. Can someone help me with converting converting the features from fc8_kevin to binary. For example, the output from fc8 is like this.

[[-0.56248081 -0.45775688 -1.30485046 -0.1604958 1.5815804 0.77081078
-0.7585181 -1.10825634 -1.4173944 -0.79846704 0.49331596 -0.42672595
0.13242441 -0.19356699 -1.40226376 -0.30305305 0.51420587 -0.613258
-0.56232476 -1.44308019 -1.60355341 -0.82647151 0.34802282 -1.08634388
-0.70180577 -1.15342426 -1.00441277 0.03183408 0.90380871 -0.03189395
-0.22369295 0.1908104 ]]

Hi,

I retrain the model on the NUSWIDE dataset. But the output binary codes for all the test and database samples are the same. Namely, the pdist2 matrix is zero matrix.

I wanna to know why this happen?

Thank you very much your paper. Is the larger the number of categories, the higher the number of hash code bits?Because I use your SSDH-VGG16-48 direct training, can't converge. The training data is in 100 categories。

Hi, I tried to use train.sh to get the training model. However, when caffe tried to load mean file from imagenet_mean.binaryproto, it gives me an error "check failed: !lines_.empty() File is empty". How can I handle this error? Thanks very much.