search

ignacio-rocco / sparse-ncnet

Implementation of Sparse-NCNet: "Efficient Neighbourhood Consensus Networks via Submanifold Sparse Convolutions"
MIT License
154 stars 18 forks source link

reproduce results #3

Closed niatzt closed 4 years ago

niatzt commented 4 years ago

Hi,

1) If I want to reproduce your results or train from scratch, which default hyper-parameters need to be modified? I assume at least set 'fe_finetune_params' to 23? And larger num_epochs? what else?

2) if I want to train on my own dataset and fine-tune on your existing model, do you have recommended 'fe_finetune_params'?

Thanks!