dividiti / ck-crowdsource-dnn-optimization Goto Github PK

We can actually create current min set of Qt libraries for Windows and Linux as a CK package (with a tag qt-min) and then whenever we compile this app, we can actually install Qt libraries via CK - this will be a canonical way of reusing artifacts.

Just one more note - we may want to compile and run this app on ARM-based Linux so we should detect ARM/x86 architecture when extracting libs. See how it is done in ck-autotuning:package:compiler-llvm-3.9.0-linux-download ...

If app has been run allready it keeps same list of engines, models and imagenet set
if some packages were installed after app has been started
seletion items keep the same but it will be usefull not to force app restart after all changes in ck env.

I tried this app on a new Surface Pro 4 installing CK via pip using Anaconda3 python ... I get the following error:

Classification program stopped prematurely. Please, select the command below, copy it and run manually from command line to investigate the issue:

python -W ignore::DeprecationWarning C:\Users\ckdnn\Anaconda3\Scripts..\ck\kernel.py run program:caffe-classification-opencl --cmd_key=use_continuous --deps.caffemodel=8a0758effe14ae49 --deps.imagenet-aux=eb8b1076f899bf92 --deps.imagenet-val=2e35bb4b714d49aa

When I run it, I gets the following:

C:\Users\ckdnn\Anaconda3\python.EXE: can't open file 'C:\Users\ckdnn\Anaconda3\Scripts..\ck\kernel.py': [Errno 2] No such file or directory

However, I can run this fine:

run program:caffe-classification-opencl --cmd_key=use_continuous --deps.caffemodel=8a0758effe14ae49 --deps.imagenet-aux=eb8b1076f899bf92 --deps.imagenet-val=2e35bb4b714d49aa

The problem is that Anaconda uses slightly different mechanism for packages so you can't find kernel there ... As I mentioned, it may be better to run ck directly rather than trying to find kernel since I spent some time figuring out how to make it portable (it's not straightforward), what do you think, @dsavenko ?

Thanks a lot!!!

@psyhtest : user would expect to see not the available dnn-proxy libs in the Caffe engine menu, but the actual installed Caffe libs that are returned by ck show env --tags=lib,caffe

Should query batch size from user (new selector at the right panel) and pass it to net

Proposal available for main contributors here
https://docs.google.com/document/d/1Sz6gqahcaOQiQi-3R4miF195N-ORAHumIc_D9RpROz8/edit#heading=h.fjhsrumqmltx

On my notebook
Core i5, 8 RAM
after 3-4 images for CPU engine I've got Segmentation fault
Valgrind log for details:
segmentation_fault_2017_02_08.txt

On ck run, we find all classification programs. But some of them may not be compiled. This case should be detected, and user should be asked whether to compile the program or skip compilation (in which case the program won't be accessible from the UI). If the user chooses to compile, the compilation should be done right away before starting the UI. If the compilation fails, the UI is still started, but without this program.

caffe-classification on Window is built and run successfully, I can see predictions for default image in the console.

dnn-proxy-caffe-cpu is built too, found and loaded by the app:

but after running it fails (and therefore the app is crashed) with message:

F0213 18:31:22.879647  2720 layer_factory.cpp:62] Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type: Input (known types: Convolution, Eltwise, LRN, Pooling, Power, ReLU, Sigmoid, Softmax, Split, TanH)

this error message is contained in logs in directory bin\logs

It seems, the message tells about section in the prototxt file

input: "data"
input_shape: {
  dim: 1
  dim: 3
  dim: 227
  dim: 227
}

prototxt file, prepared by the app, is in bin/tmp directory

Add integration with travis and automatic build for

• Linux / Ubuntu
• Windows
• MacOS ?

Users should choose VALs instead of AUXs. and the respective AUX for the chosen VAL should be picked according to the VALs tags.

1. Selected CPU proxy
2. Start execution
3. Stop execution
4. Select OpenCL type of proxy
5. Start execution
6. App crashed because unable to load another version of caffe lib while other dependencies loaded successfully:

LOG:
Loading dependencies. Count: 10
Load dep lib 10 "/opt/intel/intel-opencl-1.2-6.3.0.1904/opencl-sdk/lib64/libOpenCL.so"
Load dep lib 9 "/usr/lib/x86_64-linux-gnu/hdf5/serial/libhdf5.so"
Load dep lib 8 "/home/nikolay/CK-TOOLS/lib-clblast-development-gcc-5.4.0-linux-64/install/lib/libclblast.so"
Load dep lib 7 "/usr/lib/x86_64-linux-gnu/libopencv_core.so"
Load dep lib 6 "/home/nikolay/CK-TOOLS/lib-caffe-bvlc-opencl-clblast-trunk-gcc-5.4.0-linux-64/install/lib/libcaffe.so"
Segmentation fault (core dumped)

See loading lib code in recognizer.cpp
QString Recognizer::loadDeps(const QStringList &depLibs)

Worst prediction found with marker value 0.001
Image: "/home/anton/data1/ILSVRC2012_val/ILSVRC2012_val_00015268.JPEG"
Top-1: "0.001 - tiger shark, Galeocerdo cuvieri (3)"
Correct: "0 - tobacco shop, tobacconist shop, tobacconist (860)"

It seems that SqueezeNet can't predict any image and think they all are tiger shark

Step to reproduce (only at my ubuntu at this moment)

• Install all required ck packages as listed at read.me
• run crowdsource-video-experiments-on-desktop
• select engine = Caffe OpenCL proxy
• Press Start
• Got error
  

Steps to reproduce

1. clone
   https://github.com/dividiti/crowdsource-video-experiments-on-desktop

2. open project in QT

3. run app
   styles not aplied

Solution:
Copy or create symlink from src/app.qss to /bin/app.qss

Make preprocess script to print some informational output for cases when something goes wrong.

I've got compilatino error

ck install package:dnn-proxy-caffe-opencl
...
0) lib-caffe-nvidia-0.15-cuda  Version 0.15: 0

...
CXX/LD -o .build_release/tools/extract_features.bin
.build_release/tools/extract_features.o: In function `google::protobuf::RepeatedField<float>::Reserve(int)':
extract_features.cpp:(.text._ZN6google8protobuf13RepeatedFieldIfE7ReserveEi[_ZN6google8protobuf13RepeatedFieldIfE7ReserveEi]+0x5b): undefined reference to `google::protobuf::Arena::AllocateAligned(std::type_info const*, unsigned long)'
.build_release/tools/extract_features.o: In function `int feature_extraction_pipeline<float>(int, char**)':
extract_features.cpp:(.text._Z27feature_extraction_pipelineIfEiiPPc[_Z27feature_extraction_pipelineIfEiiPPc]+0xbe8): undefined reference to `google::protobuf::internal::fixed_address_empty_string'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::WireFormatLite::WriteStringMaybeAliased(int, std::string const&, google::protobuf::io::CodedOutputStream*)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedOutputStream::WriteStringWithSizeToArray(std::string const&, unsigned char*)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::DecrementRecursionDepthAndPopLimit(int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::ReadVarint64Fallback()'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::Arena::AddListNode(void*, void (*)(void*))'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::WireFormat::ReadPackedEnumPreserveUnknowns(google::protobuf::io::CodedInputStream*, unsigned int, bool (*)(int), google::protobuf::UnknownFieldSet*, google::protobuf::RepeatedField<int>*)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::IncrementRecursionDepthAndPushLimit(int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::UnknownFieldSet::MergeToInternalMetdata(google::protobuf::UnknownFieldSet const&, google::protobuf::internal::InternalMetadataWithArena*)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::InitEmptyString()'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::empty_string_once_init_'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::ReadVarint32Fallback(unsigned int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::WireFormatLite::WriteBytesMaybeAliased(int, std::string const&, google::protobuf::io::CodedOutputStream*)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedOutputStream::WriteVarint32SlowPath(unsigned int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::ReadLengthAndPushLimit()'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::GeneratedMessageReflection::NewGeneratedMessageReflection(google::protobuf::Descriptor const*, google::protobuf::Message const*, int const*, int, int, int, int, int, int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::CheckEntireMessageConsumedAndPopLimit(int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::BytesUntilTotalBytesLimit() const'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::ReadVarintSizeAsIntFallback()'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::io::CodedInputStream::ReadTagFallback(unsigned int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::MergeFromFail(char const*, int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::RepeatedPtrFieldBase::InternalExtend(int)'
.build_release/lib/libcaffe.so: undefined reference to `google::protobuf::internal::ArenaStringPtr::AssignWithDefault(std::string const*, google::protobuf::internal::ArenaStringPtr)'
collect2: error: ld returned 1 exit status
make: *** [.build_release/tools/extract_features.bin] Error 1
Error: Building Caffe in '/home/defremov/CK-TOOLS/lib-caffe-nvidia-cuda-0.15-nvcc-8.0.44-lib.openblas-0.2.19-85636ff-linux-64/src' failed!
CK error: [program] package installation failed!

I'm finding "Images per second" (fps) more intuitive as a performance metric than "Time per image" (especially when the unit of measurement has to be guessed:)).

I think we should just display "Images per second" (or "Image/s" if space is limited).

Shows same top predicted result text for different images
but if over this text with mouse hint shows correct text:

Check MD5: failed for "/home/nikolay/Projects/crowdsource-video-experiments-on-desktop/bin/openscience/data/87f5e203c47544d6/model/squeezenet_v1.1.caffemodel" calculated: "42c581aa4ef6d5978287df3bbb9137f1" stored: "0357e4e11d173c72a01615888826bc8e"
FAIL "Downloading file 'squeezenet_v1.1.caffemodel' from https://github.com/DeepScale/SqueezeNet/blob/master/SqueezeNet_v1.1/squeezenet_v1.1.caffemodel?raw=true into /home/nikolay/Projects/crowdsource-video-experiments-on-desktop/bin/openscience/data/87f5e203c47544d6/model" "MD5 verification failed"

@Chunosov
Please describe steps of manual packaging from original sources
Then we will try to automate and add to CI (travis)