noranta4
commented
3 years ago Concerning the data, you can find many datasets online suitable for this project, like images of 3D shapes from DeepMind. If you want to work directly with 3D point clouds you can look at something like COMA from MPI, you can also generate your own 3D version of MNIST, like in this work.
Concerning the computational cost problem, we expect every project to be feasible using Colab. Choose datasets and model size such that a training run can be completed in less than 10 hours; you should provide a demonstration, not a production model. Sometimes the runtime got disconnected, sometimes electricity goes out. Be sure to checkpoint your models during training.
PS where does the author talk about weeks or days of training? I did not spot that reference in the paper, just a footnote on page 6 saying somewhat the opposite...
colloroneluca
Hi, professor! I was thinking about the projects and I have chosen the one about VQ-Draw, I read the paper about it and found it exceptionally well written and clear. Anyway, starting to plan logistically what to do I encountered a big issue: In the project description it is stated that we should perform some experiments on different kind of data (eg. 3d data), so : -How can I retrieve this kind of data? -Since the model is really slow to train (the author talks about days or weeks!), how can we perform the above experiments? My hardware is prehistoric and google collab or similar detach the runtime after some inactive time.
I hope you can suggest some solutions/different approaches because I'm really interested in autoencoder/decoder architectures.
Have a nice evening and weekend!
Luca Collorone