bhuman / compilednn Goto Github PK

At HULKs, we are experiencing sporadic segmentation faults at:

In our code, we get a segmentation fault when we are using nn.input(0)[528]. The following assertions hold true in our use case:

assert(nn.input(0).size() == 1024);
assert(nn.input(0).dims().size() == 3);
assert(nn.input(0).dims()[0] == 32);
assert(nn.input(0).dims()[1] == 32);
assert(nn.input(0).dims()[2] == 1);
assert(nn.input(0).rank() == 3);
assert(nn.input(0).capacity() == 8195);

The internal Tensor::buffer has length 32795 and alignmentShift has an upper bound of 64. I then conclude that the buffer should be large enough for our nn.input(0)[528].

CompiledNN is used in a multithreaded program where we have separate loaded models and compiled nets per thread (should therefore be thread-safe).

It seems that some input data triggers the segmentation fault more often than other, but we don't know why.

At least the individual operation compilers need to be automatically tested for correctness. This has already been started for the UpSampling2D operation. A further step would be testing the merging of layers.

• Activation
• Arithmetic
• BatchNormalization
• Concatenate
• Conv2D
• Cropping2D
• DConv2D
• Dense
• GlobalPooling2D
• Im2Col2D
• Pooling2D
• Softmax
• UInt8Input
• UpSampling2D
• ZeroPadding2D

Summary of the original issue description in the B-Human repository:

We do not know whether inference is always correct. Some bugs are known, but there are probably more that just do not occur in the models we use. We need systematic, unit-test-like tests which create all kinds of network architectures, apply them to random data and compare the result of CompiledNN to SimpleNN. The goal is 100% path coverage of all compiler classes, also considering that some layers can sometimes be fused. The tests should be a standalone executable in the CompiledNN repository. A nice addition would be the evaluation of execution times of different model configurations. E.g., once the implementations of Im2Col and Conv1x1 are done, there need to be systematic comparisons between Conv2D and Im2Col + Conv1x1.

Currently, softmax is only implemented on flat tensors and is thus only useful as the last layer of a classification model.

In the future, the softmax operation should allow multi-dimensional input and choosing an arbitrary axis along which the operation is applied.

and use it for Dense layers and 1x1 convolutions

Hello,

I'm trying to compile a simple model

inputs = Input((120,160,1))
conv1 = Conv2D(64, 3, activation = 'relu', padding = 'same')(inputs)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(128, 3, activation = 'relu', padding = 'same')(pool1)
up1 = UpSampling2D(size = (2,2))(conv2)
merge = Concatenate()([up1,conv1]                 )
conv3 = Conv2D(64, 3, activation = 'relu', padding = 'same')(merge)
outputs = Conv2D(64, (1,1))(conv3)

model = Model(input = inputs, output = outputs) 

but I get the following error, which i don't fully understand.

/home/barth/Workspace/CompiledNN/Src/CompiledNN/CompiledNN/Operations/UpSampling2D.cpp:25: ASSERT(input.dims(2) <= settings.xmmRegs() * 4) failed
Aborted (core dumped)

Could someone help me setup the project?
i tried making a main.cpp file with the code in the readme file.
I extended the CMakeFile.txt with

add_executable(main main.cpp)
target_link_libraries(main PRIVATE CompiledNN)

I build with

mkdir _build && cd _build
cmake .. -G "Unix Makefiles"
cmake --build .

but i get the following error:

CMakeFiles/main.dir/main.cpp.o: In Funktion »NeuralNetwork::CompiledNN::apply() const«:
main.cpp:(.text._ZNK13NeuralNetwork10CompiledNN5applyEv[_ZNK13NeuralNetwork10CompiledNN5applyEv]+0x3c): Warnung: undefinierter Verweis auf »Assert::print(char const*, int, char const*, ...)«
main.cpp:(.text._ZNK13NeuralNetwork10CompiledNN5applyEv[_ZNK13NeuralNetwork10CompiledNN5applyEv]+0x41): Warnung: undefinierter Verweis auf »Assert::abort()«
collect2: error: ld returned 1 exit status
CMakeFiles/main.dir/build.make:90: recipe for target 'main' failed
make[2]: *** [main] Error 1
CMakeFiles/Makefile2:104: recipe for target 'CMakeFiles/main.dir/all' failed
make[1]: *** [CMakeFiles/main.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

Models for sequence data need 1D variants of the convolutional and pooling layers.

AC: WhistleNetMk16 can be run in CompiledNN

Hey,

I just tried your repo and the command make install fails with the following error:

[100%] Built target CompiledNN
Install the project...
-- Install configuration: ""
-- Installing: /usr/local/lib/libCompiledNN.a
-- Installing: /usr/local/include/CompiledNN/CompiledNN.h
-- Installing: /usr/local/include/CompiledNN/Model.h
-- Installing: /usr/local/include/CompiledNN/SimpleNN.h
-- Installing: /usr/local/include/CompiledNN/Tensor.h
-- Installing: /usr/local/include/CompiledNN/CompiledNN/CompilationSettings.h
-- Installing: /usr/local/include/CompiledNN/Platform/BHAssert.h
CMake Error at cmake_install.cmake:67 (file):
  file INSTALL cannot find
  "/tmp/CompiledNN/3rdParty/B-Human/Tools/Math/BHMath.h": No such file or
  directory.

make: *** [Makefile:100: install] Error 1

This file does indeed not exist in my folder structure, but there is B-Human/MathBase/ and commit b6ecd81 which might be related.

GlobalPooling2D is currently only implemented for a number of channels between 3 and 64 (or 32 on 32-bit systems).

Hey,

I'm currently trying to compile my network with your lib. As the title says I get an ASSERT(result) failed in NeuralNetwork::Model::parseJSONModel.
In the example you are giving it says:

Model model;
model.load("model.h5");

Though I could not find a load(...) function and thus used:

Model model;
model.loadKerasHDF5("model.h5");

My network is saved using model.save(...), is there anything else needed?
I am using Keras Version 2.1.3. Which version are you using and have you yet experienced incompatible versions?

It needs to be handled at runtime, not at compile time. This currently limits the number of channels to at most one full batch (plus another non-full batch).

Currently, the loading ONNX files results in the following output:

/tmp/CompiledNN/Src/CompiledNN/Formats/ONNX.cpp:37: FAIL: Not yet implemented.

ONNX would be cool as it is not so dependent on the training framework and some code seems to be there.

Is the README the only documentation? In particular, I'm looking to create a tiny_dnn-compatible 2d convolutional layer using CompiledNN, in order to benchmark it against a new one I created "tiny_jnn", but I don't want to replace the rest of our pipeline - just the one layer.

I'm trying to add a new target with an example usage of this library to the cmake file. Linking against the CompiledNN library yields this error though.

If I understand the cmake stuff correctly, this is because the B-Human subdirectory is included as PRIVATE. So targets depending on CompiledNN don't transitively include the files as well.

To fix this, I think, one has to either include the B-Human subdirectory as PUBLIC or remove the uses from the public headers of CompiledNN.

This is because the store-operations write four elements even when there are only two or three, which overwrites data that need to be read in the future.