This release contains 3 pretrained models that utilize three different architectures:
-
Small model (model_110317) | 88.06% acc
- This model is trained with the topology that is defined in the git repository. Only 10 epochs.
-
Dense model (model_dense) | 95.1% acc
- This model is trained using a more naive approach and using a greater number of convolution layers with more intermediary nodes. It is very accurate but a very heavy model to train and interact with.
-
Sparse model (model_sparse) | 87.7% acc
- This model was an attempt to see if the convolution layers could be reduced without a significant loss of accuracy. I still need to work on this model in the future, but it is quite lightweight.
Please open an issue on the repository if there are any issues with the models.