Hybrid CNN on Pawpularity Kaggle Contest
Kaggle contest link
Attempt to use the shortfused hybrid conv layer to improve a baseline VGG model with multi-modal inputs
After setting the correct data directory, run main.py using a TitanX gpu on gypsum - this works without error.
For the competition, the performance is not very satisfactory, with a public score of rmse = 22.7 on the baseline. Possible reaons include
- the outdated VGG model
- high memory complexity for more structured covariates to be used
- training from scratch on a dataset that's not large enough might not be better than transfer learning.
- There is some messy code for the hybrid conv layer implementation. Although this script ran without error when executed on gypsum using a titanX-long gpu, it threw errors when in the Kaggle environment. It's related to torch tensors being on different devices but I could not personally locate or resolve the issue.
- VGG baseline model is visibly outdated for the learning-based computer vision tasks nowadays, but the challenge is that it's non-trivial to incorporate shortfused layer to more complex CNN architectures such as EffecientNet, etc.
- Vision transformer models are very dominant across kaggle competitions and our approach is not known to be effective with transformers, at least not yet.