Script for the linear path plots. Script will first train a model on task 1. Then from this model two new models are trained with different buffers and data order up to task 5. 16 paths are calculated:
- From w1 to w2, T1 test data
- From w1 to w2, T1 train data
- From w1 to w2, on the first memory
- From w1 to w2, on the second memory
- From w2 to w2' T1 test
- From w2 to W2' T1 + T2 test
- From w2 to w2', first memory
- From w2 to w2', second memory
- From w1 to w5, T1 test data
- From w1 to w5, T1 train data
- From w1 to w5, on the first memory
- From w1 to w5, on the second memory
- From w5 to w5' T1 test
- From w5 to W5' T{1...5} test
- From w5 to w5', first memory
- From w5 to w5', second memory
Each path is by default sampled ten times. This can be changed in the script. The file linear_path_idx_[F_IDX].csv
stores an identifier for each path in the right category. The raw loss values are in linear_path_raw_[F_IDX].csv,
where each path is identifiable by its ID in the first row.
Options:
--data DATASET: which data to train on (mnist, cifar, min)--f_idx F_IDX: Index to append to file. Files should be merged manually to file with F_IDX=m before plotting.