correlllab / nn4mc_cpp Goto Github PK

• FILE: filename.cpp
• METHOD/FUNCTION: method()
• RESULT: passing/failing/need_further_test
• TEST TYPE: Unit test/Integration test

DESCRIPTION:

• input:
• output:
• desired output:

NOTES:
[write additional notes here]

generator_test_v2(79319,0x105d775c0) malloc: *** error for object 0x7fa66cc25010: pointer being freed was not allocated
generator_test_v2(79319,0x105d775c0) malloc: *** set a breakpoint in malloc_error_break to debug

This happens when try to go:

weight_generator->addWeight(it->layer->W);
...
In code_generator.cpp

I've been getting this bug for a few weeks

• Have our own in-house ONNX parser.

This doesn't happen on Linux. It must be an install issue.

HDF5-DIAG: Error detected in HDF5 (1.10.4) thread 4558149056:
#000: H5L.c line 1252 in H5Lvisit(): link visitation failed
major: Links
minor: Iteration failed
#1: H5L.c line 3481 in H5L__visit(): link visitation failed
major: Links
minor: Iteration failed
#2: H5Gint.c line 1144 in H5G_visit(): not a location
major: Invalid arguments to routine
minor: Inappropriate type
#3: H5Gloc.c line 188 in H5G_loc(): invalid group ID
major: Invalid arguments to routine
minor: Bad value

Tested/passed:

• activations.h
• activations.cpp
• conv1d.h
• conv1d.cpp
• dense.h
• dense.cpp
• maxpool1d.h
• maxpool1d.cpp
• conv2d.h
• conv2d.cpp
• maxpool2d.h
• maxpool2d.cpp

Need to test (before March 1st)

• integration testing full nn4mc flow
• finishing up all padding options

Set up testing tutorial in this issue.

Update README, examples, tutorials and documentation, GitHub wiki...

generator/neural_network_generator.cpp

Ran 8 tests (7 failed)
(attached is output test results)

• exponential (Input integers from -100 to 100): most tests failed due to decimal imprecision, tests with input larger than 88 failed due by returning 'inf' in swig

• hard sigmoid (Input integers from -100 to 100): most tests passed, tests with inputs -2, -1, 1, 2 failed due to decimal imprecision

• hyper tan (Input integers from -100 to 100): some tests passed, all tests with input between -19 and 19 failed due to decimal imprecision

• relu (Input integers from -100 to 100): all tests passed

• sigmoid (Input integers from -100 to 100): some tests passed, all tests with input between -10 and 36 failed due to decimal imprecision, tests with input larger than 88 failed by returning 'nan'

• softmax (Input integers from -100 to 100): most tests failed due to decimal imprecision, tests with input larger than 88 failed by returning '0'

• softplus (Input integers from -100 to 100): some tests passed, all tests with input between -10 and 33 failed due to decimal imprecision, tests with input larger than 88 failed by returning 'inf'

• softsign (Input integers from -100 to 100): most tests failed due to decimal imprecision

All test cases are attached
testing_module_errors.txt

Add scripts that can run in the specific operating systems that make the installation easier.

• neural_network_generator is eating up the first ########################### characters

Done. Thanks Cooper :)

• define layer sizes (input_shape and output_shape) from parser

Done.

• end to end code generator.

Done.

• Activation function lookup

Done. TODO:

• Improve parser/ input/output shape readers

Done!

• Make activation.h and activation.c

Thanks Cooper!

• Make [N][M] and [N][M][K] squisher--> - overload fwdNN( .) or make a prepare window function
  In Progress...

• Figure out what to do when user has custom activation functions

• Weights and biases have the same name--> parser's fault

Done

• write arduino code that uses nn4mc--> Examples ...

• Finish all the installation scripts

• Test dumped code

If model type is HDF5: show only one follow up uploading button
If model type is PyTorch: show two follow up uploading buttons--> one that requests python model class and another one that requests pickle file.

Convolution1D = Conv1D
Convolution2D = Conv2D
Convolution3D = Conv3D
SeparableConvolution1D = SeparableConv1D
SeparableConvolution2D = SeparableConv2D
Convolution2DTranspose = Conv2DTranspose
Convolution3DTranspose = Conv3DTranspose
Deconvolution2D = Deconv2D = Conv2DTranspose
Deconvolution3D = Deconv3D = Conv3DTranspose

Sarah TODO:

• Purchase domain

Cooper TODO:

• Add domain to hosting service

Either do it myself or hire someone to do it.

Either implement it or hire somebody to do it.