jabacat / jml Goto Github PK

The current build and test workflow that I wrote is broken in a few regards.

• It only tests MinGW on Windows.
• Linux and macOS install outdated Meson and Ninja versions
• Meson and ninja versions are not fixed
• Some other major issues that I forgot.

Assign yourself to this issue if you'd be interested in helping

Correct activation functions

• List activation functions
• Check all listed functions

Make a graph of each function and display it

Because of course it is. See here. This seems like a real mess to fix (having to add this weird compiler internal attribute to every symbol).

Use this thread to discuss how we want to design the multithreading capabilities. Thread pool? API choices? Etc

We need to add layers to matrices and make an implementation for the "apply" function. Could anyone do this by Wednesday?

We need some way to load data formats, such as CSV, Apache Parquet, HDF. There's libraries for these in C++, but I'm not sure how well they integrate. I don't know if there's a ubiquitous data science data library like pandas for C++, which might be preferable to use.

• Add members inside the Model class for training and testing data, as well as methods to add training and testing data after creation

Do this after #33. Add more loss functions than just mean squared error.

Waiting on #18.

This is a minor one, to be completed after #3 is merged.

Preferable: use CMake to create a single object file, or similar. The problem is that (obviously) there's no main function in the library, so we need to build to an object file instead of to an executable. Update scripts/build.sh for this, once PR #9 has been merged.

We should be able to see the performance of the neural network on its training data. To that end, we need a function to show the neural network average loss over the testing data, using the mean squared error at first.

Right now, when a vector is disposed of through the C API, its slots in the vectors array just keeps wasting memory.
We want it to be the case that when a new vector is made, it takes up the first free spot in the array if there is any and only extends the array if it needs to.

We would like to randomly set the initial matrix weights and bias vectors.

This is to implement the functions get_input_size, get_output_size, set_input_size, set_output_size,and is_specified.

As per the API document:

get_input_size
Scope: jml::Model::
Arguments:
Expected constraints:
Returns:
unsigned
Description:
This function queries the input size for a model.

get_output_size
Scope: jml::Model::
Arguments:
Expected constraints:
Returns:
unsigned
Description:
This function queries the output size for a model.

set_input_size
Scope: jml::Model::
Arguments:
unsigned size
Expected constraints:
size > 0
Returns: void
Description:
This function sets the input size for a model. This function errors if the model already has attached layers or training data.

set_output_size
Scope: jml::Model::
Arguments:
unsigned size
Expected constraints:
size > 0
Returns: void
Description:
This function sets the output size for a model. This function errors if the model already has attached layers or training data.

is_specified
Scope: jml::Model::
Arguments:
Expected constraints:
Returns:
bool
Description:
This function queries whether a given model is fully specified, and ready to be given training data and layers. This should return true if input and output size have been specified, and false otherwise.