This is supposed to be an assortment of small toy stuff that I fancied to build to play around with a bit.
Nothing too serious, nothing production-ready and nothing really of too much interest to other parties.
A generator for shapes (e.g. lines, crosses, ...) in various colours which gets hooked up with a shape connector to solve the classification problem of identifying the shapes. The thing to note here is that all data is generated on the fly and throw-away, i.e. the model will ever only see the data once.
Usage:
- Start TensorBoard with
tensorboard --logdir ./lightning_logs - Start the training with
python3 trainer.py
If you want to change the batch sizes or other hyperparameters, then trainer.py is your friend.
For a more user-friendly version, please see the accompanying notebook. Note that the code there is mostly copy-pasted and then expanded from the regular source code. This had to be done as including source files when running a notebook on Colab is a hassle (read: did not get it to work before I lost interest in sinking time into that). However, the notebook is a bit prettier in terms of the analysis plots.
A notebook, preferrably run on Colab to make use of their free GPU instances. Use this link to launch the notebook directly from Github (needs a Google account, though) and do not forget to set the runtime to GPU: https://colab.research.google.com/github/notger/motley_crew_of_models/blob/main/timeseries_pred_lstm/demo_lstm_regression_for_timeseries.ipynb .
Creates a synthetic time series in the form of a Markov-series and then sets up an LSTM-model to learn and try to predict these. Very standard stuff, was just curious how well it would work with larger delays.
Just a home-written transformer, as I wanted to know how they work so I thought I wrote them for myself. No real problem attached to it, this time, as I think that Transformers are applied to problems which are way out league for everyone but Meta, Google or the like.