Creating some neural networks

Hello guys,

i am getting really desperate. For my bachelor thesis i need to create simple MLP classifying anything, simple MLP for MNIST, simple CNN for MNIST and simple LSTM/GRU network. I am really close to deadline and clueless, can you guys reffer me to some super tutorials for these networks, or even help me?

Thanks in advance!

Do you need to use existing algorithms to do the classification, or implement your owns?

1 Like

Existing algorithms would do just fine :slight_smile:

there is a zoo. search for flux zoo and you will find examples for the MNIST you mentioned.

also can DM me if you need some private tutoring.

Yeah i know MODEL.ZOO but it is still pretty confusing for me…is there somewhere just simple MLP with some detialed commentary?

I think you’re going to have to be a more specific about your background and what you understand about NNs/working with NNs. If you’re comfortable creating simple MLPs/CNNs/RNNs in something like Keras (or even PyTorch), then by all means post a small sample of what you want to run in Python and we can give you some tips for porting it.

However, if you’re struggling with the concepts and how to implement them in any language, trying to learn Julia + ML libraries at the same time is very much “running before you can walk”. Since you mentioned doing a bachelor’s thesis, there are a few avenues you could try:

  1. Consult your advisor and fellow students about this. NNs are so ubiquitous that there is almost guaranteed to be someone with a decent amount of experience that can guide you to make a simple model with any language. Once that’s done, see my first paragraph above about porting it if you still have time.
  2. If there are no human resources available, the top deep learning frameworks have great introductory tutorials with detailed commentary on all different kinds of networks. Most you can even copy-and-paste directly to tweak a few things (even better if they provide a runnable notebook).
  3. Once you’ve gone through with 1) or 2), you can revisit the model zoo tutorials. The 60 minute blitz in particular tries to match the PyTorch version. It can be run directly as a script, contains lots of commentary on each step and assumes very little prior knowledge of Flux or Julia.
1 Like

Another question: Are you required to use julia for this project? I’m a big fan of julia in general and I think Flux with CUDA.jl is a really nice way to experiment with new neural network architectures. I also think julia has great promise as a production environment for tough deep learning problems. The whole SciML ecosystem is really cool.

But having said all that, if you’re under time pressure and you just need to run a simple multi-layer perceptron classifier, I would consider doing it in Python with PyTorch. Your final code likely won’t be that much simpler than it would be in julia but you’ll find a lot more tutorials that will really break down the code step by step. And you’ll find more code snippets that just work out of the box.


Unfortunately it has to be Julia with Flux. I have like 8 days left, so maybe i could handle it…

Eight days should be plenty of time. FWIW, I recently started from ground zero and built a useful classifier with Flux in less than a day. I worked from the iris example in the model zoo.

I am wondering if the Flux tutorial on JuliaAcademy may be of some help to you?

1 Like

Read the Flux documentation. They walk you through all of this. 8 days is a lot of time :slight_smile:

Hello guys,

i took holiday in work this week to make my thesis as good as possible till thursday. I did watch some videos and read some notebook on julia academy and I think i was already pretty successful with MLP and MLP for MNIST. I am working now on CNN (i think i will get that done too), but i did not find anything useful yet for LSTM/GRU. I dont understand it very well from model zoo. Can you guys help me one more time to find something on LSTM/GRU?