multiset

Loss Functions for Multiset Prediction

Running

1. Generate an MNIST-multi dataset

2. Run train.py with suitable cmd line arguments e.g:

python train.py --dataset-path data/mnist_multi_70000_min20_max50_4 --mnist-multi \
                --max-objects 4 --loss multiset_loss --use-cuda

Run train.py -h for cmd line argument details.

Note that --dataset-path and --max-objects vary based on the MNIST Multi dataset used.

Choose the loss with --loss.

When using the sequential loss (--loss ce_loss), choose an ordering strategy with --label-order.