Suggested task for Outreachy Dec 22: Task 2 - Keras to MDF

[pgleeson]

A suggested set of tasks for applicants to the Outreachy Winter 2022 call.

Motivation: Keras is a simple API for creating neural network models. We want people to be able to create a model using Keras, train it, and then export it to MDF to use it with any MDF compatible tools.

1. Install MDF locally.
2. Install Keras
3. Create a simple feedforward network with a few layers in Keras. this can be a "dummy" network, or a trained network based on an existing Keras tutorial.
4. Create another python script, which loads the trained network (i.e. with the final connection weights) and parses the structure and exports it to the equivalent network in MDF (i.e. with one node per layer which implements the activation function and edges for the connections which contain the trained weights)
5. Create a simple Jupyter notebook which demonstrates the conversion, illustrates that the keras version and the MDF version produce the same results, and shows a graphical output of the MDF equivalent.
6. Email us a link to your work.

[AniketARS]

Hi @pgleeson, in point 4 edges for the connections which contain the trained weights what does this means if we have to calculate X@W+b on the edge? because i've created parameters on edge and evaluating graph but edge is not evaluating!!

[pgleeson]

@AniketARS This is the location where the weight on an edge is incorporated:

MDF/src/modeci_mdf/execution_engine.py

Line 1186 in c4c334d

input_value = weight * value

I guess we should replace that by a check on the type of the input/weight and potentially call numpy.matmul (though the parameters could be Tensors also...).

For now, you could incorporate the weight into the target node also, and do the multiplication explicitly there.

[AniketARS]

@pgleeson if I can work on about issue, for which you suggesting of adding check?

[pgleeson]

@AniketARS Last comment wasn't very clear..

[AniketARS]

Sorry for that @pgleeson I was asking to solve the issue you mentioned in #345 (reply in thread)

[pgleeson]

Thanks @AniketARS, but the ultimate solution for that will depend on a number of other conversations we're having in the project now, specifically about how much freedom should be given to parameter types in MDF, i.e. whether the size/shape/type of input/output parameters should be specified in advance or be free for the user to choose. In the latter case it means that we will have to support scalars as well as arrays/tensors in expressions like weight * value and deal with (and document!) the various options accordingly.

So this would potentially be a task for the successful applicant to undertake in discussion with the rest of the developers during the internship. As suggested, for now you could bundle the weight into the target node, and ensure it's being handled correctly there.

[Monsurat-Onabajo]

Hello, good evening, i was working on the second task, i have installed the modeci_mdf library but i discovered i cant plot the graphs, i kept getting this error

[Monsurat-Onabajo]

Hello, i was going through my github today and i noticed that my codes for this project has been cloned by seven people, what can i do? And my code has not been checked yet so i cant make a submission

[pgleeson]

Don't worry, we'll look over your code and keep an eye out for any suggestion of plagiarism. Everyone working this task should start with a different Keras network anyway, and take a different approach to converting it.

[Monsurat-Onabajo]

Ok, thank you very much

[Monsurat-Onabajo]

Hello, As regards the completion of my final Outreachy application, as one of the requirement, I was asked to contact my mentor for a timeline of work to be accomplished on various projects and tasks to be concluded at each stage.

[pgleeson]

See #346 (comment)

[kusanele]

Good day, are we permitted to use our own dataset (upload data from our local machine to the repo) to train the model?

[pgleeson]

Sure, that's fine, as long as the code can be rerun. It shouldn't take too long to train though.